This page intentionally left blank

S CIENT IFIC SOF TWARE D E SIGN

THE OBJECT-ORIENTED WAY
This book is about software design. Although many current discussions of scientic programming focus on scalable performance, this
book focuses on scalable design. The authors analyze how the structure of a package determines its developmental complexity according
to such measures as bug search times and documentation information
content. The work presents arguments for why these issues impact
solution cost and time more than does scalable performance. The nal
chapter explores the question of scalable execution and shows how
scalable design leads to scalable execution. The books focus is on
program organization, which has received considerable attention in
the broader software engineering community, where graphical description standards for modeling software structure and behavior have been
developed by computer scientists. These discussions might be enriched
by engineers and scientists who write scientic codes. This book aims to
bring such scientic programmers into discussion with computer scientists. The authors do so by introducing object-oriented software design
patterns in the context of scientic simulation.
Dr. Damian Rouson is the manager of the Reacting Flow Research
Department at Sandia National Laboratories. He was formerly Section
Head of Combustion Modeling and Scaling at the U.S. Naval Research
Laboratory. He was assistant professor of mechanical engineering at
the City University of New York, visiting assistant professor at the
University of Maryland, and Special Scientist at the University of
Cyprus. Damian Rouson received his Bachelors in Mechanical Engineering at Howard University and his Masters and PhD in Mechanical
Engineering from Stanford University.
Dr. Jim Xia is currently a software designer and tester at the IBM
Canada Lab in Markham, Ontario, Canada. He received his PhD in
Physics from the University of Western Ontario in 1997.
Dr. Xiaofeng Xu is currently an Analyst at General Motors Corp. in
Pontiac, Michigan. At this job, he performs airow and combustion
CFD analysis to support base engine designs. He received his PhD in
Mechanical Engineering (2003) from Iowa State University in Ames
and is the author or co-author of 39 referred publications.

Scientic Software Design

THE OBJECT-ORIENTED WAY
Damian Rouson
Sandia National Laboratories

Jim Xia
IBM Canada Lab in Markham

Xiaofeng Xu
General Motors Corp

CAMBRIDGE U NI VE RS I T Y PRE S S

Cambridge, New York, Melbourne, Madrid, Cape Town,

Singapore, So Paulo, Delhi, Tokyo, Mexico City
Cambridge University Press
32 Avenue of the Americas, New York, NY 10013-2473, USA
www.cambridge.org
Information on this title: www.cambridge.org/9780521888134
Damian Rouson, Jim Xia, and Xiaofeng Xu 2011
This publication is in copyright. Subject to statutory exception
and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without the written
permission of Cambridge University Press.
First published 2011
Printed in the United States of America
A catalog record for this publication is available from the British Library.
Library of Congress Cataloging in Publication data
Rouson, Damian, 1967
Scientic software design : the object-oriented way / Damian Rouson, Jim Xia, Xiaofeng Xu.
p. cm.
Includes bibliographical references and index.
ISBN 978-0-521-88813-4 (hardback)
1. ScienceData processing. 2. EngineeringData processing. 3. Software engineering.
I. Xia, Jim, 1970 II. Xu, Xiaofeng, 1972 III. Title.
Q183.9.R68 2011
005.1dc22
2011007351
ISBN 978-0-521-88813-4 Hardback
Additional resources for this publication at www.cambridge.org/rouson.
Cambridge University Press has no responsibility for the persistence or accuracy of URLs for external
or third-party Internet Web sites referred to in this publication and does not guarantee that any content
on such Web sites is, or will remain, accurate or appropriate.

For Zendo in exchange for the many books you have written for
me. For Leilee as a tangible product of every moment of support.
DAMIAN ROUSON

To my wife, Lucy, and two lovely children, Ginny and Alan, for
their encouragement and support.
JIM XIA

To my daughter Caroline, my wife Li and parents, for their love

and support.
XIAOFENG XU

Contents

List of Figures
List of Tables
Preface
Acknowledgments
Disclaimer

page xi
xv
xvii
xxi
xxii

PART I THE TAO OF SCIENTIFIC OOP

1 Development Costs and Complexity . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.1
1.2
1.3
1.4
1.5
1.6
1.7

Introduction
Conventional Scientic Programming Costs
Conventional Programming Complexity
Alternative Programming Paradigms
How Performance Informs Design
How Design Informs Performance
Elements of Style

2 The Object-Oriented Way . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.1
2.2
2.3
2.4
2.5
2.6
2.7
2.8

Nomenclature
Object-Oriented Analysis and Design
Encapsulation and Information Hiding
Wrapping Legacy Software
Composition, Aggregation, and Inheritance
Static and Dynamic Polymorphism
OOP Complexity
More on Style

3 Scientic OOP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.1 Abstract Data Type Calculus
3.2 Analysis-Driven Design
3.2.1 Design Metrics
3.2.2 Complexity Theory
3.2.3 Information Theory

3
7
11
19
24
25
27
31

31
33
36
40
48
53
54
55
57

57
66
69
71
73
vii

viii

3.3 Still More on Style

3.4 The Tao of SOOP

77
79

PART II SOOP TO NUTS AND BOLTS

83

4 Design Patterns Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

85

4.1 Essentials
4.2 Foundations
4.2.1 Building Architecture
4.2.2 Software Architecture
4.2.3 Scientic Software Architecture
4.3 Canonical Contexts
4.3.1 The Lorenz Equations: A Chaotic Dynamical System
4.3.2 Quantum Vortex Dynamics in a Superuid
4.3.3 Burgers Equation: Shock Formation and Dissipation

85
86
87
93
98
99
100
102
104

5 The Object Pattern . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

5.1 The Problem
5.2 The Solution
5.2.1 Fortran Implementation
5.2.2 C++ Style and Tools
5.2.3 C++ Implementation of Vortex
5.3 The Consequences
5.4 Related Patterns

107
108
110
116
122
127
127

6 The Abstract Calculus Pattern . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129

6.1 The Problem
6.2 The Solution
6.2.1 Fortran Implementation
6.2.2 C++ Implementation
6.3 The Consequences
6.4 Related Patterns

129
130
131
137
140
141

7 The Strategy and Surrogate Patterns . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143

7.1 The Problem
7.2 The Solution
7.2.1 Fortran Implementation
7.2.2 C++ Implementation
7.3 The Consequences
7.4 Related Patterns

143
144
146
155
164
164

8 The Puppeteer Pattern . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167

8.1 The Problem
8.2 The Solution
8.2.1 Fortran Implementation
8.2.2 A C++ Tool: 2D Allocatable Arrays
8.2.3 C++ Implementation

167
169
170
185
191

ix

8.3 The Consequences

8.4 Related Patterns

199
201

9 Factory Patterns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202

9.1 The Problem
9.2 The Solution
9.2.1 Fortran Implementation
9.2.2 C++ Implementation
9.3 The Consequences
9.4 Related Patterns
PART III GUMBO SOOP

202
203
205
216
226
227
229

10 Formal Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231

10.1
10.2
10.3
10.4

Why Be Formal?
Problem Statement
Side Effects in Abstract Calculus
Formal Specication
10.4.1 Modeling Arrays with OCL
10.4.2 Hermeticity
10.4.3 Economy
10.5 The Shell Pattern
10.6 An Assertion Utility for Fortran
10.7 Case Study: A Fluid Turbulence Solver

231
233
237
238
238
239
242
244
245
247

11 Mixed-Language Programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251

11.1 Automated Interoperability Tools
11.2 Manual Interoperability: C++/Fortran 95
11.3 Case Study: ForTrilinos and CTrilinos
11.3.1 C Interoperability in Fortran 2003
11.3.2 Method Invocation
11.3.3 Tabulation, Construction, and Destruction
11.3.4 Polymorphism and Hierarchy Mirroring
11.3.5 Discussion

251
254
259
261
269
273
278
281

12 Multiphysics Architectures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285

12.1 Toward a Scalable Abstract Calculus
12.1.1 Amdahls Law and Parallel Efciency
12.1.2 Automatic Parallelization: Compiler Technology
12.1.3 Directive-Based Parallelization: OpenMP
12.1.4 Library-Based Parallelization: ForTrilinos and MPI
12.1.5 Intrinsic Parallelization: Fortran 2008 Coarrays
12.2 Case Studies: Multiphysics Modeling
12.2.1 Quantum Turbulence in Superuid Liquid Helium
12.2.2 Lattice-Boltzman Biouid Dynamics

285
286
292
295
297
309
317
318
320

12.2.3 Particle Dispersion in Magnetohydrodynamics

12.2.4 Radar Scattering in the Atmospheric Boundary Layer
12.3 The Morfeus Framework

325
326
330

Appendix A Mathematical Background . . . . . . . . . . . . . . . . . . . . . . . . . . . 335

A.1 Interpolation
A.1.1 Lagrange Interpolation
A.2 Linear Solvers
A.2.1 Gaussian Elimination
A.2.2 LU Decomposition
A.3 Nonlinear Solvers
A.3.1 Newtons Method in 1D
A.3.2 Newtons Method in Multiple Dimensions
A.4 Partial Differential Equations
A.4.1 The Heat Equation
A.4.2 The Burgers Equation
A.5 Numerical Analysis
A.5.1 Finite Differences
A.5.2 Numerical Methods for Differential Equations

335
335
337
337
342
343
343
344
345
345
346
347
347
350

Appendix B Unied Modeling Language Elements . . . . . . . . . . . . . . . . . 357

B.1 Use Case Diagrams
B.2 Class Diagrams
B.2.1 Classes
B.2.2 Relationships
B.3 Object Diagrams
B.4 Sequence Diagrams
B.4.1 Interactions and Messages
B.4.2 Sequence Diagrams
B.5 The Object Constraint Language
B.5.1 The Context of OCL Expression
B.5.2 Initial Value Rule
B.5.3 Specifying Object Invariants
B.5.4 Adding Pre- and Postconditions to Operations
Bibliography
Index

357
359
359
362
366
367
368
368
370
370
371
371
371
373
379

Figures

1.1
1.2
1.3
1.4
1.5
2.1
2.2
2.3
2.4
2.5
2.6
2.7
2.8
2.9
2.10
2.11
2.12
2.13
2.14
2.15
3.1
3.2
3.3
3.4
3.5
3.6
4.1

Scientic software life cycle

Sequence of models in the conventional scientic code
development process
Flow chart for a 1D, transient heat transfer analysis of a
cooling n
Fin heat conduction program
A subset of the data dependencies for a temperature
calculation without argument intent
Spacecraft radio transmitter use case diagram
Astronaut class diagram
UML Array model
Hello, world! programs
Steps for modernizing legacy Fortran programs
Use case diagram for a n heat conduction analysis system
Object-oriented heat n analysis package class interfaces
Test-driven development
Problem denition class implementation
Numerical differentiation class implementation
Heat conductor class implementation
Conductor/Differentiator composite object diagram
Function-initialized conductor inheritance hierarchy
Function-initialized conductor implementation
Function-initialized conductor test
Integrable conductor test
Integrable conductor class implementation
Field class implementation
SOOP heat conduction class diagram
Sequence diagrams for the two quadrature schemes in
Figure 3.1
Bisection method bug search
Abstract class realization: Concrete Astronaut realizes, or
implements, an Abstract_speaker interface

page 4
8
14
17
18
33
34
35
37
40
42
43
44
46
47
48
49
50
51
52
61
64
66
67
68
72
95
xi

xii

4.2
4.3
4.4

Abstract class
Phase-space trajectory of the Lorenz equation system solution
Quantum vortex lament discretization and adaptive
remeshing
4.5 Quantum vortex laments before and after reconnection
4.6 Solutions of the Burgers equation with initial condition
u = 10 sin x and periodic boundary conditions
5.1 Memory leak example adapted from Stewart (2003)
5.2 OBJECT implementation: universal memory
management parent
5.3 Quantum vortex linked list with hermetic memory
management
5.4 Vortex creation and destruction
5.5 The concrete Vortex ADT implements the
Hermetic interface
5.6 Reference counting in C++
5.7 Output tools
5.8 Quantum vortex class denition
6.1 Lorenz system integration
6.2 Fortran implementation of ABSTRACT CALCULUS applied to a
Lorenz system
6.3 C++ implementation of an abstract calculus for the Lorenz
system
7.1 Fortran class model for applying the STRATEGY and
SURROGATE patterns to extend the Lorenz model
7.2 Fortran strategy pattern implementation for Lorenz
and timed Lorenz systems
7.3 C++ strategy implementation for Lorenz and timed Lorenz
systems
8.1 Associations in the Mediator pattern
8.2 Aggregation associations in the Puppeteer pattern
8.3 A UML sequence diagram for the Puppeteers Jacobian
contribution
8.4 Fortran implementation of an atmosphere puppeteer
8.5 Fortran 2003 implementation of puppets for an atmosphere
system
8.6 Fortran 2003 implementation of puppeteer pattern applied to
an atmosphere system
8.7 A mat_t class in C++ that stimulates the functionality of 2D
allocatable real arrays in Fortran
8.8 Atmosphere class denitions in C++
8.9 Puppet class denitions in C++
8.10 C++ implementation of puppeteer pattern applied to an
atmosphere system following the same design as that by
Fortran 2003
9.1 UML class diagram of an ABSTRACT FACTORY implementation
9.2 Fortran demonstration of an abstract factory

98
102
103
104
106
113
114
115
116
116
119
124
126
131
135
140
145
154
163
168
170
171
176
184
187
191
194
199

199
205
215

xiii

9.3
10.1
10.2
10.3
10.4
10.5
10.6
10.7
10.8
10.9
10.10
11.1
11.2
11.3
11.4
11.5
11.6
11.7

Abstract factory demonstration in C++

UML class diagram of a hermetic Burgers solver
Partial implementation of a Burgers solver class
Burgers equation
RK2 predictor step
HermeticField partial class diagram
UML class diagram of a hermetic Burgers solver
HermeticField and HermeticFieldPointer class association
Fortran assertions utility
Memory allocation assertions in a Fourier-spectral
turbulence code
Memory economy assertions in a Fourier-spectral
turbulence code
Chasm stub and skeleton interfaces for language A procedure
invoking language B procedure
A at interface exported to a shadow object in language B
by a real object in language A
Sample ForTrilinos-enabled application architecture
ForTrilinos utilities
ForTrilinos interoperability test
ForTrilinos MultiVector class excerpt
ForTrilinos OBJECT hierarchy, including the Hermetic

226
233
235
236
236
239
241
243
247

SURROGATE

274

11.8 Implementation of reference-counted ForTrillinos OBJECT

hierarchy
11.9 Epetra_Row Matrix class hierarchy
11.10 ForTrilinos Epetra_Vector wrapper excerpt
12.1 Periodic, 2nd-order central differences concrete eld class
12.2 Fork-join programming model
12.3 Scalability of a central difference Burgers solver using
OpenMP directives
12.4 Parallel efciency of a ForTrilinos/MPI-based central
difference Burgers solver
12.5 ForTrilinos-based Burgers solver
12.6 Class diagram of periodic_2nd_order class using
coarray grid eld
12.7 Denition of periodic 2nd-order central difference eld class
contains a coarray global eld
12.8 Denition of eld type that stores operation results from
periodic_2nd_order class
12.9 Quantum vortices and classical vortices in turbulent
superuid liquid helium
12.10 Quantum turbulence solver architecture
12.11 The lattice of D2Q9 model
12.12 The computational domain of the aneurysm model
12.13 Vorticity contours inside the sac of aneurysm under the
effects of different shapes of stent

249
250
253
255
260
266
266
272

277
278
284
290
295
297
298
309
311
316
317
318
320
321
323
323

xiv

12.14 Class diagram of LBE solver

12.15 Class diagram of a multicomponent LBE solver
12.16 Planar projection of inertial particles in liquid-metal MHD
turbulence under the inuence of an externally applied
magnetic eld orthogonal to the viewing plane
12.17 Semisolid MHD solver architecture
12.18 Geophysical coordinate system used for deriving the
governing equations
12.19 Instantaneous contours in a vertical slice of an Ekman layer
12.20 Atmospheric boundary layer class diagram
12.21 Abstract classes in the Morfeus framework
A.1 3D Lagrange interpolation in uniform grid
A.2 Pseudocode of Gaussian elimination with partial pivoting
A.3 Gaussian elimination procedure
A.4 Gaussian elimination in C++
A.5 Thomas algorithm procedure
A.6 One iteration of Newtons method
A.7 A nite-difference stencil at the point (x0 , y0 )
A.8 Linear Lagrange polynomials for the case in which the
f values at tn and tn+1 are used to derive a numerical
quadrature scheme
B.1 Use case diagram for a n heat conduction analysis system
B.2 Class diagram for astronaut class
B.3 Class dependencies in heat conduction class diagram
B.4 Class generalizations and realizations in SURROGATE and
STRATEGY patterns for a timed Lorenz system
B.5 Association relationships in ABSTRACT FACTORY patterns
applied to a Burgers solver
B.6 HermeticField and HermeticFieldPointer class association
and adornments
B.7 In the PUPPETEER pattern, puppeteer is an aggregation of air,
cloud, and ground
B.8 An object diagram that shows the object n of class Conductor
is composed of an object diff of class Differentiator
B.9 Interactions among objects during the computation of the
2nd-order spatial derivative for forward Euler method
B.10 Sequence diagram showing operations during the computation
of the 2nd-order spatial derivative for forward Euler method
B.11 Field class diagram of the turbulence solver of Rouson et al.
(2006)

324
324

326
327
328
329
330
332
336
339
341
342
343
344
348

353
358
360
362
364
365
365
366
367
369
370
372

Tables

2.1
3.1
4.1
10.1
11.1
11.2

OOP nomenclature
page 32
Translating blackboard abstractions into software abstractions 59
Design variables supported by patterns
96
Procedural run-time distribution
250
Corresponding C++ and Fortran 95 types
256
Fortran 2003 kind parameters and the corresponding
interoperable C types
263
11.3 Dummy array declaration interoperability
268
12.1 Runtime prole for central difference Burgers solver
procedures on an IBM smp system
290
B.1 Types of actions that are commonly used in an interaction
368

xv

Preface

This book is about software design. We use design here in the sense that it applies
to machines, electrical circuits, chemical plants, and buildings. At least two differences between designing scientic software and designing these other systems seem
apparent:
1. The extent to which one communicates a systems structure by representing
schematically its component parts and their interrelationships,
2. The extent to which such schematics can be analyzed to evaluate suitability and
prevent failures.
Schematic representations greatly simplify communications between developers.
Analyses of these schematics can potentially settle long-standing debates about which
systems will wear well over time as user requirements evolve and usage varies.
This book does not chiey concern high-performance computing. While most
current discussions of scientic programming focus on scalable performance, we
unabashedly set out along the nearly orthogonal axis of scalable design. We analyze
how the structure of a package determines its developmental complexity according
to such measures as bug search times and documentation information content. We
also present arguments for why these issues impact solution cost and time more than
does scalable performance.
We rmly believe that science is not a horse race. The greatest scientic impact
does not necessarily go to the swiftest code. Groundbreaking results often occur at the
intersection of multiple disciplines, where the basic science is so poorly understood
that important insights can be gleaned from creative simulations of modest size. In
multidisciplinary regimes, scaling up to hundreds of program units (e.g., procedures,
modules, components, or classes) matters at least as much as scaling up to hundreds
of execution units (e.g., processes, processors, threads, or cores). Put another way,
distributed development matters as much as distributed computing.
Nonetheless, recent trends suggest that exploiting even the most modest hardware including the laptops on which much of this book was written requires
parallelizing code across multiple execution units. After our journey down the path
toward scalable design, the nal chapter turns to the question of scalable execution,
or how to go from running on a single execution unit to running on many. In the
xvii

xviii

Preface

process, we demonstrate how scalable design illuminates a clear path to scalable

execution.
Program organization our main concern has received considerable attention
in the broader software engineering community, where computer scientists have
developed graphical description standards for modeling software structure and behavior. They have used these standards to articulate diagrammatic templates, or
design patterns, that solve recurring problems. They have also proposed metrics to
describe software properties. However, these discussions have taken place largely
outside the purview of the domain experts who write scientic code. This book
aims to bring scientic programmers into the discussion. We do so by introducing
object-oriented software design patterns in the context of scientic simulation.
We have used each of the presented patterns in production code across a range of
applications. Since our primary expertise lies in uid turbulence, that topic provides
a common thread, but the range of phenomena we have coupled to the uid motion
should interest those working in other elds. These phenomena include electromagnetic eld behavior, quantum effects, particle transport, and biomedical device
interaction. Each is the subject of a brief case study.
Two languages dominate scientic computing: Fortran and C++. A 2006 survey of the United States Department of Defense High Performance Computing
users indicated that approximately 85% write in C/C++/C#, while nearly 60% write
in Fortran.1 Clearly there is overlap between the two communities, so we present
examples in both languages.
Historically, patterns grew out of the adaptation of design principles from building architecture to object-oriented software architecture. Since the Fortran 2003
standard provides explicit support for all of the concepts central to object-oriented
programming (OOP), a unique opportunity has emerged to put Fortran and C++ on
equal footing. With this in mind, we intend the book to be accessible to at least three
audiences:
1. Fortran programmers who are unfamiliar with OOP,
2. Fortran programmers who emulate OOP using Fortran 90/95, and
3. C++ programmers who are familiar with OOP.
While we use Fortran 2003 to introduce object orientation, we incorporate notes that
compare Fortran constructs to similar C++ constructs. This serves the dual purpose of
introducing OOP in a context familiar to the Fortran programmers and introducing
Fortran in a context familiar to the C++ programmers.
Part I introduces the object-oriented way of programming in Fortran 2003 along
with various cost and complexity analyses of scientic programming. Chapter 1 couples computational complexity theory, Amdahls law, and the Pareto principle to
motivate the need to reduce data dependencies between lines of code and to balance scalability requirements against programmability. Chapter 2 details how the
object-oriented and functional programming paradigms reduce complexity by reducing data dependencies. Chapter 3 motivates the need for a general calculus for
abstract data types, describes the implementation of such a calculus, and further
extends our complexity analysis.
1 Simon Management Group, HPC Survey Summary, July 13, 2006.

Preface

Part II presents design patterns in Fortran 2003 and C++. Included in

Chapters 59 are two domain-specic patterns for multiphysics modeling and ve
general patterns, including one language-specic Fortran pattern.
Part III surveys advanced topics. Chapter 10 augments our design pattern
schematics with formal constraints that ensure desirable runtime behavior. The growing interest in integrating multiple software applications into multiphysics packages
inevitably leads to questions of multilanguage development. Chapter 11 addresses
multilanguage interoperability via standard language constructs as well as automated
tools for compile-time and runtime integration. Chapter 12 charts a path toward
runtime scalability and takes a preliminary jaunt down that path. Chapter 12 also
surveys the multiphysics software architectures from which we culled the design patterns in Part II. We conclude Chapter 12 with a discussion of how these patterns are
being incorporated into the Multiphysics Object-oriented Recongurable Fluid Environment for Unied Simulations (Morfeus) framework, an open-source software
development project funded by the Ofce of Naval Research.
Much of Part I was written for a course on scientic software design taught by
the rst author at the University of Cyprus during a visiting faculty appointment in
the autumn of 2006. Students in the course were beginning and advanced graduate
students. The material is also suitable for a senior-level undergraduate course. The
text assumes some introductory-level familiarity with partial differential equations,
numerical analysis, and linear algebra. The examples are chosen to keep the expected
experience with these subjects comparable to that of most senior-level engineering
and physical science students. For those lacking this exposure, Appendix A summarizes the requisite mathematical background. The text should also prove useful as a
resource for practicing computational scientists and engineers interested in learning
design principles for object-oriented software construction.
Damian Rouson
Berkeley, California
Jim Xia
Toronto, Ontario
Xiaofeng Xu
Detroit, Michigan

xix

Acknowledgments

This text is an outgrowth of research the authors have conducted with graduate
students, postdoctoral researchers, and various collaborators past and present. In
particular, Dr. Helgi Adalsteinsson contributed greatly to many of the C++ examples.
Drs. Karla Moris, Yi Xiong, Irene Moulitsas, Hari Radhakrishnan, Ioannis Sarris,
and Evangelos Akylas contributed to the research codes that inspired many of this
books Fortran examples.
Many examples draw from simulations we have performed in collaboration with
several domain experts in elds ranging from quantum physics to magnetohydrodynamics, re protection engineering, and atmospheric boundary layers. These
collaborators include Profs. Joel Koplik of the City University of New York, Andre
Marshall of the University of Maryland, Stavros Kassinos of the University of Cyprus,
and Robert Handler of the Texas A&M University.
Much of the work presented in this book was funded by the Ofce of Naval
Research (ONR) Automation in Ship Systems program under the direction of program manager Anthony Seman. The authors are also deeply indebted to Dr. Sameer
Shende of the University of Oregon for providing the development platform on
which the bulk of the examples in this text were written, and to Dr. Andrew McIlroy
of Sandia National Laboratories for approving the purchase of the platform on
which the remainder of the examples were developed. We thank Daniel Strong,
also of Sandia, for his artistic expression of this books theme in the cover art.
Sandia National Laboratories is a multi-program laboratory managed and operated
by Sandia Corporation, a wholly owned subsidiary of Lockheed Martin Corporation, for the U.S. Department of Energys National Nuclear Security Administration
under contract DE-AC04-94AL85000.

xxi

Disclaimer

The opinions expressed in this book are those of the authors and do not necessarily
represent those of International Business Machines Corporation or any of its
afliates.

xxii

PA R T I

THE TAO OF S C I EN T I F I C O O P

Development Costs and Complexity

Premature optimization is the root of all evil.

Donald Knuth

1.1 Introduction
The past several decades have witnessed impressive successes in the ability of
scientists and engineers to accurately simulate physical phenomena on computers. In
engineering, it would now be unimaginable to design complex devices such as aircraft
engines or skyscrapers without detailed numerical modeling playing an integral role.
In science, computation is now recognized as a third mode of inquiry, complementing theory and experiment. As the steady march of progress in individual spheres of
interest continues, the focus naturally turns toward leveraging efforts in previously
separate domains to advance ones own domain or in combining old disciplines into
new ones. Such work falls under the umbrella of multiphysics modeling.
Overcoming the physical, mathematical, and computational challenges of multiphysics modeling comprises one of the central challenges of 21st-century science
and engineering. In one of its three major ndings, the National Science Foundation
Blue Ribbon Panel on Simulation-Based Engineering Science (SBES) cited open
problems associated with multiscale and multi-physics modeling among a group of
formidable challenges [that] stand in the way of progress in SBES research. As
the juxtaposition of multiphysics and multiscale in the panels report implies,
multi-physics problems often involve dynamics across a broad range of lengths and
times.
At the level of the physics and mathematics, integrating the disparate dynamics
of multiple elds poses signicant challenges in simulation accuracy, consistency,
and stability. Modeling thermal radiation, combustion, and turbulence in compartment res, for example, requires capturing phenomena with length scales separated
by several orders of magnitude. The room-scale dynamics determine the issue of
paramount importance: safe paths for human egress. Determining such gross parameters while resolving neither the small length scales that control the chemical
kinetics nor the short time scales on which light propagates forms an active area of
research.
3

Development Costs and Complexity

Analysis
Design

Research &
Development

Construction

Testing &
Tuning

Figure 1.1. Scientic software life cycle.

At the level of the computation, the sheer size of multiscale, multiphysics simulations poses signicant challenges in resource utilization. Whereas the largest
length and time scales determine the spatial and temporal extent of the problem
domain, the shortest determine the required resolution. The ratio of the domain size
to the smallest resolvable features determines the memory requirements, whereas
the ratio of the time window to the shortest resolvable interval determines the computing time. In modeling-controlled nuclear fusion, for example, the time scales
that must be considered span more orders of magnitude than can be tracked simultaneously on any platform for the foreseeable future. Hence, much of the effort in
writing related programs goes into squeezing every bit of performance out of existing
hardware.
In between the physical/mathematical model development and the computation
lies the software development process. One can imagine the software progressing
through a life cycle much like any other product of human effort: starting with
research and development and ending with the testing and ne-tuning of the nal
product (see Figure 1.1). However, as soon as one searches for evidence of such
a life cycle in scientic computing, gaping holes appear. Research on the scientic software development process is rare. Numerous journals are devoted to novel
numerical models and the scientic insights obtained from such models, but only
a few journals focus on the software itself. Notable examples include the refereed
journals Scientic Programming and ACM Transactions on Mathematical Software.
Further along in the life cycle, the terrain becomes even more barren. Discussions
of scientic software design rarely extend beyond two categories: (1) explanations of
a particular programming paradigm, which fall into the implementation stage of the
life cycle, and (2) strategies for improving run-time efciency, which arguably impacts
all stages of the life cycle but comprises an isolated activity only once a prototype has
been constructed in the testing and tuning phase. Articles and texts on the rst topic
include presentations of structured and object-oriented programming, whereas those
on the second include discussions of parallel programming and high-performance
computing.
Implementation-independent design issues such as the chosen modular decomposition and its developmental complexity have received scant attention in the
scientic computing community. Not surprisingly then, quantitative analyses of
software designs have been limited to run-time metrics such as speed and parallel efciency. In presenting attempts to analyze static source code organization, the
author often encounters a perception that program structure is purely stylistic of

1.1 Introduction

the You say toe-may-toe; I say toe-mah-toe variety. Even to the extent it is recognized that program structure matters, no consensus is emerging on what structural
precepts prove best.
That scientic program design warrants greater attention has been noted at
high levels. The 1999 Presidential Information Technology Advisory Committee
(PITAC) summarized the situation for developers of multidisciplinary scientic
software:
Today it is altogether too difcult to develop computational science software and applications. Environments and toolkits are inadequate to meet the needs of software developers
in addressing increasingly complex interdisciplinary problems... In addition, since there
is no consistency in software engineering best practices, many of the new applications are
not robust and cannot easily be ported to new hardware.

Fortunately, the situation with regards to environments and toolkits has improved
since the PITAC assessments. Using tools such as Common Component Architecture
(CCA), for example, one can now link applications written by different developers in
different languages using different programming paradigms into a seamless, scalable
package without incurring the language interoperability and communication latency
issues associated with non-scientic toolkits that facilitate similar linkages. Several
groups have constructed development frameworks on top of CCA to facilitate distributed computing (Zhang et al. 2004), metacomputing (Malawski et al. 2005), and
other computing models.
However, the situation with regard to software engineering best practices has
seen far less progress. Most authors addressing scientic programming offer brief
opinions to aid robustness before retreating to the comfortable territory of run-time
efciency or numerical accuracy. Without a healthy debate in the literature, it is
unsurprising that no consensus has emerged. By contrast, in the broader software
engineering community, consensus has been building for over a decade on the best
practices at least within one development paradigm: object-oriented design (OOD).
In this context, the best practices have been codied as design patterns, which are
cataloged solutions to problems that recur across many projects.
Experience indicates many scientic programmers nd the term design pattern
vague or awkward on rst hearing. However, its application to software development conforms with the rst four denitions of pattern in the Merriam Webster
dictionary:
1.
2.
3.
4.

a form or model proposed for imitation,

something designed or used as a model for making things,
an artistic, musical, literary, or mechanical design or form,
a natural or chance conguration.

Having proved useful in the past, design patterns are offered as models to be followed
in future work. The models have artistic value in the sense of being elegant and in
the sense that software engineering, like other engineering elds, is part science and
part art. And much like mechanical design patterns, some software design patterns
mimic patterns that have evolved by chance in nature.
Gamma et al. (1995) rst collected and articulated object-oriented software
design patterns in 1995 in their seminal text Design Patterns: Elements of Reusable

Development Costs and Complexity

Object-Oriented Software. They drew inspiration from the 1977 text, A Pattern Language, by architects Alexander et al., who in turn found inspiration in the beauty
medieval architects achieved by conforming to local regulations that required specic building features without mandating specic implementations of those features.
This freedom to compose variations on a theme facilitated individual expression
within the connes of proven forms.
Gamma et al. did not present any domain-specic patterns, although their books
introduction suggests it would be worthwhile for someone to catalog such patterns.
Into this gap we thrust the current text. Whereas Part I of this text lays the objectoriented foundation for discussing design patterns and Part III discusses several
related advanced topics, Part II aims to:
1. Catalog general and domain-specic patterns for multiphysics modeling,
2. Quantify how these patterns reduce complexity,
3. Present Fortran 2003 and C++ implementations of each pattern discussed.
Each of these objectives makes a unique contribution to the scientic computing eld. The authors know of only a handful of publications on
object-oriented software design patterns for scientic software (Blilie 2002;
Decyk and Gardner 2006; Decyk and Gardner 2007; Gardner and Manduchi 2007;
Markus 2006; Rouson et al. 2010) and even fewer quantitative analyses of how patterns reduce scientic programming complexity (Allan et al. 2008; Rouson 2008).
Finally, published design pattern examples that take full advantage of the new objectoriented constructs of Fortran 2003 have only just begun to appear (Markus 2008;
Rouson et al. 2010).
That Fortran implementations of object-oriented design patterns have lagged
those in other languages so greatly is both ironic and understandable. It is ironic
because object-oriented programming started its life nearly four decades ago with
Simula 67, a language designed for physical system simulation, which is the primary
use of Fortran. It is understandable because the Fortran standards committee, under
heavy inuence by compiler vendors, moved at glacial speeds to provide explicit support for object-orientation. (See Metcalf et al. 2004 for some history on this decision.)
In the absence of object-oriented language constructs, a small cadre of researchers
began developing techniques for emulating object-orientation in Fortran 90/95 in the
1990s (Decyk et al. 1997a, 1997b, 1998; Machiels and Deville 1997), culminating in
the publication of the rst text on OOP in Fortran 90/95 by Akin (2003).
The contemporaneous publication of the Fortran 2003 standard, which provides
object-oriented language constructs, makes the time ripe for moving beyond the
basic mechanics to higher-level discussions of objected-oriented design patterns for
scientic computing. Because the availability of Fortran 2003 motivates this book,
Fortran serves as the primary language for Part I of this text. Because most C++
programmers already know OOP, there is less need to provide C++ in this part
of the text. Some of the complexity arguments in Part I, however, are languageindependent, so we provide C++ translations to make this section accessible to C++
programmers. In Part II, greater effort is made to write C++ examples that stand on
their own as opposed to being translations of Fortran.

1.2 Conventional Scientic Programming Costs

A Pressing Need for 21st-Century Science

The 20th-century successes in simulating individual physical phenomena can be
leveraged to simulate multiscale, multiphysics phenomena much more simply
if we begin distilling, analyzing, and codifying the best practices in scientic
programming.

1.2 Conventional Scientic Programming Costs

The word costs plays two roles in this sections title. As a noun, it indicates that the
section discusses the costs associated with writing and running conventional scientic programs. As a verb, it indicates that the conventional approach to writing
these programs costs projects in ways that can be alleviated by other programming paradigms. Whereas this section estimates those costs, section (1.4) presents
alternative paradigms.
The software of interest is primarily written for research and development in the
sciences and engineering sciences. Although many such codes later form the core
of commercial packages used in routine design and analysis, all are grouped under
the heading scientic in this text. During the early phases of the development,
many scientic programming projects tackle problems that are sufciently unique or
demanding to preclude the use of commercial off-the-shelf software for the entire
project. Although some commercial library routines might be called, these serve supporting roles such as solving linear and nonlinear systems or performing data analysis.
Viewed from a client/server perspective, the libraries provide general mathematical
services for more application-specic client code that molds the science of interest
into a form amenable to analysis with the chosen library.
Some mathematical libraries are portable and general-purpose, such as those by
the Numerical Algorithms Group (NAG) or the Numerical Recipes series. Others
are tuned to exploit the features of specic hardware. The Intel Math Kernel Library
(MKL) targets Intel processors and scales up to large numbers of processors on
distributed-memory clusters, whereas the Silicon Graphics, Inc. (SGI) math library
scales on distributed shared-memory supercomputers with SGIs proprietary highbandwidth memory architecture.
Mathematical libraries attempt to embody the best available numerical algorithms without imposing any particular program organization. Responsibility for
program organization, or architecture, thus rests with the developers. Although the
denition of software architecture remains a subject of active debate, there is longstanding agreement on the attractiveness of modular design, so for present purposes,
architecture denotes a specic modular decomposition. This decomposition would
include both the roles of individual modules and the relationships between them. To
avoid language- and paradigm-specicity, module here denotes any isolated piece
of code whether said isolation is achieved by separation into its own le or by some
other construct.
Most scientic programmers are trained in engineering or science. Typical curricula in these elds include only an introductory programming course that teaches
the use of a particular programming language and paradigm. Teaching only the

Development Costs and Complexity

Reality

Physical
Model

Accuracy

Mathematical
Model
(continuous)

Numerical
Model
(semidiscrete)

Numerical
Model
(fully discrete)

Code
(discrete space-time
physics)

Visualization

Statistical
Analysis

Perception

Figure 1.2. Sequence of models in the conventional scientic code development process.

implementation phase differs considerably from the rest of the curriculum, wherein
students are encouraged to think abstractly about systems and to quantitatively evaluate competing models for describing those systems. In the automatic controls courses
taken by many engineering students, for example, model elements are delineated
precisely and their roles and relationships are specied mathematically, so the system
behavior can be analyzed rigorously.
Figure 1.2 situates programming within the conventional modeling process. One
starts off wishing to model physical reality. Based on observations of that reality,
one constructs a physical model, which for present purposes comprises a conceptual abstraction of reality. Both model and abstraction here connote useful
simplications that retain only those elements necessary for a given context. For
example, one might abstract a cooling n on a microprocessor chip as a thin solid
with constant properties transporting thermal energy in the direction orthogonal to
the chip. One might further assume the energy transfer occurs across a distance that
is large relative to the materials intermolecular spacing, so it can be considered a
continuum.
Although physical models are typically stated informally, Collins (2004) recently
proposed the development of a physics markup language (PML) based on the extensible markup language (XML). Formal specications in such a markup language
would complement the programming techniques presented in this book. Specically,
automatic code generation from a markup document might produce software counterparts of algebraic and differential operators specied in PML. Chapter 3 discusses
the implementation of such operators.

1.2 Conventional Scientic Programming Costs

The rst formal step in conventional scientic work involves expressing the physical model in a mathematical model with continuous variation of the dependent
variables in space and time. For the n problem, this leads to the one-dimensional
(1D), unsteady heat equation:
T
2T
= 2 ,  T = (0, Ln ) (0, tnal )
t
x
T(0, t) = Tchip

(1.1)

T(Ln , t) = Tair
T(x, 0) = Tair
where T(x, t) is the temperature at time t a distance x from the chip, is the ns
thermal diffusivity,  T is the space-time domain, Ln is the n length, tnal is the
nal time of the simulation, and where Tchip and Tair are boundary conditions.
Solving the mathematical model on a computer requires rendering
equations (1.1) discrete. Most numerical schemes employ the semidiscrete method,
which involves discretizing space rst and time second. Most spatial discretizations
require laying a grid over the spatial domain. Given a uniformly spaced grid overlaid
on , applying the central difference formula to the right-hand side of (1.1) leads to
a set of coupled ordinary differential equations for the nodal temperatures:


1
dT

dt
x2

1
2
..
.

Tchip


T +
.
2

x
..

2
Tair

1
..
.
1

(1.2)

where x Ln /N is the uniform spacing in between the N +1 grid points covering .

One can derive the same discrete system of equations via a nite volume scheme or a
Galerkin nite element method with linear-basis functions. We refer to the intervals
between nodes generically as elements and assume the reader is familiar with one or
another path that starts from equations (1.1) and leads to equations (1.2).
The semidiscrete model is rendered fully discrete by applying a time integration
algorithm to equations (1.2). The simplest case is explicit Euler, which corresponds
to a two-term Taylor series expansion:




 n+1 = T
 n + t dT  = T
 n + t AT
n +b
T
dt tn

(1.3)

 n+1 are vectors of N 1 internal grid point temperatures at time

 n and T
where T
stepstn nt and tn+1 tn + t, respectively.
Most scientic computing publications focus on the previously described modeling steps, treating their implementation in code as an afterthought, if at all. This
has numerous adverse consequences. First, several very different concepts are hardwired into equation (1.3) in ways that can be hard to separate should one stage in
the modeling process require reconsideration. Equation (1.3) represents a specic
numerical integration of a specic discrete algorithm for solving a specic partial
differential equation (PDE). Without considerable forethought, the straightforward

10

Development Costs and Complexity

expression of this algorithm in software could force major code revisions if one later
decides to change the discretization scheme, for example, the basis functions or the
PDE itself. The next section explains why the conventional approach to separating
these issues scales poorly as the code size grows.
A side effect of conating logically separate modeling steps is that it becomes
difcult to assign responsibility for erroneous results. For example, instabilities that
exist in the fully discrete numerical model could conceivably have existed in continuous mathematical model. Rouson et al. (2008b) provided a more subtle example in
which information that appeared to have been lost during the discretization process
was actually missing from the original PDE.
As previously mentioned, the code writing is the rst step with no notion of
abstraction. Except for ow charts which, experience indicates, very few programmers construct in practice most discussions of scientic programs offer no
formal description of program architecture other than the code itself. Not surprisingly then, scientic program architecture is largely ad hoc and individualistic. Were
the problems being solved equally individualistic, no difculty would arise. However, signicant commonalities exist in the physical, continuous mathematical, and
discrete numerical models employed across a broad range of applications. Without
a language for expressing common architectural design patterns, there can be little
dialogue on their merits. Chapter 2 employs just such a language: the Unied Modeling Language (UML). Part II presents design patterns that exploit commonalities
among a broad range of problems to generate exible, reusable program modules.
Lacking such capabilities, conventional development involves the reinvention of
architectural forms from scratch.
How does redevelopment inuence costs? Lets think rst in the time domain:
Total solution time = development time + computing time

(1.4)

Until computers develop the clairvoyance to run code before it is written, programming will always precede execution, irrespective of the fact that earlier versions of
the program might be running while newer versions are under development. Since
the development time often exceeds the computing time, the fraction of the total
solution time that can be reduced by tuning the code is limited. This is a special case
of Amdahls law, which we revisit in Section 1.5 and Chapter 12.
Time is money, so equation (1.4) can be readily transformed into monetary
form by assuming the development and computing costs are proportional to the
development and computing times as follows:
$solution = $development + $computing
= Ndev pavg tdev +

$computer trun
Nusers tuseful

(1.5)

where the $ values are costs; Ndev , pavg , and tdev are the number of developers, their
average pay rate, and the development time, respectively; and Nusers , trun , and tuseful
are the number of computer users, the computing time, and the computers useful
life, respectively. In the last term of equation (1.5), the rst factor estimates the
fraction of the computers resources available for a given users project. The second

1.3 Conventional Programming Complexity

factor estimates the fraction of the computers useful life dedicated to that users
runs.
There are of course many ways equation (1.5) could be rened, but it sufces
for current purposes. For a conservative estimate, consider the case of a graduate
student researcher receiving a total compensation including tuition, fees, stipend,
and benets of $50,000/yr for a one-year project. Even if this student is the sole
developer of her code and has access to a $150,000 computer cluster with a useful
life of three years, she typically shares the computer with, say, four other users.
Therefore, even if her simulations run for half a year on the wall clock, the solution
cost breakdown is
$solution = (1 developer)($50, 000/year)(1 year)

(1.6)

$150, 000 0.5 user years

5 users
3 years

(1.7)

= $50, 000 + $5, 000

(1.8)

so the development costs greatly exceed the computing costs. The fraction of costs
attributable to computing decreases even further if the developer receives higher
pay or works for a longer period likewise if the computer costs less or has a longer
useful life. By contrast, if the number of users decreases, the run-time on the wall
clock decreases proportionately, so the cost breakdown is relatively insensitive to
the resource allocation. Similar, the breakdown is insensitive to adding developers
if the development time is inversely proportional to the team size (up to some point
of diminishing returns).
For commercial codes, the situation is a bit different but the bottom line is
arguably the same. The use and development of commercial codes occur in parallel, so the process speedup argument breaks down. Nonetheless, development time
savings lead to the faster release of new features, providing customer benet and a
competitive advantage for the company.
This forms the central thesis on which this book rests:
Your Time Is Worth More Than Your Computers Time
Total solution time and cost can be reduced in greater proportion by reducing
development time and costs than by reducing computing time and costs.

1.3 Conventional Programming Complexity

Part of the reason scientic computing has traditionally focused on the earlier stages
of Figure 1.2 is that the accuracy losses at these stages are well known and rigorously
quantiable. As indicated by the gures vertical axis, each transition in the discretization process has an associated error. For example, the truncation error, e , of
the spatial discretization inherent in equation (1.2) is bounded by a term proportional
to the grid spacing:
(1.9)
e Cx3
where C typically depends on properties of the exact solution and its derivatives.

11

12

Development Costs and Complexity

Program organization may seem orthogonal to accuracy hence the horizontal

line leading to the code in Figure 1.2. Complexity arguments to be presented in
this section, however, lay the foundation for an OOP strategy in which memory
management plays a critical role. Recalling that x = Ln /N leads directly to the
fact that the number of points, N, determines the error:

eC

Ln
N

3
(1.10)

Furthermore, the maximum number of points is set by the available memory, M, and
the chosen precision, p, according to:
Nmax =

M
p

(1.11)

where typical units for M and p would be bytes and bytes per nodal value, respectively. Substituting (1.11) into (1.10) leads to the lowest possible upper bound on the
error:


pLn 3
(1.12)
sup e C
M
Whether a given program actually achieves this level of accuracy is determined by
its dynamic memory utilization. Consider two ways of computing the right-hand side
(RHS) of equation (1.3):



 n+1 = T
 n + t AT
n+b
T
(1.13)

n+b
= (I + tA) T

(1.14)

where I is the identity matrix. Working outward from inside the parenthesis, evalun ,
 n intact while computing AT
ating the RHS in equation (1.13) requires keeping T
n
n



so a time exists when three vectors are stored in memory (T , AT , and b). Adding
this storage requirement to the additional xed storage of A, t, and other values reduces the achievable N below the maximum allowed by equation (1.11) and thereby
increases the error in (1.12).
 n with the result
By contrast, evaluating equation (1.14) allows for overwriting T
n

of the product (I + tA) T , so at most two vectors need be stored at any given time.
This facilitates increasing N closer to Nmax , thereby decreasing the error. Thus,
to the extent complexity arguments lead to a strategy where memory management
plays a critical role, programming complexity plays a critical and quantiable role
in the accuracy of the results. Two programs that are algorithmically equivalent in
terms of the numerical methods employed might have different levels of error if one
program requires signicantly more dynamic memory. Chapter 12 presents much
more severe examples of this principle in which the solutions are three-dimensional
(3D) and therefore place much greater demands on the available memory.
The other way that complexity affects accuracy is by increasing the likelihood
of bugs. A statistical uncertainty argument could be constructed to quantify the inuence of bugs on accuracy, but it seems more useful to return to the theme of the
previous section by asking, How can we quantify complexity in ways that reduce
development time? Several approaches of varying sophistication are employed

1.3 Conventional Programming Complexity

throughout this book. For present purposes, lets break the development time into
the four stages:
1.
2.
3.
4.

Development of the physical, mathematical, and numerical models,

Designing the program,
Writing the code,
Debugging the code.

Because this book focuses on the design stage, one wonders how program design
inuences the debugging process. Such considerations could easily lead us down a
path toward analyzing the psychology and sociology of programmers and groups of
programmers. These paths have been trod by other authors, but we instead turn to
simple truisms about which we can reason quantitatively.
In their text Design Patterns Explained, Shalloway and Trott (2002) write:
The overwhelming amount of time spent in maintenance and debugging is on nding bugs
and taking the time to avoid unwanted side effects. The actual x is relatively short!

Likewise, Oliveira and Stewart (2006), in their text Writing Scientic Software: A
Guide to Style, write: The hardest part of debugging is nding where the bug is.
Accepting these arguments, one expects debugging time to increase monotonically
with the number of program lines, , one must read to nd a given bug:
tdebugging = f (), f (1 ) > f (2 ) if 1 > 2

(1.15)

in which case reducing the number of lines one must read reduces debugging time.
This relationship argues not so much for reducing the absolute number of program
lines but for reducing the dependencies between lines. Bugs typically present themselves as erroneous output, whether from the program itself or the operating system.
Exterminating bugs requires tracing backward from the line where the symptom
occurs through the lines on which the offending line depends.
To understand how dependencies arise, it helps to examine real code. Since this
book focuses on design, we rst construct the only design document one typically
encounters for conventional scientic code. Figure 1.3 depicts a ow chart that uses
the cooling n numerical model from section 1.2 to estimate the time required for
the n to reach steady state. At long times, one expects the solution to reach a steady
state characterized by T/t = 0 = 2 T/x2 , so the temperature varies linearly in
space but ceases changing in time. The algorithm in Figure 1.3 senses this condition
by checking whether the second derivative of the n temperature adjacent to the
chip drops below some tolerance.
Additional tertiary dependencies might be generated if any procedures listed
previously call other procedures and so forth. Dissecting the temperature calculation
according to the values it uses makes clear that shared data mediates the interactions
between program lines.
Figure 1.4 provides a sample Fortran implementation of the algorithm described
in Figure 1.3. Section 1.7 discusses several stylistic points regarding the sample code.
In the current section, the emphasis remains on dependencies between lines. For
discussion purposes, imagine that during the course of solution, program line 29 in
main (Figure 1.4(c)) prints temperatures outside the interval (Tair , Tchip ), which is

13

14

Development Costs and Complexity

Start

Input physical parameters:

, Lfin, Tchip, Tair

Input numerical parameters:

t,N

Initialize T and central

difference coefficients

Yes

t < tstable?
No
t = tstable

Calculate 2T/x2

T=T + (t)2T/x2

2T/x2 < tolerance?

No
Yes
End

Figure 1.3. Flow chart for a 1D, transient heat transfer analysis of a cooling n.

known to be physically impossible. Dissecting line 28 in main from left to right, one
nds the temperature depends directly on the following:
1. The time step read in read_numerics() at line 18 in main,
2. The second derivative T_xx calculated by differentiate() at line 27 of main,
and
3. The diffusivity initialized on line 16 of main and possibly overwritten on line 24
of main.
Furthermore, dissecting line 27 of main from left to right generates the following
secondary dependencies of T on:
4. The laplacian differencing matrix created at line 22 of main,

1.3 Conventional Programming Complexity

5. Any modications made to T by differentiate() (which can be determined to

be none after noticing the procedures pure attribute, which in turn necessitates
the intent(in) attribute on line 54 in Figure 1.4(b)),
6. The boundary conditions air_temp and chip_temp, which are initialized at line
9 in main and possibly overwritten by read_physics() on line 16 in main.
Figure 1.5 diagrams a sample interaction for a code similar to that of Figure 1.4
but without dummy argument intent specications. In programs large enough to be
broken into multiple procedures, the mechanisms for sharing data include argument
passing and le access. A certain number of data dependencies across procedures
are inevitable in modular code. In the worst-case scenario, however, the global
sharing facilitated by the Fortran common blocks in many scientic programs implies
erroneous output from any one program line could send us on a wild-goose chase
through all n 1 other lines.
With a bit of empirical data and straightforward reasoning, the previously described scenario leads to an estimate of conventional programming debugging time.
First consider that an exhaustive study of a large commercial software project by
Fenton and Ohlsson (2000) found that module defect density, r, was roughly constant across virtually the entire range of module sizes comprising the software.
Contrary to conventional wisdom, this implies the number of defects in a modular program does not depend on the module granularity. In a collection of modules
comprised of lines total, the number of defects has an expected value of r.

1
2
3
4
5
6
7
8
9
10

1
2
3
4
5
6
7
8
9
10
11
12
13
14

Figure 1.4(a)
module kind_parameters
! Type kind parameter
implicit none
private
public :: rkind,ikind,ckind
integer ,parameter :: digits=8 ! num. digits of kind
integer ,parameter :: decades=9 ! num. representable decades
integer ,parameter :: rkind = selected_real_kind(digits)
integer ,parameter :: ikind = selected_int_kind(decades)
integer ,parameter :: ckind = selected_char_kind(default)
end module
Figure 1.4(b)
module conduction_module
use kind_parameters ,only : rkind,ikind
implicit none
! Tri-diagonal array positions:
integer(ikind) ,parameter::diagonals=3,low_diag=1,diag=2,up_diag=3
contains
pure logical function found(namelist_io)
integer(ikind) ,intent(in) ::namelist_io
if (namelist_io==0) then
found=.true.
else
found=.false.
end if
end function

15

16

Development Costs and Complexity

15
16
17
18
19
20
21
22

logical function read_physics(unit,alpha,L_fin,T_chip,T_air)

real(rkind)
,intent(out) :: alpha,L_fin,T_chip,T_air
integer(ikind) ,intent(in) :: unit
integer(ikind)
:: physics_io
namelist/physics/ alpha,L_fin,T_chip,T_air
read(unit,nml=physics,iostat=physics_io)
read_physics = found(physics_io)
end function

23
24
25
26
27
28
29
30
31
32
33
34
35

logical function read_numerics(unit,dt,nodes,dx,L_fin)

real(rkind)
,intent(out) :: dt,dx
integer(ikind) ,intent(out) :: nodes
real(rkind)
,intent(in) :: L_fin
integer(ikind) ,intent(in) :: unit
integer(ikind)
:: numerics_io,elements
namelist/numerics/ dt,nodes
read(unit,nml=numerics,iostat=numerics_io)
read_numerics = found(numerics_io)
elements = nodes+1
dx = L_fin/elements
end function

36
37
38
39
40
41

pure real(rkind) function stable_dt(dx,alpha)

real(rkind) ,intent(in) :: dx,alpha
real(rkind) ,parameter :: safety_factor=0.9
stable_dt = safety_factor*(dx**2/alpha)
end function

42
43
44
45
46
47
48
49
50
51

pure function differencing_stencil(dx,nodes) result(centralDiff)

real(rkind)
,intent(in) :: dx
integer(ikind) ,intent(in) :: nodes
real(rkind),dimension(:,:),allocatable :: centralDiff
allocate(centralDiff(nodes,diagonals))
centralDiff(:,low_diag)= 1./dx**2
centralDiff(:,
diag)= -2./dx**2
centralDiff(:, up_diag)= 1./dx**2
end function

52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69

pure function differentiate(finiteDiff,T,T1st,Tlast)result(T_xx)

real(rkind),dimension(:) ,intent(in) ::T
real(rkind),dimension(:,:),intent(in) ::finiteDiff!differentiation op
real(rkind)
,intent(in) :: T1st,Tlast
real(rkind),dimension(:) ,allocatable:: T_xx
integer(ikind) :: nodes,i
nodes = size(T)
allocate(T_xx(nodes))
T_xx(1) = finiteDiff(1,low_diag)*T1st &
+finiteDiff(1,
diag)*T(1) &
+finiteDiff(1, up_diag)*T(2)
forall(i=2:nodes-1)
T_xx(i) = finiteDiff(i,low_diag)*T(i-1) &
+finiteDiff(i,
diag)*T(i ) &
+finiteDiff(i, up_diag)*T(i+1)
end forall
T_xx(nodes) = finiteDiff(nodes,low_diag)*T(nodes-1) &

1.3 Conventional Programming Complexity

70
71
72
73

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33

1
2
3
4
5
6

end function
end module

+finiteDiff(nodes,
diag)*T(nodes
+finiteDiff(nodes, up_diag)*Tlast

) &

Figure 1.4(c)
program main
use iso_fortran_env ,only : input_unit
use kind_parameters ,only : rkind,ikind
use conduction_module
implicit none
real(rkind)
,parameter :: tolerance=1.0E-06
integer(ikind) ,parameter :: chip_end=1
integer(ikind)
:: elements=4,intervals
real(rkind) :: air_temp=1.,chip_temp=1. & !boundary temperatures
,diffusivity=1.,fin_length=1.,time=0. & !physical parameters
,dt=0.01,dx
!time step, grid spacing
real(rkind),dimension(:) ,allocatable:: T,T_xx !temperature,2nd deriv
real(rkind),dimension(:,:),allocatable:: laplacian !differential op
dx=fin_length/elements
!default element size
if (.not. &
read_physics(input_unit,diffusivity,fin_length,chip_temp,air_temp))&
print *,In main: Using default physics specification.
if (.not. read_numerics(input_unit,dt,elements,dx,fin_length)) &
print *,In main: Using default numerical specification.
allocate(T(elements),laplacian(elements,diagonals),T_xx(elements))
T
= air_temp
laplacian = differencing_stencil(dx,elements)
print *,time,temperature=,time,T
dt = min(dt,stable_dt(dx,diffusivity))
do
time = time + dt
T_xx = differentiate(laplacian,T,chip_temp,air_temp)
T = T + dt*diffusivity*T_xx
print *,time,temperature=,time,T
if (T_xx(chip_end)<tolerance) exit
end do
print *,steady state reached at time ,time
end program
Figure 1.4(d)
&physics
alpha=1.,L_fin=1.,T_chip=1.,T_air=1.
/
&numerics
dt=0.01,nodes=9
/
Figure 1.4. Fin heat conduction program: (a) global constants, (b) solver, (c) driver, and (d)
input le.

Assuming programmers are conscientious, a single symptom of a problem is

sufcient to start the search for the bug that caused it, so there exists a one-to-one
relationship between bugs and symptoms. Now imagine a chronological listing of

17

18

Development Costs and Complexity

program main
real :: T(10),alpha=1.,dt=0.1,dx=0.01
...
T_xx= differentiate(laplacian,T,chip_temp,air_temp)
T = T + dt*t_xx/diffusivity
function differentiate(...,T,...)result(T_xx)
real(rkind), dimension(:) :: T
...
T_xx(i) = finiteDiff(i,low_diag)*T(i1)&
+finiteDiff(i,
diag)*T(i )&
+finiteDiff(i, up_diag)*T(i+1)
?

T(0),T(1),...,T(10)

Legend
Read
Write

Figure 1.5. A subset of the data dependencies for a temperature calculation without argument
intent. Without intent, the entire differentiate() procedure must be reviewed to determine
whether the procedure writes to T.

the program lines executed before the symptom occurs. Assume redundant occurrences are removed while ensuring that all lines that execute before the symptom
occurs are listed prior to listing the symptomatic line. Some of the lines preceding
the symptomatic line in such a listing might appear after the symptomatic line in the
source code. For example, if the symptom occurs early in the second pass through a
loop, all lines executed during the rst pass through the loop would be listed prior
to the symptomatic line.
The aforementioned assumptions are intended to avoid inating above the
total number of lines in the code and to ensure that causality connes the bug to the
lines preceding the symptom on the list. Now if a given symptom is equally likely to
occur anywhere on the list, then its expected location is /2 lines from the top, so
the bug must be in the preceding /2 1 lines.
If the bug is equally likely to be found on any of these lines, the expected value
of the number of lines that must be searched backward to nd it is (/2 1)/2. For
code with r bugs, the search algorithm just described has an expected completion
time of:


r
(1.16)
1 tline
tnd =
2 2
where tline is the average time to review one line. We dene review to include
reading it, testing it, and thinking about it over coffee. Since we do not include
the time to revise it, which might also include revising other lines, equation (1.16)
provides a lower bound on the debugging time, which provides a lower bound on the
development time. Development time thus grows at least quadratically with code size.
To estimate r, consider that Hatton (1997) found an average of 8 statically
detectable faults per 1,000 lines among millions of lines of commercially released
scientic C code and 12 per 1,000 in Fortran 77 code. Separately, he apparently
estimated that approximately two to three times as many errors occur in C++1 . He
1 Van Snyder reported Hattons C++ results in a smartly and humorously written response to a petition

to retire Fortran. See http://www.fortranstatement.com/Site/responses.html

1.4 Alternative Programming Paradigms

dened a fault as a misuse of the language which will very likely fail in some context. Examples include interface inconsistencies and using un-initialized variables.
Hattons average fault rate yields the estimate r 0.01.
Equation (1.16) provides a polynomial time estimate for completing the debugging process. Even though polynomial time estimates fall under the purview of computational complexity theory, one must be careful to distinguish the complexity of the
debugging process from the complexity of the code itself. Attempts to quantify code
complexity in useful ways particularly to predict fault rates have a checkered past.
Exacerbating the problem is the reliance of most programs on externally dened
libraries with which data must be shared. Truly pernicious bugs (which the vendors
might call features) result when code not written by the programmer modies
its arguments unexpectedly. Although such dependencies may seem inevitable,
Chapter 6 explains how to construct a library that never gets its grubby hands on
the data on which it operates.
As individual programs merge into complete software packages, Figure 1.5
becomes a plate of spaghetti, providing one interpretation of the infamous term
spaghetti code. The next section explores several philosophies for disentangling
the knots in spaghetti code. The most popular modern alternative, OOP, reduces
data dependencies by constructing modules with local data and allowing access to
that data only to procedures inside the associated module. Chapter 2 renes the
estimate in equation (1.16) by analyzing a more representative fault localization
algorithm and demonstrates how OOP reduces the search time. Chapter 3 demonstrates how an OOP strategy that is broadly useful in scientic programming further
limits search times.
The benets of the above modularization strategy are manifold, but the
strategy does have limits. Although strong evidence suggests defect rates are
roughly independent of module size as assumed implicitly in equation (1.16)
(Fenton and Ohlsson 2000), the same study provided evidence that defect rates
sometimes spike upward for the smallest modules. As module granularity increases,
one must explore additional avenues for reducing dependencies. The next section
surveys a range of possibilities.
Ravioli Tastes Better Than Spaghetti Code
Spaghetti code results partly from spaghetti dependencies. One way to reduce the
dependencies between program lines is to reorganize the program by partitioning
the data into private, local pockets. To the extent that managing the memory
in these pockets determines the maximum resolution, a programs organization
determines its accuracy.

1.4 Alternative Programming Paradigms

Many programming paradigms represent communities as much as philosophies. In
some cases, the community members have written hundreds of publications building
veritable edices of human achievement. It would be impossible to wander through
one of these edices in the allotted time and fully grasp how the community goes
about its daily work. Cameras slung unabashedly around our necks, we hope to

19

20

Development Costs and Complexity

capture a few choice snapshots, to take in the sights of most interest, to discern the
basic building blocks, to overhear a bit of the mindset expressed in conversation, and
to return home changed forever. Although the following paragraphs read at rst like
a timeline, most of these communities remain vibrant in certain corners, with good
reason and bad.
In the beginning, there was unstructured programming. Early machine language
programming employed individual lines as the basic unit of decomposition. Although
programmers were free to think in terms of higher levels of organization, no language support for such basic forms as loops existed. With the advent of Fortran as
the rst high-level programming came looping constructs (do loops) and conditional
execution blocks (if constructs).
At this level of program organization, the only commonly employed program
abstraction is the ow chart. With the resultant emphasis on program ow rather than
data, the rst few evolutionary steps beyond unstructured code involved new ways to
express and decompose algorithms. The 1958 version of Fortran, for example, added
the capability to dene a subroutine or function. The procedures constructed
therewith became the basic units of decomposition in procedural programming.
Whereas some paradigm shifts expand the developers toolset, others prune
away its rougher edges. Those aspects of structured programming and functional
programming considered here involve pruning. Given the natural emphasis on a
programs logical ow that stems from staying within the procedural paradigm, most
structured programming methods emphasize the avoidance of arbitrarily employing
discontinuous jumps from one region of code to another. Because the unconditional
goto statement offers the most severe and abrupt such jump, the goto has been
frequently maligned by structured programming adherents. One of the founders of
the structured programming movement, Dijkstra (1968), summarized some of the
basic reasons in an article entitled GOTO statement considered harmful.
Knuth (1974) pointed out that the most harmful uses of goto were already beginning to be subsumed by newer, more specic constructs, a prime example being
error handling, where it is desired to exit some branch in a programs call tree to dispatch a distant portion of code intended to deal gracefully with an aberrant situation
such as reaching the end of an input le. Knuth attempted to redirect the structured
programming dialogue toward more fundamental questions of composing programs
from collections of entities about which formal correctness proofs can be made. He
also provided a simple rule for avoiding complicated ow: It should be possible to
draw all forward jumps in a ow chart on one side and all backward jumps on the
other side with no jumps crossing each other. Most references to spaghetti code are
to programs that violate these principles. Nonetheless, as late as 1987, the debate on
the goto statement raged on, with Rubin (1987) writing an article entitled GOTO
considered harmful considered harmful that garnered multiple responding letters
to the editor, including one from a disappointed Disktra.
In addition to eschewing goto statements in this books source code, a recent
incarnation of structured programming is adopted to facilitate splitting tasks across a
number of hardware units. The Open Multi-Processing (OpenMP2 ) toolsets for parallelizing code operate on so-called structured blocks of code. The OpenMP standard
2 http://www.openmp.org

1.4 Alternative Programming Paradigms

denes a structured block as a contiguous set of statements with one point of entry
and one point of exit.
Lest we pat ourselves on the back for advancing so far from the Stone Age, a
resurgence of unstructured programming came with the advent of interactive mathematical packages. This trend has many benets. For relatively simple tasks and even
for some more complicated tasks with regularly spaced data and modest operation
counts, the benets of packages such as Matlab and Mathcad far outweigh their disadvantages. The ability to leverage hundreds of programmer-years with the press of
a button, to integrate visualization into calculations, and especially to work with units
that are automatically converted when necessary in Mathcad outweighs the difculty
of life without subroutines and functions for small code segments of a 1020 lines.
Of the approaches discussed in this section, functional programming represents the
rst that deals with the data dependencies discussed in the previous section. Functional programming emphasizes the specication of relationships between entities,
rather than specifying steps that must be done. Mathematica is the most commonly
employed functional programming language/environment with which most scientic
programmers would be familiar.
The subset of functional programming that most inuences this book is purely
functional programming, in which side effects, including modifying procedure
arguments, are completely ruled out. While we nd it impractical to rule out side
effects completely, partly for reasons explained in Chapter 3, minimizing side effects
amounts to limiting data access. Limiting data access reduces the linkages in diagrams
like Figure 1.5, which has an important impact on the debugging time estimate of
equation (1.16). We discuss this impact in Chapter 3.
Object-based programming further limits data access by encouraging encapsulation and information hiding. Although Chapter 3 provides more complete
explanations, here we dene encapsulation as the act of partitioning data and procedures for operating on that data into modules. The data referenced here is inside
a derived type of the form
type fluid
real :: viscosity
real, dimension(:,:,:), allocatable :: velocity
character, allocatable :: chemical_formula
end type
With the object-based approach, one puts procedures into a given module only if
they operate on instances (objects) of the type dened in that module or provide
services to a procedure that does. Furthermore, one generally lets only those module
procedures, typically termed methods in OOP discussions, operate on instances of the
modules data type. (Chapter 3 details the construction of methods via Fortran 2003
type-bound procedures.) The latter relationship represents data privacy, which is one
form of information hiding. Another form involves keeping some of the procedures
in the module inaccessible to code outside the module. Each of these strategies limits
data access and thereby helps reduce the debugging search time. Encapsulation and
information hiding are so closely linked that they are often considered one concept.
That concept is the rst of the three main pillars of OOP.

21

22

Development Costs and Complexity

The second pillar, polymorphism, has Greek roots meaning many faces. One
category of polymorphism takes the form of a one-to-many mapping from procedure
names to procedure implementations. Here we dene an interface as the information
summarized in a Fortran interface body: a procedures name, its argument types,
and its return type, if any. Although polymorphic variables did not enter the language until Fortran 2003, Fortran programmers had already encountered one form
of polymorphism when using intrinsic functions. For example, in writing the code:
real :: x=1.,sin_x
real ,dimension(10,10) :: a=0.,sin_a
sin_x = sin(x)
sin_a = sin(a)
where the function sin can accept a scalar argument or an array arguments and
correspondingly return either a scalar or an array. The procedure that responds to
each of the above calls must be somewhat different even though the call retains the
same form. Chapter 3 explains other forms of polymorphism and how the programmer can construct her own polymorphic procedures. Chapter 3 also analyzes the
inuence polymorphism has on programming complexity.
One form of polymorphism is so important that it is typically given its own
name: operator overloading. Again it is closely related to something Fortran 90
programmers have already encountered in writing expressions like:
real
:: x=1.,y=2.,z
real ,dimension(10,10) :: a=0.,b=1.,c
z = x + y
c = a + b
where the + operator apparently has different meanings in different contexts: the
rst being to add two scalars and the second being to perform element-wise addition
on two matrices. Operator overloading is when the programmer species how such
operators can be applied to derived types created by the programmer. Chapter 3
explains how operator overloading plays such a critical role in reducing programming
complexity that it could be considered the fourth pillar of scientic OOP (SOOP).
Object-oriented programming is object-based programming plus inheritance, the
third pillar of OOP.3 Inheritance increases code reuse by facilitating subtypes, to be
dened based on existing types. In Fortran 2003, subtypes are termed extended types
and take the form:
type ,extends(fluid) :: magnetofluid
real ,dimension(:,:,:) ,allocatable :: magnetic_field
end type magnetofluid
which species that a magnetofluid is a fluid with an associated magnetic eld.
Every instance of a magnetofluid thus inherits a viscosity, velocity, and a chemical
3 Perhaps the best denition is that object-based programming is OOP minus whatever one considers

to be the most critical feature of OOP. For example, another interpretation is that object-based
programming is OOP minus polymorphism (Satir and Brown 1995).

1.4 Alternative Programming Paradigms

formula from the fluid type that magnetofluid extends. Equally important, languages that support OOP also provide facilities for the module implementing the
magnetouid to inherit various methods from the Fluid module. By analogy with
biological inheritance, the extended type is often referred to as the child, whereas
the type it extends is termed the parent.
Component-based software engineering (CBSE) can be viewed as the next evolutionary step beyond OOP. Where OOP informs how we construct and connect
individual objects, CBSE works with components that can be comprised of collections of objects along with code not necessarily written in an object-oriented fashion.
Most of the CBSE toolsets developed for mainstream software engineering have
severe inefciencies when applied to scientic problems. Where distributing applications across the Internet is increasingly the norm in nonscientic work, parallelizing
applications across a set of closely linked computational nodes has become the norm
in scientic work. The communication patterns and needs differ considerably in these
two domains.
Fortunately, for high-performance scientic programming, the CCA facilitates
the management of interprocedural communication at run-time among programs
that might have been written in disparate languages employing equally disparate programming paradigms (Bernholdt et al. 2006). A concept closely linked with CBSE
is design by contract, which expresses the notion that software clients write specications with which software designers must comply. In fact, one denition of a
component is an object with a specication. Chapter 11 discusses a toolset for writing specications and how the specication process inuences programming in ways
that increase code robustness.
The generic programming paradigm enhances programmers ability to write
reusable algorithms. It is often referred to as programming with concepts, wherein
procedures or algorithms are expressed with a high level of abstraction in the applicable data types. When sorting a set of objects, for example, whether they are of
numeric type or the previously dened fluid type, the well-established quick-sort
algorithm can be applied, providing a well-dened comparison method between any
two objects. Appropriate use of this paradigm leads to highly reusable components,
thus rendering simpler programs.
A number of modern programming languages explicitly support generic programming. Templates in C++, such as the container classes in the Standard Template
Library (STL), are among the most commonly used generic constructs. Templates
allow one to write code in which the type being manipulated can remain unspecied
in the expression of the algorithm. Even though Fortran lacks templates, one can
emulate the simpler template capabilities in Fortran. Part II discusses some templateemulation techniques. Furthermore, Fortran supports its own version of generic programming via parameterized derived types that allow type parameters to be generic
in source. Type parameters specify attributes used in dening the derived type as in:
type vector(precision,length)
real(precision) ,dimension(length) :: v
end type
where precision and length are type parameters that can be used in a subsequent
declaration to specify the kind and dimension of v.

23

24

Development Costs and Complexity

Although not precisely programming paradigms, design patterns and formal

methods represent two software engineering subdisciplines that greatly inuence
designs. In light of the earlier presented denition of OOP, we can now state that design patterns specify the roles that the programmer-dened types play and how those
roles mesh to form a solution to some recurring problem. Object-oriented scientic
design patterns comprise the central focus of Part II of this text.
Closely related to the design by contract notion, formal methods represent
mathematically precise specications of algorithms and rigorous verications that
a particular implementation is correct with respect to the specication. Scientic
programming is already formal in its numerical algorithms. The mathematical model
to be approximated, for example, Maxwells equations for electromagnetics, provides a formal specication. Convergence proofs for the discrete approximations
employed provide a formal verication that the numerical algorithms conform to
the specication. The eld of formal methods brings this same level of rigor to nonscientic software. Chapter 10 discusses how scientic programmers can benet from
specifying formal constraints to the nonnumerical aspects of scientic software, for
example, memory management.
You Say Tomato. We Say Baloney!
Although it is common in the scientic community to think of the differences in
coding practices as purely stylistic (You say toe-may-toe. I say toe-mah-toe.),
quantitative arguments presented in the previous section suggest that structured
programming scales poorly with increasing program size. Functional, objectbased, object-oriented, generic, and component-based approaches each build
on the structured paradigm while increasing its ability to scale with package size.

1.5 How Performance Informs Design

Since this books development-time focus runs counter to the prevailing trends in
the scientic computing literature, it is incumbent on us to explain how design relates to the more common concern: performance. Any discussion of performance
ultimately turns on Amdahls law, which is discussed in detail in Chapter 12. For
present purposes, what matters is that Amdahls law determines the maximum factor by which any process can be sped up. Specically, it shows that the maximum
fractional speedup is limited by the fraction of the process that can be sped up. If
80% of the process can be sped up, the maximum speedup is:
trun
=5
(1 0.8)trun

(1.17)

where the numerator and denominator are respectively the original completion time
and the minimum time based on completely eliminating 80% of the process. This
calculation suggests that when modest speedup is acceptable, optimization efforts
should be focused on the code that occupies most of the run time. The search for
those parts leads naturally to the following rule of thumb:
Pareto Principle: For many systems in which participants share resources, roughly 20%
of the participants occupy 80% of the resources.

1.6 How Design Informs Performance

Specically, experience indicates that for many codes, approximately 20% of the lines
occupy 80% of the execution time. Though this situation might seem purely fortuitous, some form of Pareto rule always holds. In any system where participants (lines)
share resources (execution time), there will always be a number k such that k% of
the participates occupy 1 k% of the resources so long as the number of participants
is sufciently large to admit a fraction equal to the chosen percentage. Two limiting
cases are equal sharing (k = 50%) and monopoly by one line (k 100%). Hence,
the Pareto Principle suggests most systems are skewed slightly towards monopoly.
This very likely stems from the fact that, considering all possible resource allocations
for N participants, there is only one that corresponds to an equitable distribution,
whereas there are N possible monopolies.
Considering Amdahls law vis--vis Paretos rule leads to the design principle that
guides this entire book: structure 80% of the code to reduce development time rather
than run time. Focus run-time reduction efforts on the 20% of the code that occupies
the largest share of the run time. Ultimately, this means design must be as analytical
and empirical as the numerics and physics. Development time estimates must guide
the design. Solution time estimates must guide the algorithm choice. Proles of the
run-time share occupied by each procedure must guide the optimization efforts.
Rouson et al. (2008b) discussed a common case that is even far more extreme than
Paretos Principle: 79% of the run time is occupied by less than 1% of the code.
The reason is that this short code segment calls an external library that does the
heavy lifting. With increasing development of high-performance scientic libraries,
this situation can be expected to become the norm.
The percentage of the code that must be highly optimized increases with the
desired amount of speedup. When the optimization focuses on parallelization, the
process of scaling up from running on a single execution unit to running on many
thousands eventually takes the search for optimization into the darkest crevices of the
code. If one starts from code that is already parallel, however, the Pareto-inspired
rule-of-thumb holds: Targeting a vefold increase in speed is as reasonable when
going from 1 to 5 execution units as when going from 10,000 to 50,000.
May the Source Be With You
Write 80% of your source code with yourself and your fellow programmers in
mind, that is, to make your life easier, not your computers life easier. The CPU
will never read your source code without an interpreter or compiler anyway.
Arcane or quirky constructs and names written to impress on the CPU the urgency
or seriousness of your task might well get lost in the translation.

1.6 How Design Informs Performance

Users who keep pace with the speed increases of leadership-class machines pace
themselves against thousandfold speedup goals over decadal time periods. In going
from the petaops machines of the year 2010 to the exaops platforms4 projected for
2020, most of the speedup must come from distributing data and nding additional
4 Petaops and exaops correspond to 1012 and 1015 oating-point calculations per second, respec-

tively.

25

26

Development Costs and Complexity

opportunities for parallel execution. Since current trends suggest that data distribution costs will dominate computing costs at the exascale, the chosen constructs
must minimize the data distribution needs. A signicant driver for data distribution is the need to synchronize memory when one execution unit modies data that
another unit needs. Over the past decade, the computer science community has
issued many calls for functional programming as a solution that limits such communication: Side-effect-free procedures do not modify state that other procedures
need.
Any strategy that promises exaops must allow for a rapidly evolving landscape
of hardware, programming languages, and parallel programming models. Hardware
is likely to become heterogeneous with exotic combinations of processor types. Old
languages will evolve while new languages are under development. The current parallel programming toolsets will themselves evolve whereas many of their ideas will
ultimately be embedded into new and established languages. One way to mitigate
the impact of this evolution on application programmers relies on establishing highlevel, software abstractions that can remain reasonably static and self-consistent
over time. Many of the most impactful application domains, for example, energy
and climate sciences, will remain as high a priority a decade hence as a decade
prior. In these physics-based elds, the tensor form of the governing equations
will not change. Thus, tensor calculus abstractions hold great promise for achieving
abstraction invariance.
Finally, candidate abstractions must be language-agnostic. OOD patterns thus
provide a natural lexicon for expressing candidate abstractions. Developers can implement OOD patterns in any language that supports OOP, as Part II of the current
text demonstrates in Fortran 2003 and C++.
All of the reasoning in this section suggests architecting scientic software by
combining OOD patterns built around the concept of a parallel, tensor calculus
based on side-effect-free, composable functions. This forms the central theme and
ultimate aim of the current book, namely that the design aims of a book focused
on expressiveness over performance ultimately align with performance points to a
deep truth. Consider that the calculus abstractions employed throughout much of
this book are modeled after the analytical expressions used in deriving closed-form,
analytical solutions to PDEs. In writing analytical solutions, one often arrives all at
once at the solution for the whole domain or some unidirectional factor in the
solution with methods such as the separation of variables or some globally varying
term in the solution in the case of a series expansion. Thus, analytical solutions
are intrinsically parallel. One expects the same to be true of software constructs
patterned after continuous forms.

Cleanliness Is Next to Speediness

When the ultimate goal is to solve tensor PDEs, an approach intended to generate exceptionally clean code also illuminates a path to efcient parallel code.
Functional programming limits communication. OOD patterns shield the application programmer from language dependency, while the use of high-level, tensor
calculus constructs ensures the constancy of the abstractions over time.

1.7 Elements of Style

1.7 Elements of Style

By denition, a high-level language is a human tongue meant for human eyes, so one
of the ways to reduce a codes development time is to make it more easily readable
for humans. To balance the emphasis on quantitative software descriptors in the
remainder of Part I, this section offers a few qualitative tips to improve readability.
Rule 1.1. Write code that comments itself.
The suggested goal is to write code that reads like prose with the richness and
structure of natural language grammar, including nouns, verbs, prepositions, and so
on. Consider the following Fortran code:
type(gas)
:: air
real ,parameter :: viscosity=1.52E-05 ! meters/sec
if (.not.readFromDisk(air)) air = constructGas(viscosity)
and the corresponding C++:
gas
air;
const float
viscosity=1.52E-05; // meters/sec
if (! readFromDisk(air))
air = constructGas(viscosity);
It should be clear that we have dened a data structure representing some abstraction
of a gas, that the gas to be modeled is air, that the airs viscosity is constant at
1.52 105 m/s, that we will attempt to read the data structure from a disk, and that
we will construct it based on its viscosity if the read is unsuccessful. As seen here,
writing code that comments itself does not eliminate the need for comments but does
minimize it.
A common exception to writing code prosaically arises when the notation in a
given domain is sufciently expressive that the code will be clearest in that notation.
Inside procedures that solve, for example, the Navier-Stokes equations to model
gas motion, one might use u or v for the gas velocity because those symbols have
that meaning almost universally in the uid dynamics literature. Many similar examples are discussed throughout this book. Nonetheless, the key point is to avoid the
strong aversion most scientic programmers have toward using descriptive (and yes,
lengthy) names for data structures. No matter how concise, convenient, or catchy it
seems to use VTP0 as the name of an array that holds velocities, temperatures, and
pressures at time zero, such habits are to be avoided like illicit drugs. Just say no!
Rule 1.2. Name all constants.
Even worse than using cryptic names is the use of no name at all. Without an
additional comment, the meaning of the numeric constant in the following code will
be hard to decipher even by the programmer who wrote it if a long time has passed
since its writing the following Fortran code:
if (.not.readFromDisk(air)) air = constructGas(1.52E-05)
and the corresponding C++:
if (! readFromDisk(air))

air = constructGas(1.52E-05);

27

28

Development Costs and Complexity

The constant denitions in the previous examples serve dual roles as declarations
and effectively as comments.
Rule 1.3. Make constants constant.
The only downside to deciding to declare and initialize the viscosity is that
every time we declare a new data structure, we increase the possible data dependencies identied in section 1.2 as a key cause of increased programming complexity.
Declaring the viscosity constant via the parameter or const keywords in Fortran
and C++, respectively, gives the code read-only access to it, thereby reducing data
dependencies.
Rule 1.4. Minimize global data.
Rule 1.5. Make global data constant.
Rule 1.6. Provide global type parameters, including precision speciers.
Although global data sharing is to be discouraged, dening a minimal set of
values with read-only access (i.e., constants) causes no harm and, in the case of
type parameters, makes the code considerably more exible. Other globally useful
values include mathematical and physical constants. The kind_parameters module
in Figure 1.4 illustrates. For example, if every module uses this module and every
real type declaration takes the form:
real(rkind) :: x
then changing the entire software package from single- to double-precision proves
trivial from the software perspective numerical considerations aside.
Values that require coordination across multiple les also require global specication in order to avoid conicts. These might include le names and unit
numbers.
Rule 1.7. Provide global conditional debugging.
As proposed by Akin (2003), another useful global constant in Fortran is one
that facilitates toggling between production mode and debugging mode as follows:
logical ,parameter :: debugging=.true.
One can use such a parameter to perform useful debugging tasks conditionally. One
such task might be to print a message at the beginning and end of each routine.
Making debugging a global constant allows most compilers to remove any code that
depends on it when a high optimization level is employed. Lines such as:
if (debugging) print constructGas: starting
...
if (debugging) print constructGas: finished
would be removed during the dead-code removal phase of most optimizing compilers
if debugging is .false. Hence, there is no run-time penalty during production runs.
This practice proves most important in Chapter 2 and beyond, where we adopt a
programming strategy that leads to calling sequences that can only be determined at
runtime.
Rule 1.8. Declare intent for all procedure arguments.

1.7 Elements of Style

Rule 1.9. Avoid side effects when possible, particularly in a function.

Rule 1.10. Make all derived type components private.
Each of these practices reduces data dependencies, thereby directly impacting
bug search time in ways discussed throughout this chapter. Rule 1.9 also relates to the
goal of writing code that comments itself. The notation for invoking procedures that
return values resembles the notation for mathematical functions. Since mathematical
functions do not modify their arguments, code reads more clearly when the same is
true for function procedures.
Rule 1.11. Indent loops and procedure/module executable statements.
Rule 1.11 provides visual clues to code structure, indicating nesting levels in
nested constructs.
Rule 1.12. Prevent implicit typing globally.
The implicit none statement in Figure 1.4 forces the programmer to explicitly
declare the type of all variables used. Most Fortran programmers are aware of this
rule. We mention it as a reminder to include it in every module.
Figure 1.4 also illustrates several other practices we nd useful. First, we frequently align tokens in columns to expose semantic commonality. The alignment
of the commas in lines 25-28 of the conduction module provides one example. We
also attach commas to the tokens they precede, such as the intent speciers in lines
25-28. This serves as a reminder to delete the comma if one deletes the token that
necessitates the comma.
Second, as demonstrated in line 2 of Figure 1.4(b), we use the optional only
construct to modify use statements. This explicitly noties the reader of where to
nd denitions of entities used in a given module but not dened in that module.
Finally, line 2 of Figure 1.4(c) demonstrates the iso_fortran_env Fortran 2003
module that facilitates portable access to the runtime environment specically the
input_unit le unit provides portable access to the pre-connected input stream
known as standard input in Unix-like environments. We mention it here to advocate
for using standard identiers wherever possible.
Some of the previously mentioned rules warrant stricter adherence than others.
For example, naming every constant can prove cumbersome. We occasionally skip
it when we feel very condent that the meaning is clear. For example, the statement:
do file_unit=1,max_unit
would be a bit more clear (and more general) if rewritten as:
do file_unit=first_unit,last_unit
We nonetheless will use the rst form when it seems reasonable 1 will be
understood to be the rst le unit and when its value is very unlikely to change.
Another rule we adhere to without complete consistency regards character case.
Because Fortran is internally case insensitive, mixing upper and lower case can give
a false impression. For example, a reader might think that entities t and T refer to
different quantities say, time and temperature when the compiler considers them
to be the same variable. We prefer lower case most of the time, but do occasionally
switch case when we feel condent it claries more than confuses.

29

30

Development Costs and Complexity

By contrast, we have never found a good reason to violate data privacy in production code, although it is sometimes useful temporarily while debugging. The point
is to habituate oneself to writing robust code by default, rather than to impose an
unnatural level of uniformity. As Ralph Waldo Emerson said, A foolish consistency
is the hobgoblin of small minds.
Love Thy Neighbor As Thy Source
Structure your code so that it comments itself and reads like prose. Missives
that would never have impressed the CPU might win you the adoration of programmers who have to read your code. Those programmers include your future
self.

EXERCISES
1. In some systems, cooling fans stay on after the rest of the system shuts down. To
determine how long it takes to cool the ns of Section 1.2 down, consider the n
to be a lumped mass with a spatially uniform temperature, T, that obeys:
mc

dT
out
= Q
dt

(1.18)

out is the heat transfer out of

where m is the n mass, c is the specic heat, and Q
the n.
out ,
(a) Draw a ow chart using explicit Euler time advancement to read m, c, Q
Tinitial , and Tnal from the keyboard, integrate the above equation until an
input temperature is reached, and output the required cooling time.
(b) Identify the sections of your ow chart that represent structured blocks.
(c) Describe several options for decomposing your ow chart into procedures.
2. Structured blocks are dened in Section 1.4.
(a) Count the structured blocks in Figure 1.3. (Hint: Blocks can be composed of
blocks with varying levels granularity.)
(b) Design a procedural decomposition of Figure 1.3, that is, dene a set of
procedure names, arguments, argument types, and return types (if any) that
could accomplish the tasks dened by the gure.
(c) Which of the procedures dened in part (b) are purely functional?
3. Repeat problem 2 for the ow chart constructed in problem 1(a).
4. Review the code in Figure 1.4.
(a) Calculate the maximum data dependencies.
(b) Cite specic lines that are data-independent and discuss how the actual
number of dependencies relates to the maximum.

The Object-Oriented Way

Believe those who are seeking the truth. Doubt those who nd it.
Andre Gide

2.1 Nomenclature
Chapter 1 introduced the main pillars of object-orientation: encapsulation, information hiding, polymorphism, and inheritance. The current chapter provides more
detailed denitions and demonstrations of these concepts in Fortran 2003 along with
a complexity analysis. As noted in the preface, we primarily target three audiences:
Fortran 95 programmers unfamiliar with OOP, Fortran 95 programmers who emulate OOP, and C++ programmers familiar with OOP. In reading this chapter, the
rst audience will learn the basic OOP concepts in a familiar syntax. The second
audience will nd that many of the emulation techniques that have been suggested
in the literature can be converted quite easily to employ the intrinsic OOP constructs
of Fortran 2003. The third audience will benet from the exposure to OOP in a language other than their native tongue. All will hopefully benet from the complexity
analysis at the end of the chapter.
We hope that using the newest features of Fortran gives this book lasting value.
Operating at the bleeding edge, however, presents some short-term limitations. As
of January 2011, only two compilers implemented the full Fortran 2003 standard:
The IBM XL Fortran compiler,
The Cray Compiler Environment.

However, it appears that the Numerical Algorithms Group (NAG), Gnu Fortran
(gfortran), and Intel compilers are advancing rapidly enough that they are likely
to have the features required to compile the code in this book by the time of
publication (Chivers and Sleightholme 2010).
Since Fortran is the primary language for Part I of this text, we default to the
Fortran nomenclature in this chapter and the next one. Upon rst usage of a particular term, we also include a parenthetical note containing the italicized name of
the closest corresponding C++ construct. The relationship between the Fortran and
C++ constructs, however, is rarely one of exact equality. Often, both could be used in
31

32

The Object-Oriented Way

Table 2.1. OOP nomenclature

Fortran

C++

General

Derived type
Component
Class
select type

Class
Data member
Dynamic Polymorphism
(emulated via
dynamic_cast)
Virtual Member function
Base class
Subclass
Namespace
Function overloading
Destructor
Overloaded operator
Overloaded assignment
operator
Pure virtual member
function
Function prototype
Primitive type/procedure

Abstract data type

Attribute

Type-bound procedure
Parent type
Extended type
Module
Generic interface
Final procedure
Dened operator
Dened assignment
Deferred procedure binding
Procedure interface
Intrinsic type/procedure
Booch

Method, operation
Parent class
Child class
Package
Static polymorphism

Abstract method
Procedure signature
Built-in type/procedure

et al. (1999)

additional ways that do not correspond with the common usage of the other. For example, we suggest below that Fortran 2003 derived types correspond to C++ classes,
but Fortran derived types can be parameterized in the sense that parameters declared in the type denition can be used to set things like precision and array bounds
directly in a type declaration without calling a constructor1 . No analogous feature
exists in C++. Likewise, C++ classes can be templatized, that is, types can be generic
in the source and not resolved to actual types until compile-time. No analogous feature exists in Fortran, although some of the simpler features of templates can be
emulated in Fortran as explained by Gray et al. (1999) and Akin (2003).
We focus on a capability subset common to Fortran and C++. When we draw
comparisons between Fortran and C++ constructs, we imply only that the two largely
correspond when used in the manner described herein. Table 2.1 summarizes the
approximate correspondences along with some language-neutral terms generally
employed for the same concepts. Most of the latter come from the Unied Modeling Language (UML), a schematic toolset for describing object-oriented software
independently from any chosen implementation language (Booch et al. 1999). As
demonstrated in the remainder of Parts I and II, and Appendix B using UML during
the software analysis and design phase facilitates focusing on the high-level modeling
concepts rather than specic implementations thereof.
1 Section 2.2 denes and discusses the special type of procedure referred to in OOP terminology as a

constructor.

2.2 Object-Oriented Analysis and Design

radio transmitter

greet

astronaut

Figure 2.1. Spacecraft radio transmitter use case diagram.

A Rose By Any Other Nomenclature...

OOP concepts stand apart from any chosen programming language. While
UML provides for programming-language-independent descriptions of OOP constructs, the constructs implementations differ amongst languages. We focus on
features common to Fortran 2003 and C++.

2.2 Object-Oriented Analysis and Design

Object-oriented analysis (OOA) and object-oriented design (OOD) comprise
sequential steps in the modeling of software structure and behavior. In the Unied
Modeling Language User Guide, Booch et al. (1999) described several techniques for
modeling software abstractly. In the current text, we cover only those few aspects
of OOA and OOD we have found useful in our work. In our experience, these correspond to the most frequently discussed techniques in computational science and
engineering software development. These techniques have played critical roles in our
discussions with research sponsors, in responding to journal referee reports, in presenting our work to colleagues, and in providing common idioms for communicating
with codevelopers in multi-language projects.
A central activity in OOA involves determining the relevant set of actors and
their actions. UML provides a standard schema for the graphical representation of
the actors, their actions, and the system through which they interact. That schema
entails drawing a UML use case diagram. As tradition dictates for programming
discussions, we start with a program that simply prints Hello, world!
Constructing a use case diagram requires considering who might be the actor
in such a model, what might be the system with which the actor interacts, and what
uses the actor has for that system. The actor in a Hello, world! diagram might
be an astronaut transmitting a message back to the earth via the spacecrafts radio
transmitter, as shown in Figure 2.1. In this setting, the transmitter is the system and
the usage case, or use case, is greet the world, which describes the actors goal in
using the system. Use case diagrams depict each use case inside an oval and depict
associations between the actors and the use cases by connecting the two with a line.
Optionally, use cases can be encompassed by a box bearing a label for the system: the
transmitter in the current context. As depicted, one always draws the actor external
to the system.

33

34

The Object-Oriented Way

Astronaut
greeting : String
constructor
+ astronaut(String) : Astronaut
+ greet() : String
Figure 2.2. Astronaut class diagram.

A more complete OOA might incorporate additional actors, such as a navigator, along with additional use cases such as request landing location. OOA also
considers various usage scenarios; however, one scenario sufces for the current,
simple example. Given the simplicity of the current objective, we instead proceed to
the OOD phase. OOD involves decomposing a problem domain into a set of useful
classes, each of which is also termed an abstract data type (ADT). An ADT describes a collection of objects with a common state representation and functionality.
Typically, an ADT embodies an abstraction, or abstract representation of whatever
system it models. The term abstract implies the ADT retains only the features and
behaviors required to represent the actual system in a particular context. One often
regards an object as an anthropomorphic, or at least animated, entity that receives
messages and collaboratively acts with other objects to accomplish some goal driven
by the received message.
The OOD results in a specication of the ADTs that can be used to facilitate the
given use cases. UML also provides a standard schema for diagramming a designs
structure: the UML class diagram, so named because class is a common synonym
for ADT. Due to differences in the meanings of the class keywords in Fortran and
C++, we prefer the language-neutral ADT terminology; however, we use class
interchangeably.
Figure 2.2 provides a UML class diagram for an astronaut ADT. A class diagram
represents an ADT in a three-part box. The top part provides the class name. In
Figure 2.2, the class is named Astronaut.
The middle part of a class diagram describes the attributes that is, data comprising the state of each instance of the class, each instance being termed an object.
Names can be preceded by + or , indicating whether they are public or private,
respectively. Public data can be referenced by code outside the module containing
the class implementation. Private data can only be referenced inside the module.
These can also be referred to as public or private scope. In this part of a class diagram, a datums type follows its name and is separated from the name by a colon. In
Figure 2.2, the state is represented by a private String greeting.
The bottom part of a class diagram details the operations that is, the methods.
This part provides the names, scope, and return types of procedures implemented
in the class. In OOP parlance, the procedure used to initialize, or instantiate, an
object is termed a constructor. In UML, one customarily lists constructors rst and
precedes them by the stereotype name constructor (Deitel 2008). UML encloses
stereotypes in guillemets ( and ). As indicated, the astronaut class implements
a public constructor that takes a String and returns an astronaut instance. It also
implements a public greet() method that returns a String.

2.2 Object-Oriented Analysis and Design

Array
+
+
+

size(Array<Element>) : Integer
shape(Array<Element>) : Array<Element>
allocated(Array<Element>) : Boolean

(a)

type
Integer

type
Character

type
Real

enumeration
Boolean
false
true

(b)
String : Array<Character>
(c)
Figure 2.3. UML Array model : (a) Array template class, (b) supporting primitive classes, and
(c) sample instantiation.

Figure 2.3 shows the attribute and operation portions of a UML diagram are
optional, as is the listing of private information. The Integer, String, and Real types
suppress all attributes and operations. The Array class diagram suppresses only
attributes. The Boolean type suppresses only operations.
Although UML does not contain an intrinsic array type, any types dened by a
UML model are considered UML model types. Henceforth, we assume in all of our
UML diagrams the availability of a utility template class enabling the construction
of arrays of various types. By denition, a template class is a class used to construct
other classes. One often says such a class is generic or, in UML terminology, parameterized. In the template class of Figure 2.3(a), the parameter is Element, which
must represent another class. We have omitted the attributes of the Array class in
Figure 2.3(a) under the assumption they are private. In most cases, they would include array descriptors holding the values returned by the methods depicted in the
bottom box of in Figure 2.3(a).
We can also model intrinsic or, in UML terminology, primitive types as classes.
The primitive types we use in our UML diagrams correspond to those in Figure 2.3(b).
The Integer class name draws inspiration from the like-named Fortran type. The
Boolean class provides an example of an enumeration of values: true and false.
It corresponds to the Fortran logical type.
The details of resolving the String type to an actual programming language type
differ from language to language and even from implementation to implementation.
In the code examples later in this chapter, we use allocatable deferred-length character variables, for which Fortran provides automatic dynamic memory management.

35

36

The Object-Oriented Way

Figure 2.3(c) depicts another option: instantiation of a String class as an array with
Character elements (which assumes prior denition of a Character type). Finally,
Figure 2.3(a) also depicts our assumption that the available array template class utility implements the functionality of Fortran allocatable arrays and can therefore be
allocated dynamically and can return information about its size, shape, and allocation
status.
Although the Array class in Figure 2.3(a) is 1D, one can readily build up multidimensional arrays from such a class. In Fortran, there is no need to do so since the
language has built-in multidimensional arrays. In C++, building a templatized multidimensional array by wrapping STL vectors proves straightforward. We demonstrate
this in Chapter 8.
In scientic work, an ADT often represents a physical entity, such as a uid
or plasma, or a mathematical entity such as a vector eld or a grid. Before turning
to more scientic examples, the next section describes the technology that supports
constructing an ADT in Fortran 2003 with references to the corresponding C++
constructs.
Codes Are People Too
Object-oriented thinking lends itself to an anthropomorphic view of objects. They
have state (attributes) and behavior (operations). They receive messages entailing
the operations invoked on them and the arguments passed to them. An operations
result conveys an objects response. A program thus consists of a set of objects
collaborating to accomplish the users goal. A UML use case diagram summarizes
the actors in this drama along with their activities. A UML class diagram details
the supporting class structure.

2.3 Encapsulation and Information Hiding

Encapsulation and information hiding provide the mechanisms for constructing an
ADT by tightly coupling private, local data with a set of public and private procedures. Each of these procedures accesses the data, and the data is accessed
by only these procedures. This two-way relationship creates the tight coupling.
In Fortran, we accomplish encapsulation by dening a derived type (C++ class)
with type-bound procedures (C++ virtual member functions) and wrapping each
derived type in a Fortran module (C++ namespace). Each module encapsulates
only one public derived type, though there may be other private ones included in
the module. OOPs information-hiding philosophy dictates that the public derived
types components (C++ data members) each has the private attribute. Outside
the module, access to the components is provided only via public type-bound
procedures.
Figures 2.4(a) and 2.4(b) show structured and object-oriented Hello, world!
programs, respectively. The structured program exhibits nothing noteworthy other
than our choice to name the character string constant in keeping with Rule 1.2 from
Section 1.7. The object-oriented program denes an astronaut class entailing a module that encapsulates an astronaut derived type with a component named greeting
and a type-bound procedure named greet(). Following the information-hiding edict

2.3 Encapsulation and Information Hiding

1
2
3
4
5

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19

(a)
program hello_world
implicit none
character(len=12),parameter :: message=Hello world!
print *,message
end program
(b)
module astronaut_class
implicit none
private
! Hide everything by default
public :: astronaut ! Expose type & constructor
type astronaut
private
character(:), allocatable :: greeting
contains
procedure :: greet ! Public by default
end type
interface astronaut ! Map generic to actual name
procedure constructor
end interface
contains
function constructor(new_greeting) result(new_astronaut)
character(len=*), intent(in) :: new_greeting
type(astronaut) :: new_astronaut
new_astronaut%greeting = new_greeting
end function

20
21
22
23
24
25
26

function greet(this) result(message)

class(astronaut), intent(in) :: this
character(:), allocatable :: message
message = this%greeting
end function
end module

27
28
29
30
31
32
33

program oo_hello_world
use astronaut_class ,only : astronaut
type(astronaut) :: pilot
pilot = astronaut(Hello, world!)
print *, pilot%greet()
end program
Figure 2.4. Hello, world! programs: (a) structured and (b) object-oriented.

of OOP, line 3 defaults all module components to private scoping, while line 4
makes public the derived type name and the generic constructor name. The derived
type component greeting is explicitly declared private. Since Fortran type-bound
procedures are public by default, greet() also has public scope.
The main program oo_hello_world declares an astronaut named pilot and
calls a like-named constructor that returns an astronaut instance. If the type components were public or had default initializations, we could have called the intrinsic
structure constructor that Fortran 2003 provides with the same name as the type

37

38

The Object-Oriented Way

name. The result of that constructor could even be used to initialize pilot in its
declaration on line 30. The information-hiding philosophy, however, recommends
against exposing components, and Figure 2.4(b) omits any default initialization. The
gure therefore includes a user-dened constructor.
The program then calls greet(), the invocation of which shows no arguments
on the calling side. The argument received by greet() in the astronaut class implementation represents the object on which it is invoked: pilot. The Fortran 2003
standard refers to this argument as the passed-object dummy argument. In most OOP
languages, including Java and C++, the type of that argument is assumed and need
not be explicitly declared in the denition of the procedure. In these other languages,
one refers to the corresponding entity by the keyword this, so we generally choose
that to be its local name in Fortran as well. The class keyword employed in the
declaration of this facilitates inheritance, which Section 2.5 discusses, and dynamic
polymorphism, which Section 2.6 discusses.
A common model for understanding OOP views a type-bound procedure invocation as sending a message to the object on which it is invoked, that is, the passed-object
dummy argument. The message conveys a request for action. Thus, one views line
32 in Figure 2.4(b) as the main program sending the request greet to the object
pilot. That object responds by returning a character variable containing its greeting.
Figure 2.4(b) also leverages the 2003 standards augmentation of Fortrans
dynamic memory management capabilities. The greeting component of the
astronaut type is an allocatable deferred-length character string. The semantics of
allocatable entities obligates the compiler to automatically allocate sufcient space
to store whatever character string gets assigned to greeting at line 18 and likewise
for whatever string gets assigned to the greet() function result at line 24. Furthermore, the standard obligates the compiler to free any associated memory when the
corresponding entities go out of scope. The compiler must also free (or reuse) any
memory previously assigned to the left-hand-sides (LHSs) of these expressions.
This rst foray into OOP demonstrates the overhead it adds to the software
development process. Much of the additional overhead comes in the early phases
of the process: the problem analysis and design that occur before ngers reach the
keyboard. This extra effort manifests itself in discussions about the relevant actors,
their goals, and their use of the system being modeled. For sufciently small, oneoff programming tasks, the overhead associated with OOA and OOD likely prove
cost prohibitive in that the complexity of Figure 2.4(b) exceeds that of Figure 2.4(a)
by virtually any straightforward measure: the number of lines of code, procedures,
modules, language features, or programming constructs employed.
To the extent portions of Figure 2.4(b) might be more easily reused and extended, however, the additional overhead pays off many times over. In such cases,
information hiding erects a partition between code inside the astronaut class and
any outside code that relies on it. Specically, data privacy opens up the possibility
of completely changing the internal representation of the type components without affecting code external to the class. For example, suppose we wish to provide
a default initialization for the greeting component in pilot to be printed in cases
when the user forgets to instantiate objects. Because allocatable character string
declarations cannot have default initializations, we might facilitate this by switching
to the following xed-width form:

2.3 Encapsulation and Information Hiding

type astronaut
private
character(len=27):: &
greeting=Houston, we have a problem!
contains
procedure :: greet
end type
Making only this change would not force any changes to the main program. Specically, the astronaut() and greet() procedure invocations at lines 31 and 32 could
remain unchanged.
The general principle illustrated by the latter example can be articulated by considering the concept of a class interface, which comprises the collection of entities
(procedure names, argument types, variables, and constants) a class makes public. Regarding public procedures, an actual interface might comprise one or more
interface blocks, containing one or more interface specications (C++ function prototypes). In Fortran, such interface specications only need to be expressed in source
code when dealing with external procedures. External procedures are those not encapsulated in a module. Those without a corresponding interface specication are
said to have an implicit interface. The presence of an interface specication enables
the compiler to check for argument type/rank/bounds mismatches between the calling code and the called code. Such type safety is provided automatically for module
procedures. Module procedures are those contained in a module. Module procedures
are said to have explicit (though confusingly tacit) interfaces.
In order to facilitate argument type checking, we encapsulate all procedures
in this book inside modules. We therefore have no occasion to include interface
specications in programs except to link to C/C++ procedures in Chapter 11. We
employ the term interface to refer to information that, in practice, we never isolate
into one place separate from the executable code that implements that interface.2
Our class interface denition provides a succinct way to summarize one of the
primary benets of information hiding: Changes to privately scoped entities inside a
class, including changes to the internal representation of its data, have no impact on
code outside the class so long as the changes would require no modications to the
class interface. The power of this rule cannot be overstated. Section 2.7 provides an
analysis of the impact this rule has on code complexity and robustness.

Friends Dont Let Friends Share Data

Encapsulating data in an ADT and hiding it from outside code gives developers
the exibility to alter implementations without impacting client code so long as
the ADTs interface remains untouched.

2 In Fortran 2008, it is possible to move the implementation of the type-bound procedures to a separate

submodule, leaving only their interfaces in the module that denes the type. This allows complete
documentation of the interface needed by users of the module without exposing details of how the
services described in the interface are implemented.

39

40

The Object-Oriented Way

2.4 Wrapping Legacy Software

Projects need not be object-oriented from their inception in order to benet from
OOA and OOD. In fact, a principle motivation behind OOP in Fortran 2003 stems
from the prospect of seamlessly interfacing to the vast collection of Fortran programs already in use. These codes embody an indeterminately large number of
people-years worth of inherited knowledge. Reproducing this knowledge in new
code, besides being inefcient, breaks down the trust the legacy codes have earned
through years of use. This section addresses the process of wrapping legacy Fortran
in modern, object-oriented Fortran, while Chapter 11 covers wrapping C++ with
object-oriented Fortran.
Figure 2.5 delineates a strategy for modernizing legacy Fortran programs. This
strategy has been adapted from one Norton and Decyk (2003) proposed. Whereas
they discussed updating Fortran 77 codes to take advantage of Fortran 90/95 features,
Figure 2.5 further contemplates transition to Fortran 2003 and potential inclusion
in a CBSE-style framework. The rst several steps might be considered steps in
code refactoring, a term that most commonly refers to the process of tidying code.
The rst step addresses standard compliance. Because Fortran 90 retained all of
Fortran 77 as a subset, Fortran 95 deleted only a small subset of Fortran 77, and
Fortran 2003 contains all of Fortran 95, most standard-conforming Fortran 77 code
also conforms to the later standards. The rst step thus revolves primarily around
removing compiler-specic extensions.
The second step in Figure 2.5 addresses obsolescent and deprecated features.
The Fortran 95 standard equates obsolescence with redundancy, declaring features
as obsolescent if Fortran 90 and 95 provided better alternatives. (Features that were
already obsolescent by this denition in Fortran 77 were deleted in Fortran 95.) These
included xed-source form (now superseded by free-source form, which removes
the signicance of columns in source text), computed go to statements (superseded

Legacy Fortran code

Address Compliance
Replace non standardcompliant features

CBSE
Incorporate new code into a
component framework,
Integrate with larger
projects

Improve type safety

Address
Obsolescent/Deprecated
Features
Wrap common blocks,
Prefer use over include,
Replace statement functions
...

Wrap external procedures,

Eliminate implicit typing and
implicit interfaces,
...

OOA and OOD

Consider use cases,
Construct class diagrams,
Write method calls to
legacy procedures, ...

Add new capabilities

Create allocatable
dynamic memory,
Write final procedures,
...

Classes and Components

Figure 2.5. Steps for modernizing legacy Fortran programs (adapted from Norton and Decyk
[2003])

2.4 Wrapping Legacy Software

by the case construct), statement functions (superseded by internal functions, i.e.,

functions preceded by a contains statement), and ve other features.
In addition to the ofcially obsolescent features, Metcalf et al. (2004) deprecated
several constructs and recommended they be declared obsolescent in a future standard. The features listed in Figure 2.5 fall into this category: common statements and
include lines. Norton and Decyk (2003) described how Fortran 77 common blocks
with static data can be reorganized to facilitate dynamic memory allocation. The
process involves moving the declaration into a module, writing constructors, and
replacing static data with allocatable entities. This adds considerable exibility and
functionality, including the methods in the Array class of Figure 2.3 and the ability
to gracefully handle exceptions such as insufcient available memory.
Wrapping include lines in Fortran 90 modules also adds functionality and provides greater safety. A module can grant access to entities declared in the included
code, but can also facilitate restricting access either via the private attribute inside the module or via an only clause in the use statement in client code. Wrapping
included code in a module also prevents the accidental declarations that can occur
when included legacy code contains common statements (see Metcalf et al. [2004]).
The third step in Figure 2.5 addresses type safety. Grouping external procedures
into a set of module constructs gives them explicit interfaces and thereby precludes the
argument type mismatches mentioned in Section 2.3. We prefer this approach to the
construction of actual interface specications. The lack of information redundancy
between an actual procedure and an associated interface specications eliminates
the need to maintain consistency between the two during code revisions. Additional
type safety derives from employing the implicit none statement to prevent implicit
variable declarations.
A side benet one incurs from the wrapping process will likely be greater type
safety. Fortran 77 provided no way to explicitly declare a procedure interface. As a
result, that version of the language uses a sequence association calling convention
when passing arrays between procedures. There are many problems inherent in such
calls due to the lack of type, rank, or bounds checking. Many legacy Fortran 77 codes
took artful advantage of this situation. For example, as Norton and Decyk (2003)
pointed out, many legacy Fotran 77 codes passed two-dimensional (2D) actual arguments that are received as 1D dummy arguments inside the called procedure. Modern
wrappers can employ a generic name (via a generic interface) for such a procedure,
resolving that name to the name of any one of multiple actual procedures that can
accept the desired argument the unsafe alternative being to make calls directly to
an external legacy procedure and to thereby abandon type/rank/bounds checking on
all arguments, not just the array in question.
In addition to being safer, wrapper interfaces can be simpler. Modern Fortrans
intrinsic functions for inquiring about array properties obviate the need to pass
these properties as arguments. Furthermore, the rich data structures and methods
attendant to objects also shortens argument lists.
The fourth step in Figure 2.5 contemplates new capabilities. We strongly advocate liberal use of allocatable constructs for dynamic memory management.
The language standards obligate compilers to free the memory associated with
allocatable entities when they go out of scope, thus preventing memory leaks with
any compliant compiler. Fortran 2003 greatly expanded the list of entities that can

41

42

The Object-Oriented Way

fin analyzer

specify
problem
system
architect

design
numerical
method
numerical analyst

predict heat
conduction
thermal analyst

Figure 2.6. Use case diagram for a n heat conduction analysis system.

have the allocatable attribute to include derived types and their components. We
have found that this capability eliminates the vast majority of cases where we would
otherwise use the pointer attribute. Nonetheles, when derived types do contain
pointers, Fortran 2003 final procedures must be used to free associated memory
when the pointer name goes out of scope.
The fth step in Figure 2.5 turns to the chief subject of this book: objectorientation. To illustrate this process, we return to the structured heat n conduction
of Chapter 1 and consider how to wrap that code in an object-oriented n analyzer
framework. Figure 2.6 provides a use case for the usage of such a framework. The
actors contemplated there include a system architect, a numerical analyst, and a
thermal analyst. The architect uses the n analyzer to specify the problem, providing
the numerical analyst with a set of relevant physical parameters, including tolerances
and boundary and initial conditions. The numerical analyst uses this specication to
develop a mathematical model in the form of a discrete set of equations that can be
solved on a digital computer. The thermal analyst uses the resulting mathematical
model to predict the performance of various ns.
Having a use case diagram in hand enables us to think abstractly about the
purpose of the software without tying us to a particular design or implementation.
Starting from a given use case diagram, different software development strategies
might generate designs and implementations bearing little resemblance to each other.
Conventional practice would likely proceed directly to the structured program implementation in Figure 1.4. Another strategy might model classes after the actors as
in the astronaut design of Section 2.3. A third might model classes after the activities.
An OOD following the latter path might lead to the class interfaces of Figure 2.7,
including a problem denition class, numerical differentiation class, and a heat conduction class. Chapter 3 presents a fourth approach that leads to a very different
class decomposition.
The Problem class described in Figure 2.7(a) provides a constructor that takes
an Integer le handle as its only argument. Although the implementation details

2.4 Wrapping Legacy Software

Problem
constructor
+ problem(le_handle : Integer) : Problem
+ boundary_vals() : Array<Real>
+ diffusivity() : Real
+ nodes() : Integer
+ time_step() : Real
+ spacing() : Real
(a)
Differentiator
constructor
+ differentiator(specication : Problem) : Differentiator
+ laplacian(T :Array<Real>,Tboundaries :Array<Real>) :Array<Real>
(b)
Conductor
constructor
+ conductor(d :Differentiator,spec :Problem,T :Array<Real>) :Conductor
+ heat_for(time_step : Real)
+ temperature() : Array<Real>
+ time_derivative() : Real
(c)
Figure 2.7. Object-oriented heat n analysis package class interfaces : (a) problem specication
class, (b) numerical differentiation class, and (c) heat conduction class.

remain unspecied by the interface, the corresponding le presumably contains the

information required to initialize the private Problem attributes. The Problem class
publishes operations that return physical parameters such as the boundary temperatures and the material diffusivity. Other operations return numerical parameters
such as the number of nodes, the time step size, and the grid spacing.
The Differentiator class described in Figure 2.7(b) provides a constructor
that takes a Problem argument and returns a Differentiator. It also provides a
.laplacian. operation. That operation returns an array of second derivatives given
arrays of internal temperatures and boundary temperatures.
The Conductor class described in Figure 2.7(c) provides a constructor that instantiates a Conductor from a Differentiator, a Problem, and a temperature array.
Conductor also provides an operation that conducts heat for a specied time (one
time step). A Conductor can also return an array of n temperatures and the time
derivative of the temperature at the chip boundary.
Having the class interfaces that result from an OOD provides an opportunity for
test-driven development. This strategy advocates for writing the test that veries a

43

44

The Object-Oriented Way

1
2
3
4
5
6
7

(a)
program fin_test
use iso_fortran_env ,only : input_unit
use kind_parameters ,only : rkind
use problem_class
,only : problem
use differentiator_class ,only : differentiator
use conductor_class
,only : conductor
implicit none

8
9
10
11
12

real(rkind), parameter :: tolerance=1.0E-06

type(problem)
:: specification
type(conductor) :: fin
type(differentiator) :: finite_difference

13
14
15
16
17
18
19

specification
= problem(input_unit)
finite_difference = differentiator(specification)
fin = conductor(finite_difference,specification)
print (a,5g9.3),initial temperature = ,fin%temperature()
call fin%heat_for(specification%time_step())
print (a,5g9.3),final temperature
= ,fin%temperature()

20
21
22
23
24
25
26
27
28

1
2

1
2
3
4

if (abs(fin%time_derivative())<tolerance) then
print (2(a,es9.3)),|dT/dt|=,fin%time_derivative(),<,tolerance
print *,In main: test passed. :)
else
print (2(a,es9.3)),|dT/dt|=,fin%time_derivative(),>,tolerance
print *,In main: test failed. :(
end if
end program
(b)
&physics alpha=1.,L_fin=1.,T_chip=1.,T_air=1. /
&numerics dt=0.01,nodes=3 /

initial temperature = 1.00

final temperature
= 1.00
|dT/dt|=0.000E+00<1.000E-06
In main: test passed. :)

(c)
1.00
1.00

1.00
1.00

1.00
1.00

1.00
1.00

Figure 2.8. Test-driven development: (a) main program, (b) input, and (c) output for adiabatic
test.

programs correctness before writing the program. To demonstrate, we consider the

adiabatic case, wherein spatially constant temperatures preclude heat transfer and
therefore correspond to a zero time derivative according to
T = const

T
= 2 T = 0
t

(2.1)

Figures 2.8(a)-(b) depict a main program and an input le that tests for this condition
given a set of classes that implement the interfaces of Figure 2.7. Figure 2.8(c) shows
the resulting output.

2.4 Wrapping Legacy Software

Figures 2.92.11 present implementations of the interfaces depicted in Figure 2.7.

Each implementation wraps parts of the structured program in Figure 1.4. As discussed in the remainder this chapter, repackaging the structured program in this
new object-oriented framework enhances its extensibility, exibility, and stability.
The enhanced extensibility derives at least in part from inheritance, the subject of
Section 2.5. The increased exibility derives in part from facilitating polymorphism,
the subject of Section 2.6. Stability derives in part from limiting data dependencies, the subject of Section 2.7. Chapter 3 demonstrates how a very different OOD
engenders even greater exibility.
Cake From A Box Never Tasted So Good
Transitioning from structured programming to OOA, OOD, and OOP does not
require mixing all your own ingredients and baking everything from scratch.
Wrapping legacy software preserves the trust that software has earned from years
of use. Appropriate wrapping increases type safety, simplies procedure calls, and
adds extensibility, exibility, and robustness in ways covered in the remaining
sections of this chapter.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27

Figure 2.9
module problem_class
use kind_parameters ,only: ikind,rkind
implicit none
private
public :: problem
type problem
private ! Default values:
integer(ikind) :: num_nodes=4
real(rkind):: air_temp=1.,chip_temp=1.!boundary temperatures
real(rkind):: alpha=1.,length=1.
!physical parameters
real(rkind):: dt=0.1,dx
!numerical parameters
contains
! constructor computes default
procedure :: boundary_vals
procedure :: diffusivity
procedure :: nodes
procedure :: time_step
procedure :: spacing
end type
interface problem
procedure spec ! constructor
end interface
contains
type(problem) function spec(file)
use conduction_module &
,only : read_physics,read_numerics,stable_dt
integer(ikind) :: file
integer(ikind) :: elements_default

28
29
30
31

elements_default = spec%num_nodes+1
spec%dx = spec%length/elements_default ! default element size

45

46

The Object-Oriented Way

32
33
34

if (.not. read_physics(file
&
,spec%alpha,spec%length,spec%chip_temp,spec%air_temp)) &
print *,In problem constructor: Using default physics spec.

35
36
37
38

if (.not. read_numerics(file
&
,spec%dt,spec%num_nodes,spec%dx,spec%length)) &
print *,In problem constructor: Using default numerics spec.

39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64

spec%dt = min(spec%dt,stable_dt(spec%dx,spec%alpha))
end function
pure function boundary_vals(this)
class(problem) ,intent(in) :: this
integer(ikind) ,parameter :: end_points=2
real(rkind) ,dimension(end_points) :: boundary_vals
boundary_vals = (/this%chip_temp,this%air_temp/)
end function
pure real(rkind) function diffusivity(this)
class(problem) ,intent(in) :: this
diffusivity = this%alpha
end function
pure integer(ikind) function nodes(this)
class(problem) ,intent(in) :: this
nodes = this%num_nodes
end function
pure real(rkind) function time_step(this)
class(problem) ,intent(in) :: this
time_step = this%dt
end function
pure real(rkind) function spacing(this)
class(problem) ,intent(in) :: this
spacing = this%dx
end function
end module
Figure 2.9. Problem denition class implementation.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18

Figure 2.10
module differentiator_class
use kind_parameters ,only: ikind,rkind
implicit none
private
! Hide everything by default.
public :: differentiator ! Expose type/constructor/type-bound procs.
type differentiator
private
real(rkind),dimension(:,:),allocatable::diff_matrix
contains
procedure :: laplacian ! return Laplacian
procedure :: lap_matrix ! return Laplacian matrix operator
end type
interface differentiator
procedure constructor
end interface
contains
type(differentiator) function constructor(spec)
use problem_class ,only : problem

2.4 Wrapping Legacy Software

19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42

use conduction_module ,only : differencing_stencil

type(problem) ,intent(in) :: spec
integer(ikind) ,parameter :: diagonals=3
allocate(constructor%diff_matrix(spec%nodes(),diagonals))
constructor%diff_matrix = &
differencing_stencil(spec%spacing(),spec%nodes())
end function
pure function laplacian(this,T,Tboundaries)
use conduction_module, only : differentiate
class(differentiator), intent(in) :: this
real(rkind) ,dimension(:) ,intent(in) :: T
real(rkind) ,dimension(:) ,allocatable :: laplacian
integer(ikind) ,parameter :: end_points=2
real(rkind), dimension(end_points), intent(in) :: Tboundaries
allocate(laplacian(size(T)))
laplacian = differentiate(& !compute derivative at internal points
this%diff_matrix,T,Tboundaries(1),Tboundaries(2) )
end function
pure function lap_matrix(this)
class(differentiator), intent(in) :: this
real(rkind) ,allocatable ,dimension(:,:) :: lap_matrix
lap_matrix = this%diff_matrix
end function
end module
Figure 2.10. Numerical differentiation class implementation.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27

Figure 2.11
module conductor_class
use kind_parameters ,only : ikind,rkind
use differentiator_class, only : differentiator
implicit none
private
! Hide everything by default
public::conductor ! Expose type and type-bound procedures
integer(ikind), parameter :: end_points=2
type conductor
private
type(differentiator) :: diff
real(rkind) :: diffusivity
real(rkind) ,dimension(end_points) :: boundary
real(rkind) ,dimension(:), allocatable :: temp
contains
! internal temperatures
procedure :: heat_for
procedure :: temperature
procedure :: time_derivative
end type
interface conductor
procedure constructor
end interface
contains
type(conductor) function constructor(diff,prob,T_init)
use conduction_module ,only : read_physics
use problem_class
,only : problem
type(differentiator) ,intent(in)
:: diff
type(problem)
,intent(in)
:: prob

47

48

The Object-Oriented Way

28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63

real(rkind), dimension(:) ,optional :: T_init

constructor%diff = diff
constructor%diffusivity = prob%diffusivity()
constructor%boundary = prob%boundary_vals()
if (present(T_init)) then
if (size(T_init)/=prob%nodes())stop In conductor: size mismatch.
constructor%temp = T_init
else; allocate(constructor%temp(prob%nodes()))
constructor%temp = constructor%boundary(1)
end if
end function
function temperature(this)
class(conductor), intent(in) :: this
real(rkind), dimension(:), allocatable :: temperature
temperature = (/ this%boundary(1), this%temp, this%boundary(2) /)
end function
subroutine heat_for(this,dt)
class(conductor) ,intent(inout) :: this
real(rkind)
,intent(in)
:: dt
real(rkind) ,dimension(:) ,allocatable :: T_xx !2nd derivative
allocate(T_xx(size(this%temp)))
T_xx = this%diff%laplacian(this%temp,this%boundary)
this%temp = this%temp + dt*this%diffusivity*T_xx
end subroutine
real(rkind) function time_derivative(this)
class(conductor)
:: this
real(rkind) ,dimension(:) ,allocatable :: T_xx_end
real(rkind)
:: T_2
if (size(this%temp)>1) then
T_2=this%temp(2)
else; T_2=this%boundary(2)
end if
T_xx_end = this%diff%laplacian(this%temp,(/this%boundary(1),T_2/) )
time_derivative = this%diffusivity*T_xx_end(1)
end function
end module
Figure 2.11. Heat conductor class implementation.

2.5 Composition, Aggregation, and Inheritance

Central to OOP lies the goal of code reuse. Most reuse in OOP takes one of three
forms: composition, aggregation, or inheritance. Each represents a relationship between two or more classes. Composition involves placing one or more instances of
one class inside the denition of another class as a component of the second class.
Composition can be thought of as a whole/part relationship.
In addition to conveying spatial information about the location of one object
inside another, composition harbors temporal connotations, implying coincident
lifetimes. The construction and destruction of the whole coincide with the construction and destruction of its parts. Hence, composition usually models relationships in
which the part loses its usefulness in the absence of the whole.
Composition need not be universal. Although every Conductor dened by
Figure 2.11 is composed of a Differentiator, not every Differentiator instance is part

2.5 Composition, Aggregation, and Inheritance

n:Conductor
diff

diff:Differentiator
diff_matrix = (/1.,-2.,1./)
Figure 2.12. Conductor/Differentiator composite object diagram.

of a Conductor. The finite_difference object in the main program (fin_test in

Figure 2.8) stands alone even though the conductor() constructor stores a copy of
it named diff in fin. The compiler destroys diff when fin goes out of scope (at
the program termination in this case).
Since not every Differentiator is part of a Conductor, their composite relationship might be best thought of as a relationship between two objects rather than
between two classes. A UML object diagram renders an instantaneous snapshot of
each objects state and its relationship to other objects. One portrays a relationship in
UML as a line connecting objects in an object diagram (or classes in a class diagram).
As Figure 2.12 shows, a software designer can indicate composition by adorning
the connecting line with a closed diamond at the whole end of the whole/part
relationship.
In object diagrams, one underlines the object and class names and separates them
by a colon. Furthermore, the object state at the time of the snapshot can be indicated
by assigning values to the objects attributes. The values assigned to diff_matrix
in Figure 2.12 correspond to the simple case of a single internal node with unit grid
spacing between that node and the boundaries.
When one class is composed of another, a software designer places a closed
diamond at the whole end of the whole/part relationship as shown in Figure 2.12.
If no symbol appears at either end (as in the bottom relationship in Figure 2.13), the
diagram conveys no information about the nature of the relationship. The designer
can also adorn both ends with a number, or number range, indicating the multiplicity,
which is the allowable number of instances at either end. The absence of explicit
multiplicities implies single instances.
Aggregation generalizes the composition concept by decoupling the lifetimes of
the whole and the part. As with composition, the ends of the relationship connecting
line can be adorned with a multiplicity indicating the number of possible instances
at either end. Authors frequently refer to aggregation as a has a relationship in
the sense that one class has an instance of another class. Since composition represents a specic instance of aggregation, another term for composition is composite
aggregation.
Inheritance denotes a very specic type of composite aggregation: one in which
the whole supports the entire public interface of the part. Authors frequently refer
to inheritance as an is a relationship in the sense that one class is a specic subcategory of another class because by default it retains the state and any nonoverridden
behavior of the parent class. Fortran provides inheritance via type extension. When
a child type extends a parent type, the compiler automatically inserts an instance of

49

50

The Object-Oriented Way

Conductor

Fun_conductor
constructor
+ fun_conductor(le_handle :Integer,temp_dist :Procedure) :Fun_conductor
type
Procedure
Figure 2.13. Function-initialized conductor inheritance hierarchy.

the parent type as an component in the child type. The child type can override any of
the parents methods by implementing them itself. Otherwise, the compiler handles
any invocations of the parents methods on the child by invoking these calls on the
parent. The Fortran language standard refers to the child as an extended type.
Consider a developer who wants to construct a Conductor by passing a temperature distribution function. We designate this new class fun_conductor as a
shorthand for function-initialized conductor. Since the existing Conductor constructor in Figure 2.11 accepts an optional argument containing nodal temperature
array, we can leverage this ability in the fun_conductor() constructor by accepting an argument that designates a temperature distribution function. The constructor
could sample that function at the node locations and pass those samples to the parent
types conductor() constructor.
The open triangle in Figure 2.13 indicates that a Conductor generalizes
a Fun_conductor. One usually models inheritance as generalization in UML.
Fun_Conductor enhances the Conductor interface by dening a new constructor and
simplies the interface by hiding the Problem and Differentiator objects inside the
fun_conductor() constructor. The test program in Figure 2.15 passes a le handle
and temperature distribution function to the constructor, delegating the acquisition
of all other construction-related data to the constructor.
Figure 2.13 declares the temperature distribution function to be of primitive
type Procedure, which might represent any of several language mechanisms for
passing a procedure argument. In C++, one might pass a function pointer. In Fortran, one might pass a procedure dummy argument (as in our implementation in
Figure 2.14), a procedure pointer, or a C function pointer (see Chapter 11). An OOP
purist might encapsulate the mechanism in a class, reserving the right to change the
private implementation details.

1
2
3
4
5
6
7

module fun_conductor_class
use kind_parameters ,only : ikind,rkind
use conductor_class ,only : conductor
implicit none
private
public::fun_conductor
type ,extends(conductor) :: fun_conductor

2.5 Composition, Aggregation, and Inheritance

8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44

private
real(rkind) :: dt
contains
procedure :: time_step
end type
interface fun_conductor
procedure constructor
end interface
contains
type(fun_conductor) function constructor(file,T_distribution)
use differentiator_class ,only : differentiator
use problem_class
,only : problem
integer(ikind) ,intent(in)
:: file
type(differentiator)
:: diff
type(problem)
:: prob
real(rkind), dimension(:) ,allocatable :: T
real(rkind)
:: dx
integer(ikind)
:: i
interface
pure real(rkind) function T_distribution(x)
use kind_parameters ,only : rkind
real(rkind) ,intent(in) :: x
end function
end interface
prob = problem(file)
diff = differentiator(prob)
dx
= prob%spacing()
constructor%dt = prob%time_step()
allocate(T(prob%nodes()))
forall(i=1:prob%nodes()) T(i) = T_distribution(i*dx)
constructor%conductor = conductor(diff,prob,T)
end function
real(rkind) function time_step(this)
class(fun_conductor) :: this
time_step = this%dt
end function
end module
Figure 2.14. Function-initialized conductor implementation.

Figure 2.14 shows a sample implementation of a function-initialized conductor that inherits the type-bound procedures of the Conductor class implemented
in Figure 2.11. Since linear temperature distributions also satisfy equation 2.1 and
therefore dene a steady-state solution, we can use a test analogous to the one used
for the Conductor class. Figure 2.15 demonstrates such a test along with the input
and output les.
The convenience of inheritance lies in the automated compiler support for
applying the operations of the parent to the instances of the child. The typebound procedures temperature() and time_derivative() invoked on fin in
Figure 2.15(a) result in calls to the corresponding procedures in the Conductor
implementation of Figure 2.11, even though the passed-object dummy argument

51

52

The Object-Oriented Way

1
2
3
4
5
6
7
8
9
10
11

(a)
module initializer
implicit none
private
public :: linear
contains
pure real(rkind) function linear(x)
use kind_parameters ,only : rkind
real(rkind) ,intent(in) :: x
linear = 1. - x
end function
end module

12
13
14
15
16
17
18
19
20

program fun_fin_test
use iso_fortran_env
,only : input_unit
use kind_parameters
,only : rkind
use fun_conductor_class ,only : fun_conductor
use initializer
,only : linear
implicit none
real(rkind) ,parameter :: tolerance=1.0E-06
type(fun_conductor)
:: fin

21
22
23
24
25
26
27
28
29
30
31
32
33

1
2

1
2
3
4

fin = fun_conductor(input_unit,linear)
print (a,5g9.2),initial temperature = ,fin%temperature()
call fin%heat_for(fin%time_step())
print (a,5g9.2),final temperature
= ,fin%temperature()
if (abs(fin%time_derivative())<tolerance) then
print (2(a,es9.3)),|dT/dt|=,fin%time_derivative(),<,tolerance
print *,In main: test passed. :)
else
print (2(a,es9.3)),|dT/dt|=,fin%time_derivative(),>,tolerance
print *,In main: test failed. :(
end if
end program
(b)
&physics alpha=1.,L_fin=1.,T_chip=1.,T_air=0. /
&numerics dt=0.01,nodes=3 /

initial temperature =
1.0
final temperature
=
1.0
|dT/dt|=0.000E+00<1.000E-06
In main: test passed. :)

(c)
0.75
0.75

0.50
0.50

0.25
0.25

0.0
0.0

Figure 2.15. Function-initialized conductor test: (a) main program, (b) input, and (c) output
for steady-state test.

this those procedures receive is of type fun_conductor rather than the declared
conductor. The next section describes this ability to pass objects of types other than
declared type.

2.6 Static and Dynamic Polymorphism

2.6 Static and Dynamic Polymorphism

Morpheus, the ancient Greek god of dreams, had the ability to take many different
human forms when appearing in dreams. Similarly, polymorphism connotes the ability of a given construct to take many object forms or procedural forms. To facilitate
taking different object forms, several objects might be able to respond to the same
message that is, the same method invocation either by implementing their own
version of the named method, by manually delegating to another object to respond,
or by letting the compiler automatically delegate to an ancestor. To facilitate taking
different procedural forms, several different procedures might have the same name,
whether or not these procedures are class operations.
Polymorphism takes two forms: static and dynamic. Both appeared in code in
the previous section. The current section highlights these occurrences and explains
the related concepts.
Static polymorphism denotes situations in which the responder to a message can
be determined at compile time, its identity henceforth xed throughout execution.
Many intrinsic functions are statically polymorphic. The expression exp(A) can refer
to the exponential of a real scalar, a complex scalar, a real matrix, or a complex matrix,
depending on the declared type of A. Each case invokes a different procedure and
returns a result of a different type.
Thus far, our code examples have included only one constructor per derived
type. Furthermore, our constructors are not type-bound procedures, so they are not
inherited by extended types. For this reason, they cannot be invoked on children.
We can use static polymorphism to enable the invocation of the parent Conductor
constructor on instances of the child Fun_Conductor by adding a procedure to the
interface statement in lines 13-15 of Figure 2.14 as follows:
interface fun_conductor
procedure constructor, conductor_constructor
end interface
and adding a module procedure that delegates construction to the parent instance
that the compiler automatically inserts as a component of the child:
type(fun_conductor) &
function conductor_constructor(diff,prob,T_init)
use differentiator_class ,only : differentiator
use problem_class
,only : problem
type(differentiator) ,intent(in)
:: diff
type(problem)
,intent(in)
:: prob
real(rkind), dimension(:) ,optional :: T_init
conductor_constructor%conductor = conductor(diff,prob,T_init)
end function
This effectively emulates inheritance of the parent constructor.
Dynamic polymorphism denotes situations in which the determination of
which procedure will respond to a given message can be delayed until runtime. Consider again the Conductor type-bound procedures temperature() and
time_derivative() invoked on fin in Figure 2.15. When writing these procedures

53

54

The Object-Oriented Way

in the Conductor class implementation of Figure 2.11, we originally planned for the
passed-object dummy argument to be of type Conductor. We compiled and ran that
code in the n test of Figure 2.8(c). Dynamic polymorphism, as designated by the
class keyword, enables the reuse of that compiled code, without recompilation, in
the subsequent writing of the Fun_Conductor class. The compiler generates object
code that can detect at runtime the actual argument passed and then dispatch code
capable of invoking the specied method on the passed object. The underlying technology is also referred to as dynamic dispatching. It proves especially useful when
the source code is unavailable, as might be the case in using a polymorphic library
constructed using the techniques in Chapter 6.
Fortran links the concepts of inheritance and dynamic polymorphism by requiring programmers to replace the usual type keyword with its class counterpart in
passed-object dummy argument declarations. The class keyword can be applied if
and only if the passed-object dummy argument is extensible. The actual argument
passed must then be of the type named in the class construct or any type that extends that type, including all generations of descendants. Such a chain of extended
types comprises a class hierarchy or inheritance hierarchy.

2.7 OOP Complexity

Each of the techniques discussed so far reduce development time in some way. A
little analysis helps determine which features of OOP offer the greatest benet. It
can be argued that inheritance and polymorphism attack problems that scale linearly with the size of the code (Rouson and Xiong 2004). Given a piece of code
that is inherited across G generations of a parent/child class hierarchy, inheritance
reduces the number of times this code must be replicated from G to 1. Given P
common procedures to be performed on C classes of objects, polymorphism reduces the size of the protocol for communicating with those classes from CP to P.
Note that the variables G, C, and P each appear with a unit exponent in the above
expressions.
By contrast, encapsulation and information hiding attack a problem that scales
quadratically with code size as shown in equation (1.16). Specically, enforcing strict
data privacy reduces the number of linkages in Figure 1.5, thereby limiting the number of lines one must visit to nd the bug. Specifying argument intent limits the type
of linkages by giving some procedures read-only access or write-only access to data.
To reect this in the debugging time estimate, we rewrite equation 1.16 as
tnd =


rc  c
1 tline
2 2

(2.2)

where c is the number of lines with write access to the solution vector. In OOP, c
corresponds to the number of lines in a class (assuming data privacy with one class
per module in Fortran). Hence, signicant reductions can only be accomplished if
class size restrictions can be enforced. Although it might seem draconian to restrict
class size across a project, Chapter 3 outlines a strategy that very naturally leads to
roughly constant c across the classes.

2.8 More on Style

2.8 More on Style

In our Fortran examples throughout this book, we typically avoid construct ending
names such as the second occurrence of foo in the following:
interface foo
procedure bar
end interface foo
Not using the ending name precludes the mistake of subsequently changing the rst
foo to new_foo, for example, without making the same change to the second foo.
Though easily rectied, it seems best to avoid the problem in the rst place when the
second occurrence serves little useful purpose. An exception to this practice arises
when the number of lines separating the beginning and ending of a construct likely
exceeds one page on a screen or in print, as happens with some long procedures
and main programs. In these instances, the second occurrence of the construct name
provides a useful reminder of which construct is ending.
We make occasional use of the forall statement, as in Figures 1.4 and 2.14. We
use this construct for two primary purposes. First, it occasionally proves more compact than a corresponding do loop because it has a single-statement form that does
not necessitate an end statement. Second, it indicates opportunities for parallelism.
Occurrences of forall can range over multiple array indices in a single construct
and apply conditional masks as in:
forall(i=1,imax/2, j=1,jmax, c(j) > 0.

) a(i,j) = b(i) + c(j)

where no assignment takes place for j values corresponding to positive c(j). In

addition to providing a compact syntax useful in a rst pass through non-critical
code, the semantics of forall require that it be possible for each of the resulting
assignments to take place concurrently. Unfortunately, the attendant code speedup
is typically moderate at best and negative at worst. Hence, the value forall adds in
clarity, conciseness, and concurrency must be considered in the context of the given
codes prominence in the overall performance prole.
We have chosen to write object constructors as standalone functions rather than
type-bound procedures. We consider this approach the most natural for Fortran
programmers because Fortrans intrinsic structure constructors take the same form.
The intrinsic structure constructors which are named after the type and have arguments named after the type components can only be used when the derived type
components are public or have default initialization.
Had we chosen to make the constructors type-bound procedures, the inheritance
of parent constructors emulated in Section 2.6 would be handled by the compiler
automatically. If the type-bound procedure constructor were to remain a function,
however, it would have to return a polymorphic result to allow for extensibility. This
would necessitate replacing the intrinsic assignment with a dened assignment. We
deem such machinations overkill for the simple examples discussed so far.
A nal alternative would be for the type-bound procedure to be a subroutine.
With that option, however, we lose the resemblance to Fortrans intrinsic structure
constructors. We also prefer the clarity of instantiating an object by assigning to it the
result of a constructor named after the derived type. Nonetheless, the stated benets

55

56

The Object-Oriented Way

must be weighed against the possibility that a compiler might make an unnecessary
temporary version of the object before copying it into the left-hand-side object and
destroying the temporary return instance. In many cases, optimizing compilers can
eliminate such temporaries.

EXERCISES
1. Verify that the corresponding tests fail with nonlinear initial temperature
distributions in (a) Figure 2.8(a) and (b) Figure 2.15(a).
2. In some systems, cooling fans stay on after the rest of the system shuts down.
To determine how long it takes to cool the ns of Section 1.2 down, consider
the n to be a lumped mass with a spatially uniform temperature, that obeys

mc dT
dt = Qout , where m and c are the n mass and specic heat and Qout is the
heat transfer rate.
(a) Draw a UML class diagram for a lumped mass class capable of determining
the time necessary to reach a nal cooling temperature from a given starting
temperature.
(b) Write a main program that uses the above class to output the cooling time.
List a few subclasses one might later desire to build from your lumped mass
class.

Scientic OOP

Software abstractions should resemble blackboard abstractions.

Kevin Long

3.1 Abstract Data Type Calculus

A desire for code reuse motivated most of Chapter 2. Using encapsulation and information hiding, we wrapped legacy, structured programs in an object-oriented
superstructure without exposing the legacy interfaces. We used aggregation and composition to incorporate the resulting classes into components inside new classes. We
also employed inheritance to erect class hierarchies. Finally, we employed dynamic
polymorphism to enable child instances to respond to the type-bound procedure
invocations written for their parents.
Much of the code in Chapters 12 proved amenable to reuse in modeling heat
conduction, but few of the object-oriented abstractions presented would nd any
use for nonthermal calculations. Even to the extent the heat equation models other
phenomena, such as Fickian diffusion, calling a procedure named heat_for() to
solve the diffusion equation would obfuscate its purpose. This problem could be addressed with window dressing creating a diffusion-oriented interface that delegates
all procedure invocations to the conductor class. Nonetheless, neither the original
conductor class nor its diffusion counterpart would likely prove useful in any simulation that solves a governing equation that is not formally equivalent to the heat
equation. This begs the question: When reusing codes, what classes might we best
construct from them?
Most developers would agree with the benets of breaking a large problem into
smaller problems, but choosing those smaller problems poses a quandary without a
unique solution. If the primary goal is reuse, however, one can draw inspiration from
the OOP proclivity for designing classes that mimic physical objects and mathematical constructs. When the entities being mimicked exhibit high levels of reusability in
their respective domains, one naturally expects similar levels of reusability in their
software counterparts.
In physics-based simulations, for example, elds and particles are nearly ubiquitous. Students rst acquire the concept of a eld as a function of space and time.
57

58

Scientic OOP

They get exposed to eld instances starting with scalar elds such as pressure and
temperature. These concepts get generalized to include vector elds such as electric
and magnetic vector elds. Ultimately, further generalization leads to higher-order
tensor elds such as stress elds in continuum mechanics and energy tensor elds in
relativity.
The apparent reusability of the eld concept makes it a prime candidate for encapsulation in software. A scalar eld class constructed to model thermodynamic
pressure could be reused to model absolute temperature or material density. Furthermore, one can imagine an ADT hierarchy in which scalar eld objects become
components of a vector eld class.
A similar statement can be made for the particle concept. A neutral particle class
used to model atmospheric aerosols might be reused to develop a charged particle
class for studying electric precipitators. Inheritance might be employed to add a
new charge component to a derived type, while adding a type-bound procedure that
computes electric and magentic eld forces on the particle.
Dening a set of algebraic, integral, and differential operators for application to
elds enables compact representations of most physical laws. Fortran supports userdened operators (C++ overloaded operators), the implementations of which can
evaluate approximations to the corresponding mathematical operators. The results
of expressions containing user-dened operators can be assigned to derived type instances either via a language-dened intrinsic assignment or by a dened assignment
(C++ overloaded assignment operator). We refer to the evaluation of expressions
involving the application of mathematical operators to ADT instances as abstract
data type calculus or ADT calculus.
Supporting ADT calculus has been a recurring theme across numerous scientic software projects over the past decade, including the Sophus library at the
University of Bergen1 (Grant, Haveraaen, and Webster 2000), the Overture project
at Lawrence Livermore National Laboratory2 (Henshaw 2002), and the Sundance3
and Morfeus4 projects that originated at Sandia National Laboratories (Long 2004;
Rouson et al. 2008). In many instances, these projects allow one to hide discrete
numerical algorithms behind interfaces that present continuous ones. As noted on
the Sundance home page, this lets software abstractions resemble blackboard abstractions. Users of these ADTs can write ordinary differential equations (ODEs)
and PDEs formally in a manner that closely resembles their expression informally
in blackboard discussions.
Table 3.1 depicts software abstractions that might correspond to various blackboard ones. The rst row of the software abstraction column declares a variable T,
its time derivative dT_dt, and its Laplacian laplacian_T to each be scalar elds. The
second row provides a notation that might be used to dene boundary conditions.
The third row provides a software notation analogous to the subscript notation frequently used for partial derivatives a time derivative in the case shown. The fourth
row demonstrates forward Euler time advancement. The fth row shows how the
result of the T%t() operator might be calculated. In doing so, it also depicts the full
1
2
3
4

http://www.ii.uib.no/saga/Sophus/
https://computation.llnl.gov/casc/Overture/
http://www.math.ttu.edu/klong/Sundance/html/
http://public.ca.sandia.gov/csit/research/scalable/morfeus.php

3.1 Abstract Data Type Calculus

Table 3.1. Translating blackboard abstractions into software abstractions

Blackboard abstraction

Software abstraction

T = T(x, y, z, t), T/t, 2 T

T(x = 0, y, z, t) = T0
Tt T/t
 t +t
T n+1 = T n + tnn
Tt dt
T n + Tt t
T/t = 2 T
2 T +Txx + Tyy + Tzz

type(field) :: T,dT_dt,laplacian_T
call T%boundary(x=0,T0)
T%t()
T = T + T%t()*dt
dT_dt = alpha*laplacian(T)
laplacian_T = T%xx() + T%yy() + T%zz()

expression of a PDE as it might appear inside an ADT that uses the field class
to represent heat conduction physics. The nal row demonstrates the evaluation of
the PDE RHS by invoking the aforementioned subscript analogy. This expression
would likely appear as the result of a function implemented inside the field class
itself.
Figures 3.13.3 refactor the n heat conduction problem using ADT calculus.
Figure 3.1 offers a selection of tests of forward Euler quadrature:
T n+1 = T n + T/t|tn t
= T n + 2 T n t

(3.1)
(3.2)

or backward Euler quadrature:

T n+1 = T n + T/t|tn+1 t

(3.3)

= T n + 2 T n+1 t
1

Tn
= 1 t 2

(3.4)
(3.5)

where lines 27 and 30 in Figure 3.1 provide the corresponding software expressions.
In the case of implicit time advancement schemes such as backward Euler, the
choice of software abstractions requires considerable thought to preserve clarity and
generality without undue overhead in development time and runtime. One question
regarding line 30 in Figure 3.1 concerns which terms to encapsulate as objects and
which to represent using intrinsic data types. A preference for objects makes the
code easier to maintain: Many details private to the employed classes can be changed
without affecting the client code.
In our implementation, the * and - operators return real arrays, whereas
the .inverseTimes. operator returns an integrable_conductor object. Practical
concerns dominated our choice: The arrays contain more data than would t in
an integrable_conductor, whereas the result of .inverseTimes. contains the right
amount of data for encapsulation in an integrable_conductor, as must be the case
because it produces the RHS that ultimately gets assigned to fin.
The implementation of the .inverseTimes. operator invoked at line 30 also
reects an important efciency concern: Rather than computing a matrix inverse and

59

60

Scientic OOP

matrix-vector product, it performs Gaussian elimination and back substitution. Thus,

the operator name, chosen by direct analogy to equation (3.5), need not correspond to
the actual solution method so long as the two produce equivalent results. The choice
stems from the associated operation counts, with Gaussian elimination and back
substitution possessing a lower operation count than matrix inversion and matrixvector multiplication (Strang 2003).

1
2
3
4
5
6
7
8
9
10
11
12
13
14

Figure 3.1
program integrable_fin_test
use iso_fortran_env
,only : input_unit
use kind_parameters
,only : rkind,ikind
use initializer
,only : linear
use problem_class
,only : problem
use integrable_conductor_class ,only : integrable_conductor
implicit none
real(rkind) ,dimension(:,:) ,allocatable :: I ! identity matrix
real(rkind) ,parameter
:: tolerance=1.0E-06
real(rkind)
:: dt=0.1
! default time step
type(integrable_conductor):: fin
! heat equation solver
integer(ikind)
:: scheme
! quadrature choice
type(problem)
:: specs
! problem specifications
namelist /test_suite/ scheme

15
16
17
18
19
20
21
22

enum ,bind(c)
enumerator forward_euler,backward_euler
end enum
if (.not. get_scheme_from(input_unit)) scheme=forward_euler ! default
specs = problem(input_unit)
fin
= integrable_conductor(specs,linear)
print (a,5g9.2),initial temperature = ,fin%temperature()

23
24
25
26
27
28
29
30
31
32

dt = specs%time_step()
select case (scheme)
case(forward_euler)
fin = fin + fin%t()*dt
case(backward_euler)
I = identity(fin%rhs_operator_size())
fin = ( I - fin%rhs_operator()*dt ) .inverseTimes. fin
case default; stop In main: no method specified.
end select

33
34
35
36
37
38
39
40
41
42
43
44
45

print (a,5g9.2),final temperature

= ,fin%temperature()
if (abs(fin%time_derivative())<tolerance) then
print (2(a,es9.3)),|dT/dt|=,fin%time_derivative(),<,tolerance
print *,In main: test passed. :)
else
print (2(a,es9.3)),|dT/dt|=,fin%time_derivative(),>,tolerance
print *,In main: test failed. :(
end if
contains
logical function get_scheme_from(file)
integer(ikind) ,intent(in) :: file
integer(ikind) ,parameter :: found=0

3.1 Abstract Data Type Calculus

46
47
48
49
50
51
52
53
54

integer(ikind)
:: namelist_status
read(file,nml=test_suite,iostat=namelist_status)
if (namelist_status == found) then
get_scheme_from = .true.
else
get_scheme_from = .false.
end if
rewind(file) ! future reads start from the file head
end function

55
56
57
58
59
60
61
62
63
64

pure function identity(n)

integer(ikind) ,intent(in) :: n
integer(ikind)
:: i
real(rkind) ,dimension(:,:) ,allocatable :: identity
allocate(identity(n,n))
identity = 0.
forall(i=1:n) identity(i,i) = 1.
end function
end program
Figure 3.1. Integrable conductor test.

The structure of line 30 reects a desire to decouple the time integration code
from the specic PDE solved. A more equation-specic form might be:
real(rkind) :: alpha
alpha= specs%diffusivity()
fin= (I - alpha*fin%laplacian_operator()*dt) .inverseTimes. fin
where we assume the existence of a suitably dened type-bound procedure that
returns a matrix-form nite difference approximation to the Laplacian operator. The above implementation would tie the time advancement code and the
integrable_conductor interface to the heat equation. For a similar reason, the main
program obtains its own copy of the time step from the problem object, keeping
the physics model free of any of the details of the quadrature scheme used for time
advancement.
Handling derived type components inside dened operators requires discipline
to perform only those operations on each component that are consistent with the
meaning of the operators. When writing a time-integration formula, this suggests
each component must satisfy a differential equation. For alpha in Figure 3.2, the
effective differential equation is d/dt = 0. For stencil in Figure 3.3, we instead
choose to make it a module variable. Giving it module scope enables its use throughout the Field class without incorporating it into the field derived type. We also
give stencil the protected attribute so only procedures inside the enclosing module can modify it. We recommend this practice for all module variables that lack
the private or parameter attributes. We designate stencil protected in case it
becomes public in a future revision.
Whereas Chapter 2 presented the use of UML diagrams in OOA and OOD,
these diagrams also nd use in reverse engineering existing codes. This need arises

61

62

Scientic OOP

when handed legacy software without design documents. It also arises with new software for which the development path appears sufciently straightforward to proceed
to OOP without going through OOA and OOD as might occur in ADT calculus, given the close correspondence between the program syntax and the high-level
mathematical expressions being modeled.
Even in the above cases, UML diagrams can illuminate subtle aspects of the software structure. Figure 3.4 depicts two new types of relationships in a class diagram
reverse engineered from the code in Figures 3.23.3. Whereas the composition and
inheritance relationships in Chapter 2 could be discerned simply by looking at the
source-code denitions of the corresponding derived type denitions, the relationships in Figure 3.4 require deeper inspection of the code. The dashed arrows drawn
to the Problem class denote dependencies that result from argument passing. The
solid line drawn between the Differentiator and Field classes indicates the use of
an object that is neither aggregated explicitly nor inherited implicitly: It has module
scope.
UML diagrams can also explain software behavior. Important behavioral aspects
of ADT calculus relate to the depth of the resulting call trees and the uidity with
which the trees change with the addition or rearrangement of terms in the generating
formulae. The UML sequence diagrams in Figure 3.5 depict the calling sequences
for the forward and backward Euler time-integration formulas at lines 27 and 30 of
Figure 3.1. Following the message-passing interpretation of OOP, one can read a
sequence diagram from top to bottom as a series of messages passed to the objects
in a simulation.
The evaluation of the forward Euler expression fin + fin\%t()*dt involves
rst sending the message t() to fin instructing it to differentiate itself with respect
to time. This results in fin sending the message xx()" to its field component
T_field, which in turn creates a new, temporary Field object. This new Field initializes its allocatable array component by sending the message laplacian() to the
Differentiator object stencil. Subsequently, the temporary Field invokes the division operator on itself represented as a loopback in the sequence diagram. Finally,
fin relays its multiplication and addition messages * and + to T_field to complete
formation of the RHS of the forward Euler expression.
Figure 3.5(b) diagrams the sequence of calls in the evaluation of the backward Euler expression (I-fin\%rhs_operator()) .inverseTimes. fin. Upon
receiving the message rhs_operator(), fin queries its T_field component for its
second-order nite difference matrix. T_field in turn requests the corresponding
matrix from stencil. Because the compiler handles the ensuing intrinsic array subtraction operation, the sequence diagram does not show this step. Finally fin invokes
the .inverseTimes. operator on itself by relaying the same message to its T_field
component.
The relationship between scientic research and ADT calculus is not unique:
ADT calculus could nd use outside science (in nancial mathematics for example)
and other programming paradigms nd use in science. Nonetheless, it appears to
be the one major, modern programming idiom that originated amongst scientic
programmers. Thus, when we refer to Scientic OOP (SOOP), we refer to ADT
calculus. As a novel paradigm, SOOP warrants its own analysis and scrutiny. The
next section provides such analysis and scrutiny.

3.1 Abstract Data Type Calculus

Coming Soon to a Screen Near You!

With ADT calculus, your screen becomes your blackboard. Writing a high-level
language resembles using a symbolic runtime environment such as Mathematica,
Maple, or Mathcad. The techniques described in this chapter demonstrate how
to support the natural syntax of these packages in your own source code.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34

Figure 3.2
module integrable_conductor_class
use kind_parameters ,only : ikind ,rkind
use field_class
,only : field, distribution
implicit none
private
public :: integrable_conductor
type integrable_conductor
private
type(field) :: T_field ! temperatures
real(rkind) :: alpha
! thermal diffusivity
contains
procedure
:: t
procedure
:: rhs_operator
procedure
:: rhs_operator_size
procedure
:: temperature
procedure
:: time_derivative
procedure ,private
:: product
generic
:: operator(*) => product
procedure ,private
:: total
generic
:: operator(+) => total
procedure ,private ,pass(rhs) :: inverseTimes
generic
:: operator(.inverseTimes.) => inverseTimes
end type
interface integrable_conductor
procedure constructor
end interface
contains
type(integrable_conductor) function constructor(spec,T_distribution)
use problem_class ,only : problem
type(problem) :: spec ! problem specification
procedure(distribution) T_distribution
constructor%T_field = field(spec,T_distribution)
constructor%alpha
= spec%diffusivity()
end function

35
36
37
38
39

pure integer(ikind) function rhs_operator_size(this)

class(integrable_conductor) ,intent(in) :: this
rhs_operator_size = this%T_field%field_size()
end function

40
41
42
43
44
45

function rhs_operator(this)
class(integrable_conductor) ,intent(in) :: this
real(rkind) ,dimension(:,:) ,allocatable :: rhs_operator
rhs_operator = this%T_field%xx_matrix()
end function

63

64

Scientic OOP
46
47
48
49
50
51

function temperature(this)
class(integrable_conductor) ,intent(in) :: this
real(rkind) ,dimension(:) ,allocatable :: temperature
temperature = this%T_field%nodal_values()
end function

52
53
54
55
56
57
58

pure type(integrable_conductor) function product(this,factor)

class(integrable_conductor) ,intent(in) :: this
real(rkind)
,intent(in) :: factor
product%T_field = this%T_field * factor
product%alpha
= this%alpha
* factor
end function

59
60
61
62
63
64
65

pure type(integrable_conductor) function total(lhs,rhs)

class(integrable_conductor) ,intent(in) :: lhs
type(integrable_conductor) ,intent(in) :: rhs
total%T_field = lhs%T_field + rhs%T_field
total%alpha
= lhs%alpha
+ rhs%alpha
end function

66
67
68
69
70
71
72

pure function t(this) result(d_dt)

class(integrable_conductor) ,intent(in) :: this
type(integrable_conductor)
:: d_dt
d_dt%T_field = this%T_field%xx()*this%alpha
d_dt%alpha
= 0.0_rkind
end function

73
74
75
76
77

pure real(rkind) function time_derivative(this)

class(integrable_conductor) ,intent(in) :: this
time_derivative = this%alpha*this%T_field%xx_boundary()
end function

78
79
80
81
82
83
84
85

type(integrable_conductor) function inverseTimes(lhs,rhs)

real(rkind) ,dimension(:,:) ,allocatable ,intent(in) :: lhs
class(integrable_conductor)
,intent(in) :: rhs
inverseTimes%T_field = lhs .inverseTimes. rhs%T_field
inverseTimes%alpha = rhs%alpha
end function
end module integrable_conductor_class
Figure 3.2. Integrable conductor class implementation.

1
2
3
4
5
6
7
8
9
10

Figure 3.3
module field_class
use kind_parameters
,only : ikind, rkind
use differentiator_class ,only : differentiator
implicit none
private
abstract interface
pure function distribution (x) result(ret)
import
real(rkind), intent(in) :: x
real(rkind) ret

3.1 Abstract Data Type Calculus

11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50

end function
end interface
type field
private
real(rkind) ,allocatable ,dimension(:) :: node ! internal points
contains
procedure
:: nodal_values
procedure
:: field_size
procedure ,private :: product
generic
:: operator(*) => product
procedure ,private :: ratio
generic
:: operator(/) => ratio
procedure ,private :: total
generic
:: operator(+) => total
procedure ,private ,pass(rhs) :: inverseTimes
generic
:: operator(.inverseTimes.) => inverseTimes
procedure
:: xx
! 2nd-order spatial derivative
procedure
:: xx_boundary ! 2nd-order boundary derivative
procedure
:: xx_matrix
! matrix derivative operator
end type
interface field
procedure constructor
end interface
public :: field, distribution
type(differentiator) ,protected
:: stencil
integer(ikind)
,parameter
:: end_points=2
real(rkind) ,dimension(end_points) :: boundary
contains
type(field) function constructor(spec,sample)
use problem_class ,only : problem
type(problem) :: spec
procedure (distribution) sample
real(rkind)
:: dx
integer(ikind) :: i
stencil = differentiator(spec)
dx
= spec%spacing()
allocate(constructor%node(spec%nodes()))
forall(i=1:spec%nodes()) constructor%node(i) = sample(i*dx)
boundary = spec%boundary_vals()
end function

51
52
53
54
55

pure type(field) function xx(this)

class(field) ,intent(in) :: this
xx%node = stencil%laplacian(this%node,(/boundary(1),boundary(2)/))
end function

56
57
58
59
60
61
62

pure real(rkind) function xx_boundary(this)

class(field) ,intent(in) :: this
type(field)
:: this_xx
this_xx = this%xx()
xx_boundary = this_xx%node(1)
end function

63
64
65

pure function xx_matrix(this)

class(field) ,intent(in) :: this

65

66

Scientic OOP
66
67
68

real(rkind) ,dimension(:,:) ,allocatable :: xx_matrix

xx_matrix = stencil%lap_matrix()
end function

69
70
71
72
73

pure integer(ikind) function field_size(this)

class(field) ,intent(in) :: this
field_size = size(this%node)
end function

74
75
76
77
78
79

pure function nodal_values(this)

class(field) ,intent(in) :: this
real(rkind) ,allocatable ,dimension(:) :: nodal_values
nodal_values = this%node
end function

80
81
82
83
84
85

pure type(field) function ratio(this,denominator)

class(field) ,intent(in) :: this
real(rkind) ,intent(in) :: denominator
ratio%node = this%node/denominator
end function

86
87
88
89
90
91

pure type(field) function product(this,factor)

class(field) ,intent(in) :: this
real(rkind) ,intent(in) :: factor
product%node = this%node*factor
end function

92
93
94
95
96
97

pure type(field) function total(lhs,rhs)

class(field) ,intent(in) :: lhs
type(field) ,intent(in) :: rhs
total%node = lhs%node + rhs%node
end function

98
99
100
101
102
103
104
105
106
107

function inverseTimes(lhs,rhs)
use linear_solve_module ,only : gaussian_elimination ! See Appendix A
real(rkind) ,dimension(:,:) ,allocatable ,intent(in) :: lhs
class(field)
,intent(in) :: rhs
type(field)
,allocatable :: inverseTimes
allocate(inverseTimes)
inverseTimes%node = gaussian_elimination(lhs,rhs%node)
end function
end module field_class
Figure 3.3. Field class implementation.

3.2 Analysis-Driven Design

One expects programs that employ ADT calculus to exhibit a degree of elegance in
some sense, which naturally generates two fundamental questions: to what degree
and in what sense? This section presents steps toward answering these questions.
The stated aim requires developing quantitative, analytical ways to describe
scientic software architecture. Attempts to do so in the broader software engineering community have met with some controversy (Fenton and Ohlsson 2000;
Schach 2002; Shepperd and Ince 1994). Much of the controversy swirls around the

3.2 Analysis-Driven Design

Integrable_conductor
<<constructor>>
integrable_conductor(Problem,Procedure)

Field
Differentiator

<<constructor>>
field(Problem,Procedure)

Problem

Figure 3.4. SOOP heat conduction class diagram.

validity and utility of measurable properties of software, or design metrics, when

employed as
1. Fault rate predictors,
2. Programmer productivity evaluators, and
3. Development time estimators.
For example, in programs with modular architectures, one might expect fault
rates to increase monotonically with module size; yet, as noted in Section 1.2,
Fenton and Ohlsson detected nearly constant fault rates independent of module size
in their study of two releases of a major telecommunications software project. In fact,
the only signicant departure from this behavior occurred at the smallest module size,
where the fault rates dropped precipitously in one release and spiked upward in the
subsequent release. They hypothesized that the latter behavior stemmed from the
increased ratio of interface content to body content for the smallest modules, so that
complexity in the body had been pushed into the interface.
As tools for evaluating productivity, design metrics suffer from potential manipulation by programmers and misinterpretation by their managers. To use the
simplest example, were one to use the much maligned source lines of code (SLOC)
as the measure of productivity, even novice programmers could manipulate the result by inserting dead code or splitting steps that could easily be combined. On the
managers side, a tendency to task the more experienced developers with the more
difcult problems might yield the lowest SLOC for the most adept people.
The use of design metrics to estimate development time provokes possibly the
most pessimistic backlash. For example, Lewis (2001) argued that neither a programs size nor its complexity can be objectively estimated a priori and likewise for its
development time and the absolute productivity of its developer. Lewis backed these
claims with algorithmic complexity theory. The basic argument regarding program
size and complexity follows: For a string s, the associated algorithmic complexity,
or Kolmogorov complexity, K(s), dened as the shortest program that produces s,
is provably noncomputable. Taking s to be the desired program output suggests
that the smallest program (including input data) that produces this output cannot be
estimated.
Alternatively, easily derived upper bounds on the minimum program size, such
as the size of a program that simply prints the output string, do not prove useful. The
lack of useful, computable bounds on program size implies a similar statement for
development time because it depends on program size as well as a similar statement

67

68

Scientic OOP
fin:integrable_conductor
fin + fin%t()*dt

T_field:field

stencil:differentiator

t()
xx()

<<create>>

:field
laplacian()

operator(/)

operator(*)
operator(*)
operator(+)
operator(+)

(a)

fin:integrable_conductor

T_field:field

stencil:differentiator

(I rhs_operator()*dt)
.inverseTimes. fin

rhs_operator()
xx_matrix()

lap_matrix()

inverseTimes(Ihs,rhs)

inverseTimes(Ihs,rhs)

(b)
Figure 3.5. Sequence diagrams for the two quadrature schemes in Figure 3.1: (a) forward
Euler and (b) backward Euler.

3.2 Analysis-Driven Design

for programmer productivity as measured by program size divided by development

time. Ultimately, Lewis concluded that even approximate estimations cannot be
constructed theoretically and must therefore be judged purely on empirical grounds.
None of the above precludes calculating algorithmic complexity, or other formal
properties of software, within restricted domains wherein certain program properties can be guaranteed. For example, Chaitin (1996) demonstrated how to run
algorithmic information theory on a computer by restricting the program to being
expressible in so-called pure LISP. In this context, one assumes programs to be
self-delimiting that is, to contain information about their lengths. In such a system,
one can construct successively sharper lower bounds on a programs algorithmic
complexity.
Somewhat ironically, Kirk and Jenkins (2004) provided empirical evidence of
the usefulness of the very entity, algorithmic complexity, that Lewis employed in
arguments against theoretical estimates. They took the concatenation of a compressed piece of source code and the compression software that produced it as an
upper bound on K(s). Furthermore, they used code obfuscators to estimate the
worst-case complexity and compared this with the complexity added by the developer during a revision cycle. Obfuscators alter source code in ways that make it more
difcult for humans to understand and more difcult to reverse engineer. Using the
compression software bzip25 , they estimated the information content of obfuscated
les at varying levels of obfuscation. Based on their results, they proposed analogies
between their information measures and thermodynamic entropy. They further suggested that these analogies could lead to studying material-like phase transitions as
software becomes more brittle through stronger coupling between modules.
This section builds on the complexity discussions in Chapters 1 and 2 by exploring
three avenues for analyzing ADT calculus: object-oriented design metrics, computational complexity theory, and information theory. Each subsection analyzes the
problem of time advancement in the context of the forward and backward Euler
evaluation lines in Figure 3.1. Subsection 3.2.1 employs two design metrics that
appear frequently in the OOP literature: coupling and cohesion. Subsection 3.2.2
explores scaling arguments and a more realistic bug search time estimate than that
considered in the complexity sections of Chapters 12. Finally, subsection 3.2.3 develops two information-theoretic arguments that quantify another development-time
process: developer communications.
3.2.1 Design Metrics
The proponents of design metrics aim to improve software engineering practices by
encouraging code designs that are coherent, exible, and robust. In many regards,
metrics share these purposes with ADT calculus. Not surprisingly then, the implementation of ADT calculus tends to foster trends in certain metrics. Specically,
ADT calculus implementations tend to exhibit high cohesion within classes and low
coupling between classes.
The term cohesion describes the extent to which a classs procedures relate to
each other in purpose. Stevens et al. (1974) ranked various types of cohesion from
5 http://www.bzip.org

69

70

Scientic OOP

weakest to strongest. Coincidental cohesion, the weakest form, occurs when parts of
a system have no signicant relationship to each other. Functions in classical, procedural mathematics libraries exhibit coincidental cohesion in that most procedure
pairs have in common only that they evaluate mathematical functions. Functional
cohesion, the strongest form, occurs when each part of a system contributes to a
single task.
In the discrete time-advancement schemes, each of the operators and type-bound
procedures contributes to time advancement. This guarantees functional cohesion
at least in a minimal ADT that implements only the requisite calculus.
The term coupling describes the extent to which different ADTs depend
on each other. As with cohesion, one can rank types of coupling from loosely
coupled to tightly coupled. Authors typically rank data coupling as the loosest
form (Pressman 2001; Schach 2002). Data coupling occurs when one part of a system depends on data from another part. Control coupling, a moderately tighter form,
arises when the logical ow of execution in one part depends on ags passed in from
another part. Content coupling, which typically ranks as the tightest form of coupling (Pressman; Schach), occurs when one violates an abstraction by providing one
ADT with direct access to the state or control information inside another ADT. An
especially egregious form of content coupling can occur when it is possible to branch
directly into the middle of a procedure inside one ADT from a line inside another.
Because Fortran requires dened operators to be free of side effects (all arguments must have the intent(in) property), a developer can write the time
advancement expressions in Figure 3.1 without any information about the state or
ow of control inside the objects and operators employed. ADT calculus thus allows
for strict adherence to data privacy, which precludes content coupling.
Furthermore, depending on the design of the ADT calculus, one can write discrete time advancement expressions in which nearly all the procedure arguments are
instances of the class being advanced. This is the case for the forward Euler algorithm in Figure 3.1. The arguments arise implicitly when the compiler resolves the
statement
fin = fin + fin%t()*dt
into a procedure call of the form
call assign(fin,total(fin,product(t(fin),dt)))
where total(), product() and t() are the dened addition, multiplication, and
time differentiation procedures, respectively, in Figures 3.23.3 and where assign()
could be a suitably dened assignment procedure.6 Thus, in the absence of any ags
internal to the derived type itself, no control coupling arises when employing an
ADT calculus. This leaves only the loosest form of coupling: data coupling. At least
in terms of the top-level semantics, even data coupling is kept to a minimum given
that all operators are either unary or binary, so at most two arguments can be passed.
6 We explicitly reference the dened assignment here for clarity. In practice, Fortrans intrinsic as-

signment sufces unless the derived type contains a pointer component, in which case one needs to
dene an assignment if one desires a deep copy of all the RHS data into the LHS object. Otherwise, the intrinsic assignment executes a shallow copy in which only a pointer assignment occurs,
resulting in the LHS and RHS pointer components sharing a target.

3.2 Analysis-Driven Design

Whereas cohesion is essentially qualitative, coupling is more easily quantied.

Following Martin (2002), we dene an ADTs afferent couplings (Ca) as the number
of other ADTs that depend on it, whereas its efferent couplings (Ce) are the number
of ADTs on which it depends. Its instability
I

Ce
Ce + Ca

(3.6)

vanishes when modications to other ADTs do not impact the ADT in question (a
completely stable situation) and approaches unity when changes to any other ADT
can potentially affect the one in question (a completely unstable situation). In an
exercise at the end of this chapter, we ask the reader to calculate the instability of
the packages in our various n models. Chapter 8 discusses the instability metric in
the much richer context of multiphysics modeling.
Building up the instability metric from the afferent and efferent couplings represents a subtle shift from measuring properties of software to measuring properties
of the software development process. Although one could argue that the couplings
themselves have little importance in static code, it would be hard to downplay their
potential impact on the code revision process. The instability metric captures that potential impact. The next two subsections also focus on development-time processes:
bug searches and developer communications.

3.2.2 Complexity Theory

Chapter 1 analyzed a bug search algorithm, estimating that the average number of
lines searched to nd all bugs in a traditional, procedural code grows quadratically
in for code with the lines operating on a globally shared data set. The bug search
algorithm considered in Chapter 1 has the virtue of being easy to analyze but the
vice of not resembling what most programmers do. The following bisection method
takes us a step closer to how developers actually debug code:
1. Bracket the suspected offending code with statements that verify the satisfaction
of necessary constraints. An appropriate set of constraints will be satised at
the beginning of the code segment (as indicated by the true conditions in
Figure 3.6) but not satised at the end (as indicated by the false conditions in
Figure 3.6).
2. Bisect the code segment and insert constraint checks at the midpoint.
3. Form the next segment from the code between the midpoint and the end point at
which the result is the opposite of the midpoint result. In Figure 3.6, for example,
the verication passes at the rst midpoint, so the next segment lies between the
midpoint and the end point.
This process can be repeated until the segment is reduced to a single line of code.
That line is the rst place where the constraint checks fail.
The bisection search draws inspiration from the analogous method for nding
a root of a function of a single variable. In that setting, a necessary and sufcient
condition for convergence to a particular root is the existence of exactly one sign
change in the initially bracketed interval. That sign change indicates the presence of
a root.

71

72

Scientic OOP
True

Source code listing

True

2
1

False

False

Figure 3.6. Bisection method bug search.

The corresponding condition for the bisection bug search is the presence of
exactly one place where the constraint statements change from truth to falsehood.
A sufciently tight bracketing of the bug in a chronological listing of executed lines
would likely sufce. The number of lines n bracketed at the nth iteration of a bisection search provides as a measure of the maximum distance between the lines
being tested (the end points of the nth segment) and the bug. Because the bisection
methods cuts the size of the interval in half at each iteration, the maximum distance
from the bug at iteration n is
n = 0 /2n

(3.7)

where 0 represents the initialize segment size.

The search converges when the code segment reaches a single line, which
happens after
n = log2 0

(3.8)

iterations. At the beginning of the bisection search (n = 0), two lines are checked,
those being the rst and last lines of the suspect code. At each subsequent iteration,
one new line is checked, that being the midpoint line. Thus, the search for all r bugs
terminates after
searched = r(2 + log2 0 ) r(2 + log2 )

(3.9)

lines, or asymptotically O( log2 ) lines maximum.

Following the arguments in section 2.7, the number of unique lines tested in
the course of the search is very likely bounded by the size of the class that contains
the problematic data: 0 c for a class with c lines sharing access to private data.
The authors experiences suggest this inequality holds in ADT calculus. Although
occasional excursions outside the class in question aid in understanding the ow
of control and ensure that the client code supplies correct data to the class, the
excursions prove brief and straightforward due to the simplicity of the expressions

3.2 Analysis-Driven Design

that generate the call tree and due to these software expressions close resemblance
to the blackboard expressions from which they derive. Hence
searched rc(2 + log2 c)

(3.10)

where the inequality must be considered approximate.

An even more exciting feature of ADT calculus appears when we factor c into
the product of the number of procedures p in the suspect class and the classs average
procedural line density c/p. Two reasons motivate this factoring. First, reviewing
the forward Euler expression
fin = fin + fin%t()*dt
suggests that the number of procedures one must implement to advance any ADT
over time would be the same as those dened for fin. These procedures number
three to four: the addition operator (+), the time differentiation procedure (t()), the
multiplication operator (*), and, if necessary, a dened assignment (=). Chapter 8 describes a strategy for building a multiphysics package in which one can time-advance
a coupled group of single-physics ADTs using a single expression analogous to the
forward Euler expression for fin. With ADT calculus, the same three procedures
must be dened for each single-physics abstraction. Thus, p remains constant as the
size of the software project grows to incorporate new physics.
Second, the algorithm implemented inside the aforementioned procedures is
often the same for each ADT. The * operator multiplies an objects state vector by a
constant, while the + operator adds the state vectors of two objects, and = copies an
objects state vector. Because the algorithm is the same for each class, the number of
lines required to express the algorithm will be very nearly the same. Hence, remains
very nearly constant across the project. This leads to rewriting equation (3.10) as
searched rp[2 + log2 (p)]

(3.11)

and to the assertion that the maximum number of lines searched remains very nearly
constant independent of the size of the overall project.
The above analysis focuses on testing for the presence of a bug during program
execution. This would appear to neglect the possibility of errors in the input data.
However, sufciently restrictive validation constraints should detect such errors during the two initial tests that start the bisection search namely the one on the rst
executed line.

3.2.3 Information Theory

In his seminal paper on information theory, Shannon (1948) quantied the information content in telecommunication signals. It seems reasonable then to build on
his foundation to describe the content in other types of communications such as
communications between developers. Of interest here is the amount of information
developers communicate in supplying new classes that extend the capability of a multiphysics framework as described in the previous subsection. In OOP, the minimum
information each developer must provide is the interface to the class that developer contributes. In the time-integration problem, that interface contains a derived

73

74

Scientic OOP

type along with the aforementioned mathematical operators, time differentiator, and
dened assignment.
Shannon reasoned as follows about the set of all possible messages that can be
transmitted between two locations (two developers in the present case):
If the number of messages in the set is nite then this number or any monotonic function
of this number can be regarded as a measure of the information produced when one
message is chosen from the set, all choices being equally likely.

The underlying notion here is that communication hinges on novelty. If only one
message is possible, the transmission always contains that message and no new information arrives. Imagine an analog electrical transmission line on which the voltage
remains xed. Such a line cannot be used to transmit information. It follows that the
degree to which the voltage can vary, and likewise the number of possible states into
which the line can be placed, the more novel is the choice of a particular voltage, the
more information can be conveyed.
Shannon went on to choose the logarithm as the monotonic function because it
satises several conditions that match our intuitive understanding of information. For
example, consider a digital transmission line that can transmit one of two messages,
say 0 or 1. Adding two identical new lines would cube the number of states available
to the collective system: There would be 23 possible messages. Intuitively, however,
one expects the amount of information to triple in going from one line to three. The
logarithm rescales the number of states to match our intuition: log2 23 = 3 log2 2 = 31,
where the base of the logarithm determines the size of one unit of information. In
digital systems, the natural unit is the binary digit, or bit.
More generally, Shannon dened information entropy, a measure of information
content, as
N

H
pi log pi
(3.12)
i=1

where pi is the frequency of occurrence of the ith token (e.g., a character or keyword)
in some large sample representative of the domain of interest, and where N is the
number of unique tokens. In Shannons entropy, log pi represents the amount of information communicated in each occurrence of the ith token. In probability theory,
summing the product of a quantity and its probability over the entire space of probabilities provides its expected value or mean. Taking all choices (tokens) as equally
likely leads to pi = 1/N and therefore
H = log N

(3.13)

which returns us to the simpler denition: the logarithm of the number of possible
messages.
To relate these concepts to source code, consider that the information received
over time on a transmission line can also be spread over space in a program source
le. Thus one draws an analogy between the probability of occurrence of a token
in a signal and its probability of occurrence in a program. One can measure the
probability by calculating the tokens frequency of occurrence in a signal (or program
le) sufciently long to provide accurate statistics.

3.2 Analysis-Driven Design

Considering the information entropy of an ADT calculus elucidates how the

calculus simplies developer communications. First, consider that the class interfaces
comprise the minimum amount of information the developer of each class must
communicate to users of that class. In ADT calculus, wherein each class that supports
the calculus must implement the same operators, the class interfaces exhibit a high
degree of repetition. To see this, imagine extending our heat conduction solver to
include one-dimensional convection via the PDE
2T
T

= (Tu ) + 2
t
x
x

(3.14)

where u u/(c) is a constant ow speed divided by the material density and specic
heat capacity. To enable forward and backward Euler time advancement, we might
build the new type as follows:
module integrable_fluid_ADT
implicit none
private
public :: integrable_fluid
real(rkind) ,parameter :: u_star =1.0_rkind
type ,extends(integrable_conductor) :: integrable_fluid
private
contains
procedure :: t
procedure :: rhs_operator
procedure ,private ,pass(rhs) ::
inverseTimes
generic :: operator(.inverseTimes.) => inverseTimes
end type
interface integrable_fluid
procedure constructor
end interface
contains
! implementation omitted
end module
where we have overridden three of the parent types type-bound procedures: t,
rhs_operator, and inverseTimes. Two compute the new time derivative and the
RHS operator matrix. A third solves the linear system, allowing for a more sophisticated Gaussian elimination process, for example, one with pivoting, due to potential
changes in the matrix properties. The new type inherits the parent types remaining
type-bound procedures.
The character/keyword patterns that recur in the integrable_conductor and
integrable_fluid types would occur in any ADT designed for forward and backward Euler time advancement. Considering each character as a token yields high
pi and low H when adding integrable_fluid type to the package. This slow
information growth reects in very succinct human communications between programmers. The test writer simply informs the developers of the required operators.
The developers provide interfaces supporting those operators.

75

76

Scientic OOP

Taking larger groups of characters as the tokens yields even lower H. At one
extreme, the entire listing of integrable_conductor type-bound procedures from
the contains to the end type could be one token. This token would recur across
all ADTs that support the desired calculus without inheriting any procedure bindings. The resulting low H values would indicate the possibility of very high le
compression, the signicance of which we consider next.
The dependence on the token set demonstrates the nonabsolute nature of information entropy H. Kolmogorov complexity K provides a more general quantication
of information in that it does not depend on a probability model (which depends on
the choice of tokens). Although the cost of this generality lies in the noncomputability
of K, estimating K by compressing les links K to H, which is computable. Shannons
relative entropy provides the link:
The ratio of the entropy of a source to the maximum value it could have while still
restricted to the same symbols will be called its relative entropy. This is the maximum
compression possible when we encode into the same alphabet.

Thus, the best available compression software yields the tightest estimate to the
Kolmogorov complexity as well as the best estimate of the relative entropy.7
Equation (3.12) is formally analogous to the thermodynamic entropy of statistical
mechanics, wherein it describes the information one obtains by determining the
microscopic state of a collection of matter (as determined by the quantum states of
its constituent parts) when it is in a particular macroscopic state. The thermodynamic
limit treats all microscopic states as equally likely so H = k ln N, where the Boltzmann
constant k facilitates converting the logarithm to the natural base and thereby sets
the unit of information.
Kirk and Jenkins (2004) took the thermodynamic analogy further. They estimated the entropy of Java byte code. They also showed that other properties of
the compressed code behave like other thermodynamic variables such as volume,
temperature, and pressure. They plotted these variables against each other for an
ensemble of student projects. We will not pursue an empirical study of source code
entropy in the current text, but an exercise at the end of Chapter 8 presents a very
simple, quantiable entropy measure.
The Audacity of Hope
Pessimism abounds regarding the utility, and even the possibility, of reasoning
quantitatively about software properties. Yet heuristic arguments with varying
degrees of sophistication suggest properties of the software development process
are amenable to analysis. OOD metrics suggest ADT calculus makes packages stable under revision. Complexity theory suggests it renders bug search times nearly
independent of project size. Information theory suggests it reduces the required
developer communications by limiting the growth in the interface information
content incurred when adding new abstractions.
7 Encapsulation and information hiding also reduce interface content in any OOP project due to the

shorter argument lists they facilitate. SOOP augments these strategy by limiting most argument lists
to one or two for the unary and binary operators, respectively, of an ADT calculus.

3.3 Still More on Style

3.3 Still More on Style

The codes in Figures 3.13.2 present several new constructs along with new structural
and behavioral features. Lines 16-18 in Figure 3.1 present this books rst use of
one of the Fortran 2003 features designed to promote interoperability with the C
programming language. The enum construct assigns distinct integer values to the
names listed in the enumerator clause. The bind(c) attribute ensures that the type
kind parameter is interoperable with C in the sense that the bit representation of the
values corresponds to that which the companion C compiler uses for an enumeration.
Chapter 11 details several additional, useful C interoperability features in Fortran
2003.
Given the depth of the call trees in ADT calculus, one must take care to avoid
inefciencies associated with excessive function invocations. One straightforward
strategy involves collecting large amounts of coarsely partitioned data at the lowest
levels of abstraction (corresponding to the deepest invocations in the call tree) and
pushing computationally intensive calculations down to those same abstractions.
With this strategy, the computational work accomplished inside the most deeply
nested procedures greatly exceeds the cost of the function invocations required to
get there.
As we detail next, the chief penalties in terms of execution speed and memory utilization relate to the separation of operations that might be more efcient if
combined. Consider once more the forward Euler expression:
fin = fin + fin%t()*dt
The aforementioned predilection for coarse-grained data implies the presence of a
sizeable array component inside the time-differentiation result. Considering most
modern processors ability to perform a multiplication/addition pair in a single clock
cycle, separating the scaling of this result by the factor dt from the addition of the
resulting scaled array and the array component inside fin represents a potential
doubling of the number of clock cycles required to evaluate the forward Euler RHS
expression. On a closely related note, if these array components exceed the size of
a processors memory cache, looping through an entire array means retiring results
of the multiplication to main memory before reloading them for the addition.
Surmounting these performance hurdles requires combining operators. This can
happen in three ways. First, in some instances, optimizing compilers can combine
the operators during their interprocedural optimization stage. Second, programmers
can dene combined operators explicitly in the source code. A combined forward
Euler operator might take the form
fin = fin%plus_times(fin%t(),dt)
where plus_times scales the fin\%t() results array component by dt and adds the
scaled array it to the corresponding fin array component in a single loop. A third
approach available in C++ is to write expression templates that teach the compiler
how to resolve the original forward Euler syntax into an invocation of a combined
procedure such as plus_times. As always, the choice of strategy should be informed

77

78

Scientic OOP

by empirical proles of the relative runtime share of various processes in the simulation. On many problems of modest size, the gains will not justify the loss in semantic
elegance and the attendant sacricing of interface clarity.
Optimizing compilers are very likely to be efcient with intrinsic assignment of
RHS objects to the LHS ones. Whereas naive compiler implementations might form
the RHS results array component before copying it into the LHS, more sophisticated compilers often recognize the opportunity to overwrite the LHS in the process
of constructing the RHS result. Doing so, however, might require separating the
multiplication and addition steps, because overwriting the fin component during
the multiplication phase (or likewise during the combined addition/multiplication
phases) generates the wrong result. This subtle interplay between memory management and operation count underscores one of the reasons such optimizations are
best left to the compiler wherever possible.
The issue of creating a temporary object to hold the RHS result represents one
example of the larger issue of how to handle temporaries created as the result of all
of the operators and type-bound procedures in an expression. The compiler knows it
will overwrite the LHS of an assignment, but one cannot safely expect similar insight
regarding the fate of other type-bound procedure results. One strategy involves
writing all ADT calculus procedures such that they return intrinsic types. In the
forward Euler expression, having fin\%t, operator(*), and operator(+) all return
real arrays, exposes a great deal of information to the compiler that can be used
for the kinds of optimizations that might otherwise require combined operators.
This has the drawback, however, of limiting the programmer to the intrinsic data
structures. There would be no way to write a procedure that works only with the
nonzero elements of sparse arrays, for example. Chapter 10 outlines a strategy that
enables users to eliminate most temporaries while retaining control over the data
abstractions.
A common theme in object-oriented software design (the theme of Part II of the
current text) is a preference for aggregation over inheritance. Both technologies involve an instance of one object serving as a component of another. With aggregation,
the programmer explicitly imposes this structure. With inheritance, the compiler creates this structure automatically. In many instances, however, employing inheritance
resembles using a machete to slice butter. Much of the power is wasted if the two
types do not share a signicant portion of their interfaces and the writer of the child
type must override most or all of the type-bound procedures. An even worse case
occurs when some of the procedures in the parent types interface have no reasonable
meaning in the child types interface.
This dilemma exposes the ambiguities of the has a and is a metaphors for
aggregation and inheritance relationships, respectively. These metaphors are interchangeable in theory. Nonetheless, aggregation often feels more natural in practice.
To wit, saying an integrable_conductor is a field that satises the heat equation
is essentially equivalent to saying an integrable_conductor has a field that satises the heat equation. Yet several of the procedures bound to the field type
would need different interfaces for clarity in the integrable_conductor. It makes
little sense to talk about a conductors second derivative. Furthermore, some of
the procedures exposed in the field interface would prove completely superuous in the integrable_conductor interface. For example, the ADT calculus the

3.4 The Tao of SOOP

integrable_conductor supports for time advancement does not require a division

operator (/).
Figure 3.3 marks this books rst use of the keyword abstract. It plays two roles
in Fortran 2003. In an abstract interface such as at line 6 in Figure 3.3, it enables
the declaration of a procedure signature, comprising the procedure arguments attributes, types, and, if it is a function, its return type and attributes. In doing so,
the abstract interface facilitates the use of the dummy procedure argument at line
42 in Figure 3.3. At runtime, one can pass as an actual argument the name of any
procedure with matching argument attributes, types, and return type. As compared
to the Fortran-95-style procedure argument in Figure 2.14, this new form avoids
error-prone, duplicate interface declarations. Even an external procedure could be
substituted for the declared one.
Section 4.2.2 describes the second usage of abstract, wherein it serves as a type
attribute. An abstract type denes a specication to which its extended types must
conform. One can never instantiate an abstract type, and its type-bound procedures
might or might not have default implementations. In lieu of procedure implementations, abstract types often have deferred bindings that specify an abstract interface
to which any extended types type-bound procedures must conform.
A nal salient aspect of the sample code is the aforementioned usage of variables
with module scope at lines 35-37 in Figure 3.3. An alternative would be to include
the relevant variables in the type denition, which would have the chief drawback
of requiring that the variables be initialized or computed in every procedure that
constructs a new field object. These denitions or calculations prove superuous
when all elds share a common state (as with C++ static variables) that does not
change with time. One must take care to remember, however, that module variables
must become type components when different instances need different values as
might be the case with the stencil variables for so-called staggered grid calculations,
wherein different variables might be sampled at different spatial locations.

3.4 The Tao of SOOP

The Chinese symbol for tao most often translates as way or path. Part I of the
current text attempts turn the focus of the scientic programmer from the numerical
algorithms and their runtime cost toward their high-level expression in software and
its development costs. This largely represents a redirection of developers attention
from the properties of a program (such as its accuracy and speed) to the properties
of the programming (such as the complexity of the associated bug search process
and the requirement that process places on developers to clearly communicate their
designs to each other). In other words, Part I shifts attention from the product to the
path.
The Merriam-Webster dictionary offers the following alternative denition8 :
the art or skill of doing something in harmony with the essential nature of the thing

For many hundreds of years, the essence of scientic discourse could be thought of
metaphorically as a collective discussion at the blackboard or on a handwritten page.
8 http://www.merriam-webster.com/dictionary/tao

79

80

Scientic OOP

The advent of scientic computing, however, has largely turned this open dialogue
into sorcery. Individual researchers or small research groups develop codes that are
largely indecipherable to the uninitiated. This has ramications both for the openness
of the scientic dialogue and the ability to independently verify codes and validate
their output.
Via ADT calculus, SOOP empowers researchers to build software constructs
that express the essential nature of the mathematical constructs they represent. In
many respects, the syntax of ADT calculus resembles many commercial, symbolic
computation engines such as Mathematica, Mathcad, or Maple. Although some of
these packages evolved from open versions, their modern incarnations are proprietary. In this regard, ADT calculus aims to bring the nonprofessional programmer
the person who sees scientic insights as their primary product rather than software
behind the curtain to see how their code can more closely resemble the natural syntax
that professional programmers provide them via the aforementioned packages.

EXERCISES
1. The Crank-Nicolson time advancement algorithm applied to the heat equation
takes the form
 


t 2 1
t 2 n
n+1
n
T +
(3.15)
= 1

T
T
2
2
Modify Figure 3.1 to include a third enumeration named crank_nicolson. Add a
case(crank_nicolson) to the select case construct in the same gure to include
a software expression of the form
fin = ( I - fin%rhs_operator()*(alpha*dt/2.) &
.inverseTimes. ( fin + fin%t()*(alpha*dt/2.)
where I is an appropriately sized identity matrix and fin%rhs_operator()
returns a nite difference matrix operator approximating a 1D Laplacian operator. Test the new scheme by dening a new, nonlinear initial condition and
advancing to steady state until the the magnitude of the value returned by
fin%time_derivative() falls below tolerance.
2. Draw a UML sequence diagram to describe the sequence of execution corresponding to the above Crank-Nicolson code.
3. Use the Fortran 2003 pass attribute in the product type-bound procedure
binding in Figure 3.2 to enable the products
fin%rhs_operator()*(dt/(alpha*2.))
and
fin%t()*(alpha*dt/2.)
to be written in the reversed order
(alpha*dt/2.)*fin%rhs_operator()
and

3.4 The Tao of SOOP

(alpha*dt/2.)*fin%t()
respectively.
4. Add a debugging parameter of logical type to the kind_parameters module
(or, for clarity, create a new module) and modify each module in the integrable
conductor model to give it access to this parameter via a use statement. Insert
lines of the form
if (debugging) print *,In (procedure name): start
and
if (debugging) print *,In (procedure name): end
at the beginning of each procedure throughout the model.
(a) Set debugging=.true. and print the calling sequence in the integrable
conductor model.
(b) Print the calling sequence again after reordering the terms in the forward
Euler expression, for example
fin = fin%t()*dt + fin
Try other rearrangements of this expression.
5. Calculate the afferent couplings (Ca), efferent couplings (Ce), and instability (I)
of all classes in the n heat conduction models in Chapters 23. Which approach
appears to be the most stable in terms of the average ADT stability across the
design?

81

PA R T II

SOOP T O N U T S A N D B OL TS

Design Patterns Basics

There is one timeless way of building.

Christopher Alexander

4.1 Essentials
Whereas code reuse played an important role in Part I of this text, design reuse
plays an equally important role in Part II. The effort put into thinking abstractly
about software structure and behavior pays off in high-level designs that prove useful
independent of the application and implementation language. Patterns comprise
reusable elements of successful designs.
The software community typically uses the terms design patterns and objectoriented design patterns interchangeably. This stems from the expressiveness of
OOP languages in describing the relationships and interactions between ADTs. Patterns can improve a codes structure and readability and reduce its development costs
by encouraging reuse.
Software design patterns comprise four elements (Gamma et al. 1995):
1. The pattern name: a handle that describes a design problem, its solution, and
consequences in a word or two.
2. The problem: a description of when to apply the pattern and within what context.
3. The solution: the elements that constitute the design, the relationships between
these elements, their responsibilities, and their collaborations.
4. The consequences: the results and trade-offs of applying the pattern.
Although there have been suggestions to include additional information in identifying a pattern, for example, sample code and known uses to validate the pattern
as a proven solution, authors generally agree that elements 2-4 enumerate the three
essential factors in each pattern. Alexander et al. (1977) referred to a slightly different three-part rule as a relation between a certain context, a problem, and a
solution.
The current text takes a soup-to-nuts approach, providing the three essential
elements followed by complete, compilable Fortran and C++ examples. We choose
simplicity and clarity over speed and semantic power. Sophisticated programmers
85

86

Design Patterns Basics

might exploit more advanced language features to enhance our codes generality,
speed it up, or reduce its resource utilization. We encourage readers to revise and
extend our examples. Publishing versions that employ more advanced programming
techniques, for example, generic programming, or that address the performance
needs of larger scale applications, might launch a much needed dialogue about
scientic software design.
Demonstrating patterns in code completes a cycle. All patterns start out as working designs that developers reused across well-engineered projects without conscious
recognition of their universality. The code came rst. Gamma et al. (1995) distilled
recurring themes out of the resulting source code after observing those designs
positive impact on projects. Thus, going from ubiquitous truths to simple examples
brings us full circle: Software begets patterns beget software. An additional, unique
step the current text adds to the cycle involves theoretical analysis along the lines of
Section 3.2.
Pattern recognition requires perspective. A frequently used program construction in one language might be an intrinsic facility in another. For instance, as pointed
out by Gamma et al. (1995) and demonstrated by Decyk et al. (1997b1998), one
might construct patterns for inheritance, encapsulation, or polymorphism in a procedural language, whereas these exist in object-oriented languages. Thus, the patterns
one sees at a particular level of abstraction reect the programming capabilities
supported at that level.
At the source code level, intrinsic language constructs represent patterns supported by the creators of the language. Thus, whereas patterns enable us to see the
forest in the trees, such a view treats trees as an intrinsic construct. Were one to attempt to construct trees from scratch using designer genes, the tree concept itself
would represent a biological design pattern.
Patterns comprise tried-and-true constructions well worn by the time of their
inclusion in a pattern canon, or pattern language. The subsequent chapters of Part
II present a pattern menagerie for which we make none of the comprehensiveness
claims that might justify calling it a pattern language. Rather, Part II whets the
readers appetite for patterns with a few we have found broadly useful.
Its Easy Being Green
OOD patterns emerge like the independent recurrence of a solution to a common biological problem in multiple evolutionary tree branches. In this sense, they
resemble organic growth in nature. Moreover, as the construction of natural beings, software bears consideration as a product of nature. Copying time-honored
designs that naturally arise in one environment eases the creation of designs that
feel natural in another.

4.2 Foundations
This section sketches the evolution of patterns research from its origins in building
architecture through the identication of patterns in software architecture to their
recent penetration into scientic software architecture.

4.2 Foundations

4.2.1 Building Architecture

As the quote at the beginning of this chapter suggests, patterns hail from time
immemorial. Their modern identication and labeling began with architecture
professor Christopher Alexander1 and his collaborators at the Center or Environmental Structure three-volume series: Volume 1, The Timeless Way of
Building (Alexander 1979); Volume 2, A Pattern Language: Towns, Buildings,
Construction (Alexander et al. 1977); and Volume 3, The Oregon Experiment
(Alexander, Silverstein, Angel, Ishikawa, and Abrams 1975).
Opposite to the numbering of the volumes, the chronological order of publication mirrors the progression outlined in Section 4.1: Practice precedes patterns
precedes theory. The Oregon Experiment describes an architectural planning project
conducted for the University of Oregon. As such, it shows an actual implementation
of the patterns and theory in the other two volumes. A Pattern Language describes
more than 250 archetypal patterns patterns with lasting value, as likely to be used
centuries hence as centuries prior. The Timeless Way of Building (hereafter, Timeless Way) puts forth a theoretical framework and eshes out and contextualizes the
pattern concept.
Timeless Way is at once poetic and philosophical. Alexander intersperses short,
italicized paragraphs with the body of the text. He refers to these as headlines. He
recommends one method of skimming the book by reading each chapters start and
nish along with the headlines in between. Read in this manner, the expressiveness
and brevity of the headlines sometimes give them the feeling of the stanzas of a poem.
The books emphasis on the way evokes Taoism. Its occasional use of seeming contradictions feels reminiscent of Zen Buddhism, wherein koans provoke the
reader to reconcile paradoxes as in the famous koan: Two hands clap and there is
a sound; what is the sound of one hand?2 Likewise, in order to purge ourselves of
the constrictions imposed by other design methods, Alexander urges the reader to
learn a new discipline only to give it up:
To purge ourselves of these illusions, to become free of all the articial images of order
which distort the nature that is in us, we must rst learn a discipline which teaches us the
true relationship between ourselves and our surroundings.
Thence, once this discipline has done its work, and pricked the bubbles of illusion which
we cling to now, we will be ready to give up the discipline, and act as nature does.

This is immediately reminiscent of another Zen Buddhist concept: shoshin, or

beginners mind, an openness to possibilities that might seem closed off to an expert
schooled in the ways of the eld.
Timeless Way concerns itself with a process for releasing ones innate understanding of an unnamed quality that breathes life into buildings and towns and makes them
whole. That process starts with observing patterns of events that occur in these places.
As these patterns unfold over time, so do they in space. Their spatial characteristics
dene characteristic geometries that designers and inhabitants can weave together
into a complete building or town. This dialogue between designers and inhabitants is
1 The rst recipient of a Ph.D. in architecture from Harvard University, Alexander was also awarded

the rst Gold Medal for Research by the American Institute of Architects and is now Professor
Emeritus of Architecture at the University of California, Berkeley.
2 Oral tradition attributes this koan to Hakuin Ekaku, 16861769.

87

88

Design Patterns Basics

one of the hallmarks of the timeless way. It is a dialogue spoken in pattern languages.
Because each person possesses a pattern language unique to the events and cycles of
their own life, facilitating mutual understanding necessitates constructing a common
pattern language.
Several aspects of Alexanders writing strike chords that resonate across many
schools of philosophical and scientic thought, Eastern and Western, ancient and
modern. In stating, To seek the timeless way, we must rst know the quality without a name, he acknowledges that fundamental truths evade precise articulation in
everyday language. A second universal theme is Alexanders useage of dialectics:
He maintains that vibrant architecture can only emerge from an interplay of ideas
between its designers and its intended inhabitants. Pattern languages give the interested parties a set of design elements along with rules for combining those elements to
extract designs from a combinatorial space of possibilities. Languages allow innite
variation in many of the ultimate details while still maintaining thematic consistency.
For instance, Timeless Way outlines a pattern language for stone houses in the South
of Italy and sketches several plans that could be instantiated from the language.
In addition to identifying the fundamental nature of patterns and their importance, Alexander postulates that the distinction between living and dead patterns
lies in the extent to which everyone shares an understanding of them. In cultures
where the patterns are alive, all people share a basic understanding of building. Such
was the case in agricultural societies. Farmers felt comfortable building barns. Home
owners felt comfortable laying out their homes. While this might seem farfetched in
modern times, Alexander points out that as recently as fty years ago in Japan, for
example, every child learned how to lay out a house.
By contrast, the specialization required in industrial societies tends to kill
patterns:
Even within any one profession, professional jealousy keeps people from sharing their
pattern languages. Architects, like chefs, jealously guard their recipes so that they can
maintain a unique style to sell.
The languages start out by being specialized, and hidden from the people; and then within
the specialties, the languages become more private still, and hidden from one another
and fragmented.

In such settings, so much wisdom fails to propagate down through the generations
that even the specialists begin to make obvious mistakes. Alexander gives examples
from his campus where two seminar rooms lack the geometry and lighting to foster vibrant, comfortable discussions. The architects who designed the room forgot
several basic rules of thumb:
Rooms need natural light from two directions.
Windows too low create silhouettes of occupants as viewed by other occupants.
A reective display surface orthogonal to an adjacent window reects the

windows view into its viewers eyes.

One modern response to this situation involves centralized control over urban design.
Centralization, however, fails to generate vibrant buildings and towns due to its
insufcient responsiveness to peoples needs. It exacerbates the very problem it
seeks to solve.

4.2 Foundations

Timeless Way denes architectural design patterns in terms of the context within
which they occur, a conict that arises in that context, and the way in which the
pattern resolves the conict. A sample contextconictresolution triplet reads
communal roomconict between privacy and community alcove opening off
communal room. This three-part logic imparts a degree of empiricism: One can
verify whether the context exists, whether the conict arises, and whether the pattern
resolves the conict. Alexander asserts that this empiricism moves the statement
that a pattern is alive from the realm of taste to the realm of objective veriability.
Alexanders nal step in dening a pattern is naming it. Thus, the aforementioned
triplet becomes the ALCOVE pattern.
Whereas patterns share their empirical nature with physics, they share another
quality with computer science: invariance. As Chapter 10 describes, the formal methods subdiscipline of computer science deals in part with logical statements that must
be true throughout the execution of an algorithm. These are invariants of the algorithm. The search for patterns represents a search for invariant aspects of designs
that can be observed across all valid implementations of the pattern.
Timeless Way advocates precision in specifying patterns. An ENTRANCE TRANSITION pattern describes the junction at which the indoors meets the outdoors.
Examples of precision in dening this pattern include stipulations that doors be 20
feet from the street, that windows facing the transition region not be visible from the
street, that the character of surfaces differ from that in the adjoining spaces, and that
it include a glimpse of something entirely hidden from the street. These stipulations
imbue patterns with a property many consider to be among the most fundamental
in all of science: falsiability. Their precision opens them up to challenge.
The book also acknowledges the elusive nature of even some precisely dened
concepts. In one sense, an entrance transition never exists in its own right but merely
describes a relationship between two other entities: indoors and outdoors. Specically, it describes their intersection. In fact, the transition described might even occur
fully indoors. This points up a paradox Alexander identies in the goal of making designs whole. Wholeness implies limits and self-containment, and yet patterns might
exist solely as relationships or as characteristics of a particular part of a design, say,
the part of the indoors closest to the outdoors.
Lest patterns appear overly formal and sterile, Timeless Way connects them intimately with the realm of social science. The ultimate test of a pattern comes in the
way it makes occupants feel. This exposes one of the reasons patterns vary among
cultures. Nonetheless, within the appropriate culture or subculture, Alexander reports a remarkably high degree of unanimity about the feelings patterns evoke in
people.
Not every concept from Timeless Way maps neatly or obviously to software. This
seems particularly true in its discussion of construction. Nonetheless, contemplating
the relationship between the two elds generates useful insights. Consider Alexanders comparison of the emergence of a design to the growth of an organism from
an undifferentiated clump of cells. Rather than stitching together preformed parts,
the organism comprises a single, whole entity throughout its maturation. Different
organs arise in constant contact with each other. Though innitely more complex,
this process seems reminiscent of a technique the current authors employ in program
construction: We replicate a skeleton ADT multiple times before differentiating it

89

90

Design Patterns Basics

for the specic role it will ultimately play. In Fortran, such a skeleton might take the
form:
module skeleton_ADT
! Manage entity scope/lifetime
use
,only :
! Surgical use of module entities
implicit none
! Prevent implicit typing
private
! Hide everything by default
public :: skeleton
! Expose type/constructor/methods
type skeleton
private
! Hide data
! derived type components
contains
procedure ::
! Public methods
final :: destructor ! Free pointer-allocated memory
end type
interface skeleton
! Generic constructor name
procedure constructor
end interface
contains
type(skeleton) function constructor()
! return new instance
end function
elemental subroutine destructor(this)
type(skeleton) ,intent(inout) :: this
! deallocate any allocated pointer components
end subroutine
end module skeleton_ADT
Such a skeleton is not viable in the sense that it cannot be compiled. It simply
provides a scaffolding containing the minimum number of elements required for
a well-designed ADT. These include the enforcement of explicit typing, default
privacy, modular scoping, and memory management. What it does not contain is
the equivalent of DNA molecules driving a process of self-organization. Only the
programmer can drive that process.
The skeleton ADT contains this books rst elemental procedure. Fortran
requires rank and parameter matches for automatic invocation of nal subroutines.
If a program instantiates an array of skeleton objects, destructor will only be
called if a nal subroutine explicitly accepts an array argument of matching rank or
if destructor is elemental and therefore supports all ranks.
Timeless Way concludes by providing the kernel of the way. True to form, it
does so by pointing out a paradox: By enforcing a discipline based on reconnecting
your design to its living purposes, a pattern language frees you from its very own
strictures. Once you have mastered it, you no longer need it. You can create beautiful
designs by relying on your innate skills now that you have learned to trust them.
Alexander refers to this state of mind as egoless.
Moving from Timeless Way to the second volume, A Pattern Language, one nds
fewer specics of direct utility in software. Yet the structure and manner in which one
uses the book carries over. In A Pattern Language, Alexander et al. (1977) catalog

4.2 Foundations

253 patterns and order them from the largest to the smallest. Each intermediate
pattern helps complete larger patterns that come before it. Each is itself completed
by smaller patterns that come after it. The progression of patterns goes from regions and towns to clusters of buildings, buildings themselves, building subunits, and
construction details.
Each pattern comprises, in order:

A picture of an archetypal example of the pattern.

A paragraph that contextualizes the pattern within larger patterns.
A one- to two-sentence headline summarizing the patterns essence.
The body of the pattern, detailing its empirical background, evidence of its
validity as a solution, and its range of manifestations.
The solution, including the physical and social relationships required to solve
the problem.
A diagram.
A list of smaller patterns required to complete or enhance the current pattern.
Thus, A Pattern Language eshes out each pattern rather completely, providing
much of the information required of a trusted manual for design and construction.
Alexander and his colleagues suggest a methodology for reading and using the
book in an actual project. Scanning the master sequence of patterns, one nds and
reads the largest pattern describing the overall project scope. One then marks on the
master list those smaller patterns delineated at the end of this largest pattern. Next,
one reads the smaller patterns and progressively marks the even smaller patterns
they list. This cascade generates a network of patterns. To prevent unmanageable
growth of this network, the authors suggest skipping the smaller patterns that seem
of doubtful value. However, they suggest inserting your own patterns where no
appropriate one exists in the language.
Alexander et al. (1977) tag their patterns in a way that distinguishes their certainty in each patterns capturing of an invariant that is a property common to
all possible solutions of the problem the pattern addresses. In one set, they believe
they have found an inescapable feature of all well-formed solutions. In a second set,
they believe the pattern gets us closer to this inescapable set but that other possible
solutions almost certainly exist. Finally, in a third set, they express certainty that
they have not found an invariant. They also suggest modifying patterns to suit your
purposes. The resulting pattern network represents a pattern language unique to
your project.
Interestingly, the frontmatter in A Pattern Language makes the aforementioned
connection to poetry concrete. In a section entitled THE POETRY OF THE LANGUAGE,
Alexander and colleagues distinguish poetry from prose by the density of interpretations each writing style affords. Prosaic English most often conveys one,
straightforward meaning. English poems, by contrast, harbor manifold interlocking messages even in single sentences or words. The authors argue that a building
that strings together loosely connected patterns resembles prose whereas one that
overlaps multiple patterns in space captures many interpretations in its nooks. Such
density engenders profundity. As the authors state, compressing many patterns into
the smallest possible space encourages economy and leads to buildings which are
poems.

91

92

Design Patterns Basics

The third volume, The Oregon Experiment, illustrates the application of the
theory expounded by Timeless Way and the structural template laid out in A Pattern
Language to a specic architectural project at the University of Oregon. It would at
rst seem that the more concrete nature of building design and construction would
lend itself to fewer direct comparisons with software design and construction. Yet the
chief force driving the adoption of patterns in one setting mirrors that in the other:
managing growth. The Oregon Experiment arose from the crisis brought about by
rapid growth over a single decade from a century-long size of a few thousand students
to a student body of 15,000. Recognizing that computer programs experienced a
contemporaneous period of rapid increase in scale, leading to what is often termed
the software crisis, it starts to seem less farfetched that the work of Alexander and
colleagues ultimately inuenced software design.
The Oregon Experiment comprises six chapters, each expounding a single principle the authors believe will probably have to be followed in all communities where
people seek comparably human and organic results. The six principles comprise organic order, participation, piecemeal growth, patterns, diagnosis, and coordination.
The authors set the principle of organic order up in opposition to the totalitarian rule
imposed by traditional master plans. Such plans attempt to map out the organization
of a site envisioned for the distant future. Organic order relies instead on a pattern
language adopted by thousands of local projects meeting immediate needs. A critical
driver of organic order is a planning board small enough to keep meetings focused
and balanced enough between users and decision makers such that real, informed
decision making happens in the room.
With the case for user participation already detailed in the rst two volumes, The
Oregon Experiment focuses on a specic example of user participation in designing a
set of new buildings for the School of Music. The embodiment of user participation
takes the form of a user design group. This group drafts pattern-inspired design
schematics before architects enter the project.
Piecemeal growth harbors two important implications. First, it acknowledges the
need for constant repair of existing structures as part of the growth process. Second,
it suggests new pieces be added slowly, in small increments over time. The authors
suggest enforcing piecemeal growth by distributing a capital budget evenly across
budget categories. With the same amount of funds allocated for small projects in
aggregate as for large projects in aggregate, there will be many more small projects
than large ones.
Of the 250 patterns in A Pattern Language, The Oregon Experiment cites 200
relevant to the university. Of these, they nd 160 of appropriate scope for individual user groups to employ those dealing with building rooms and gardens, for
example while 37 are of sufciently large scope they require university-wide, multiproject agreement to complete. These include BIKE PATHS AND RACKS and PROMENADE
patterns, for example. They also identify 18 domain-specic patterns of special interest to universities, for example, CLASSROOM DISTRIBUTION and REAL LEARNING IN
CAFES patterns. In keeping with the empirical philosophy of Timeless Way, the authors suggest planning boards adopt new patterns or revise old patterns only when
the patterns have experimental or observational support.
Diagnosis refers to an annual process of reevaluating which spaces are alive
and which are dead. As with the timeless nature of patterns, diagnosis has deep

4.2 Foundations

roots. Alexander and colleagues provide evidence of annual surveys of 13th-century

Siena, Italy. Laypersons, as opposed to building specialists, populated the survey
committee. The committee drafted plans for specic projects aimed at improving
important public works.
Coordination happens when the planning board considers newly proposed
projects with explicitly articulated relationships to the currently adopted patterns
and diagnosis. The board prioritizes the submitted projects, judges them by the
extent of their conformance to the current patterns and diagnosis, and partitions the
list so that each project competes only with those of a similar size.
As these concepts apply to scientic software, the organic growth and user participation principles suggest scientic software projects might most often succeed
when the code designers are embedded in the organizations that will ultimately use
the code. Piecemeal growth acknowledges the importance of ongoing maintenance
and incremental expansion. The necessity for domain-specic patterns suggests tailoring some abstractions to the scientic domain of interest. The need for diagnosis
encourages gathering users annually to inform the global view of the health of various
parts of the software. We have observed these practices across several of the bestrun scientic software projects. The coordination mechanism, however, frequently
differs depending on the extent to which funding is central or distributed.

4.2.2 Software Architecture

Although University of Oregon alumnus Kent Beck and coauthor Ward Cunningham introduced the design pattern concept to the computer science community in
1987 (Beck and Cunningham 1987), the idea rst gained a substantial following with
the publication of Design Patterns: Elements of Reusable Object-Oriented Software
by Gamma et al. (1995), often referred to as the Gang of Four or GoF. The rst part
of the GoF book denes design pattern in the context of software and demonstrates the utility of patterns via a case study. The second part catalogs general
design patterns of three varieties: creational, structural, and behavioral. Creational
patterns deal with object construction. Structural patterns describe the composition
of collections of classes with the chief compositional mechanism being inheritance.
Behavioral patterns describe class or object interactions and responsibilities.
Much as Alexander and his colleagues did, the GoF employ a literary analogy
for their work, noting:
Novelists and playwrights rarely design their plots from scratch. Instead, they follow patterns like Tragically Flawed Hero (Macbeth, Hamlet, etc.) or The Romantic Novel
(countless romance novels).

As they point out, object-oriented software designers follow patterns as well. Much
like common literary themes, different patterns arise in different genres or domains.
The GoF explained that their book only captures a fraction of the likely knowledge
of expert designers. They made clear they were not attempting to present patterns for
concurrency or distributed programming, for example; nor did they cover patterns
specic to particular application domains such as user interfaces, device drivers, or
databases.

93

94

Design Patterns Basics

Inspired by Alexander, the GoF draw analogies between software architecture and building architecture, comparing objects and interfaces to walls and doors.
Expanding on the three-part rule, the GoF include several additional features of the
patterns they present, including:
Pattern Name and Classication: the name captures the essence of the pattern

and becomes part of the design lexicon, while the classication species the
patterns role (creational, structural or behavioral) and scope (whether they
determine static, compile-time class relationships or dynamic, run-time object
relationships):
Intent: what a pattern does, its rationale, and the problem it solves.
Also Known As: pattern aliases.
Motivation: an archetypal scenario.
Applicability: situational context.
Structure: schematic descriptions of class relationships and interactions.
Participants: classes or objects involved.
Collaborations: the roles of the participating classes.
Consequences: design trade-offs.
Implementation: issues that arise in constructing the solution.
Sample Code: illustrative code fragments.
Known Uses: examples from real systems.
Related Patterns: differences from, and relationships to, other patterns.

While we do not adopt the GoFs complete structure for describing patterns, we
include various elements of this structure in our pattern descriptions.
The GoF identied the key drivers that complicate OOD: encapsulation, granularity, dependency, exibility, performance, evolution, and reusability. They aptly
pointed out that these drivers often oppose each other, sometimes in complicated
multidimensional ways. The desire to reduce dependencies between abstractions
complicates encapsulation. Performance demands often discourage high levels of
granularity in scientic applications where efciency necessitates long loops over
large data sets. Frequent reuse of a given class early in the evolution of a package
might inhibit the packages exibility in later generations.
Two guiding principles reverberate throughout the GoF patterns:
1. Favor object composition over class inheritance.
2. Program to an interface, not an implementation.
Section 3.3 discusses the rst principle. Illustrating the second principle requires introducing a new concept: an abstract class. An abstract class denes the interface that
its child classes must implement. No instances of an abstract class can be instantiated.
In its purest form, an abstract class contains no data and defers the implementation
of all its methods to its child classes. Any class that is nonabstract is termed concrete.
In Fortran and C++, the language mechanisms typically employed to relate abstract classes to concrete ones match those employed to relate parent classes to their
children: type extension in Fortran and inheritance in C++. Whereas a concrete child
class inherits from its parent class any state and behavior that it does not specically
override, no such state and behavior exist if the parent class encapsulates no data and
provides no default procedure denitions. This is the case for pure abstract classes

4.2 Foundations

Abstract_speaker


Astronaut
Figure 4.1. Abstract class realization: Concrete Astronaut realizes, or implements, an
Abstract_speaker interface.

(C++ pure virtual classes). In such a case, one more naturally refers to the child class
as implementing or realizing the interface dened by the abstract parent. UML models a realization relationship with a dashed line adorned by an open triangle at the
interface end, as demonstrated in Figure 4.1. Furthermore, UML denotes abstract
class names in bold italics.
Fortran abstract types facilitate creating abstract classes. Figure 4.2 demonstrates
the denition of an abstract type containing no components, and specifying a deferred
binding speak, leaving the implementation of the corresponding method to extended
types. It constrains those implementations to conform to the abstract interface talk.
Figure 4.2 further depicts the extension of the abstract_speaker type to create an
astronaut with the same purpose as the like-named type in Figure 2.4.
The construction of an abstract type as a base for extension in Figure 4.2(a)
does not enhance the concrete type implemented in Figure 4.2(b). It enhances the
say_something() client code in Figure 4.2(c), enabling it to be written using the
abstract type with knowledge of neither what concrete type actually gets passed
at runtime nor what procedure actually gets invoked on that type. At runtime, any
concrete type that extends abstract_speaker can be passed to say_something. This
exibility is a hallmark of design patterns and explains why abstract class construction
plays a recurring role across many patterns.
Neither Fortran nor C++ formally denes class interfaces in a way that decouples
this concept from a classs implementation. In both languages, however, one can
approximate the class interface concept with abstract classes. The GoF advocate
that all extensions of an abstract class merely dene or redene the abstract methods
of their parent abstract class without dening any new methods. All child classes
so dened share an identical interface. This enables the writing of client code that
depends only on the interface, not on the implementation.
As the GoF pointed out, one must instantiate concrete classes eventually. Creational patterns do so by associating an interface with its implementation during
instantiation, thereby abstracting the object creation. We describe a creational pattern, FACTORY, in Chapter 9. Other GoF patterns we have found useful in our research
include the behavioral patterns STRATEGY and TEMPLATE METHOD, and the structural
pattern FACADE. Each of these is the subject of, or inspiration for, a section in Part
II of the current text. FACTORY METHOD encapsulates object construction for a family
of subclasses. STRATEGY represents a family of interchangeable algorithms. FACADE
presents a unied interface to a complicated subsystem. TEMPLATE METHOD species an algorithm in skeleton form, leaving child classes to implement many of the
algorithms steps.

95

96

Design Patterns Basics

Table 4.1. Design variables supported by patterns
Pattern

Variable

Factory Method
Proxy
Composite
Facade
Iterator
Mediator
Strategy
Template Method

subclass of instantiated object

location or method of accessing an object
object structure and composition
interface to a subsystem
how an aggregates elements are accessed
how and which objects interact with each other
an algorithm
steps of an algorithm

The essence of a pattern often lies in the exibility it facilitates by enabling

variation in design elements. The GoF tabulated the design variations supported by
each of their patterns. Table 4.1 excerpts the list of GoF patterns that appear in or
inspire patterns in the current text.
The GoF outline how this exibility enhances software applications, toolkits,
and frameworks. One might dene an application as any collection of code that
includes a main program. Applications benet from the increased internal reuse,
maintainability, and extensibility that design patterns foster. The GoF dene toolkits as useful classes that user codes call, that is, the OOP counterpart to conventional
subroutine libraries. Toolkits benet from the loose coupling that design patterns
encourage. Such loose coupling makes toolkits more reusable across disparate applications. The GoF dene frameworks as domain-specic system architectures that call
user code. Frameworks benet from the increased consistency and clarity exhibited
by pattern-based project.
Part I of the current text demonstrates application design and construction, focusing primarily on heat conduction. Refactoring that application in terms of design
patterns would allow for varying the numerical methods or the objects that encapsulate those methods or those objects properties, including their external view and
internal structure.
Part II corresponds closely to toolkit design and construction. The designs presented herein can be reused across a plethora of applications. As a natural analog
to library development, for example, the designs in Chapters 67 could be reused to
build a numerical quadrature library.
Part III outlines framework development. Chapter 11 outlines the development
of Fortran interfaces to a C++ solver framework. Chapter 12 sketches the design of
the Morfeus multiphysics Fortran framework.

1
2
3
4

module speaker_class
implicit none
private
public :: say_something

Figure 4.2(a)

5
6
7

type ,abstract ,public :: speaker

contains

4.2 Foundations
8
9

procedure(talk) ,deferred :: speak

end type

10
11
12
13
14
15
16
17

abstract interface
function talk(this) result(message)
import :: speaker
class(speaker) ,intent(in) :: this
character(:) ,allocatable :: message
end function
end interface

18
19

contains

20
21
22
23
24

subroutine say_something(somebody)
class(speaker) :: somebody
print *,somebody%speak()
end subroutine

25
26

1
2
3
4
5

end module
Figure 4.2(b)
module astronaut_class
use speaker_class ,only : speaker
implicit none
private
public :: astronaut

6
7
8
9
10
11
12

type ,extends(speaker) :: astronaut

private
character(:) ,allocatable :: greeting
contains
procedure :: speak => greet
end type

13
14
15
16
17
18
19
20
21
22

interface astronaut ! Map generic to actual name

procedure constructor
end interface
contains
function constructor(new_greeting) result(new_astronaut)
character(len=*), intent(in) :: new_greeting
type(astronaut) :: new_astronaut
new_astronaut%greeting = new_greeting
end function

23
24
25
26
27
28
29

1
2

function greet(this) result(message)

class(astronaut) ,intent(in) :: this
character(:) ,allocatable :: message
message = this%greeting
end function
end module
Figure 4.2(c)
program abstract_hello_world
use astronaut_class ,only : astronaut

97

98

Design Patterns Basics

3
4
5
6
7
8

use speaker_class ,only : say_something

implicit none
type(astronaut) :: pilot
pilot = astronaut(Hello, world!)
call say_something(pilot)
end program
Figure 4.2. Abstract class: (a) abstract speaker, (b) extended type, and (c) driver.

4.2.3 Scientic Software Architecture

Object-oriented design patterns apparently rst entered in the scientic software
literature early in the 21st century (Blilie 2002). Gardner and Manduchi (2007) published the rst book-form coverage of applying OOD patterns in scientic research.3
Their book covers a case study on the design and construction of an open-source
waveform browser for use in data acquisition system on plasma fusion experiments.
Gardner and Manduchi implement their designs in Java to leverage the Java Virtual Machines platform-independence and Javas explicit support for parallelism.
Part 1 of their book outlines a rst implementation of a waveform browser dubbed
PreEScope. Part 2 refactors PreEScope into a pattern-based EScope, itself a simplied version of the 30,000-line, public-domain jScope that is in production use at
several plasma fusion laboratories.
Gardner and Manduchi design a FACADE that provides a unied interface to two
subsystems in EScopes graphical user interface (GUI). One subsystem handles EScopes display and control. The other interacts with the data server. The facade has
two parts: an abstract facade class that client code uses, and a concrete facade class
that interacts with the subsystems. The concrete facade fullls requests the client
submits via the abstract facade.
Gardner and Manduchi employ a MEDIATOR to communicate information between classes within the subsystems. Doing so eliminates the need to trace complicated linkages between classes to understand the communication patterns. All
information passes through the mediator.
Gardner and Manduchi use the TEMPLATE pattern to express in a parent superclass those parts of an algorithm that are common to a collection of subclasses. The
superclass delegates the subclass-specic portions of the algorithm to the subclasses.
They use a TEMPLATE to separate the classes that deal with drawing various parts of
graphs: the waveform, the axes, the labels, and any messages.
Gardner and Manduchi discuss a collection of creational patterns under the
rubric Factory patterns. These include FACTORY METHOD, ABSTRACT FACTORY,
BUILDER, PROTOTYPE, and SINGLETON. Chapter 9 of the current text describes how
3 Given

the importance of parallel programming in high-performance scientic

computing, it is tempting to give precedence to Patterns for Parallel Programming (Mattson, Sanders, and Massingill 2005). Although most of their examples are scientic,
they aim to capitalize on the industry-wide trend toward parallelism and to impact programming
practice beyond science. Furthermore, they do not limit themselves to a specic programming
paradigm. As such, their use of the term pattern refers more broadly to recurring themes in
program design independent of whether those programs are object-oriented.

4.3 Canonical Contexts

to use a FACTORY METHOD to add exibility to the spatial differencing scheme

incorporated into our heat equation solvers in Part I.
Each of the previous mentioned patterns covered by Gardner and Manduchi
has an analog in the current text. The ABSTRACT CALCULUS pattern of Chapter 6
functions as a domain-specic facade. Our implementations of the STRATEGY pattern
in Chapter 7 effectively embed TEMPLATE patterns. Our PUPPETEER pattern plays the
role of a MEDIATOR.
Other GoF patterns that Gardner and Manduchi cover include DECORATOR,
ADAPTER, OBSERVER, ADAPTER, STATE, CHAIN OF RESPONSIBILITY, and ABSTRACT FACTORY. They also propose one new pattern as a variation on a GoF pattern useful in
their application domain: ARTICULATED FACADE. Finally, they touch on concurrent
programming patterns that arose after the GoFs seminal text.
In terms of presenting design patterns in scientic programming, the book by
Gardner and Manduchi sits closer to the current book than any other of which we
are aware. The application they study, however, differs considerably from that of
prime interest herein: multiphysics simulation. Moreover, their language selection
precludes the core technology we feature: ADT calculus. Java neither allows the
overloading of existing operators, as Fortran and C++ do, nor the denition of new
operators, as Fortran allows. Although this makes Java unsuitable for current purposes, the language has much to recommend it. Java stands as the only one of the
three languages that was object-oriented, parallel, and distributed from conception.
It therefore conforms much more closely with OOP philosphy. Nearly every data entity in Java is an object, and even the simplest Hello, world! program is inherently
object-oriented. Furthermore, the multithreading capabilities of Java long preceded
the expected entry of similar facilities in the upcoming Fortran 2008 and C++0x standards. The network socket model intrinsic to Java lacks standard counterparts for
distributed programming in Fortran and C++. Finally, the garbage collection for all
dynamic memory eliminates one of the most vexing challenges encountered by C++
programmers and, to some extent, by Fortran programmers.
Theres a There There
Writing scientic software traditionally comprised no more than expressing algorithms in code. Algorithms describe a codes temporal structure, saying nothing
about its spatial structure. Without a notion of how to locate oneself spatially
relative to other parts of the code,theres no there there in the words of author
Gertrude Stein. Thinking architecturally leads to representations of the code layout in terms of class and object relationships. Design patterns provide a common
language for describing the resulting spatial structures along with behavioral and
creational structures. This school of thought took root in the building architecture
community before making the leap to software architects and nally to scientic
software architecture.

4.3 Canonical Contexts

When the remaining chapters of Part II refer to the problem, they generally refer to
a dilemma that arises in designing software to embody a mathematical model. From

99

100

Design Patterns Basics

a design pattern viewpoint, the mathematical problem statement itself is simply the
context within which the software design problem occurs. Dening a set of canonical
contexts here facilitates focusing the rest of Part II more tightly on the software
problems, their solutions, and their consequences.
A desirable canon would contain contexts that are simple enough to avoid distracting from the software issues yet complex enough to serve as a proxy for the
contexts commonly encountered in scientic research. The chosen ones involve solving coupled ordinary differential equations, coupled integro-differential equations,
and a partial differential equation. All exhibit nonlinearity. In contexts involving temporal dependence, the time integration algorithms applied in these contexts range
from straightforward explicit Runge-Kutta schemes to multistep, semi-implicit, and
fully implicit schemes. In contexts involving spatial dependence, the nite difference methods employed likewise range from the explicit and straightforward to the
implicit and more complicated. Three important complications missing from our
canonical contexts are addressed in Part III of this text: complicated geometry and
boundary conditions and runtime scalability.
Section 4.3.1 provides the context for Chapters 68. Section 4.3.2 provides the
context for Chapter 5. Section 4.3.3 provides the context for Chapter 9.

4.3.1 The Lorenz Equations: A Chaotic Dynamical System

The software complexity issues that patterns address become most acute as a package
scales up to large numbers of classes. In scientic simulation code, this increase in
scale often involves adding new physics. Starting from a classical, Navier-Stokes
turbulence simulation package, one might add particle tracking to study multiphase
ow, magnetic eld transport to study magnetohydrodynamics (MHD), scalar elds
to study molecular diffusion, or quantum vortices to study superuidity. Each of
these topics forms a case study in Chapter 12 in Part III.
In most instances, the resulting multiphysics package greatly exceeds the size
desirable for a sample pattern demonstration. We therefore choose a very simple
mathematical model that can be formulated in a way reminiscent of a multiphyics
model. We defer discussion of true multiphysics solvers to Part III. The current
section aims to describe the mathematical motivation and to provide some basic
insights into its dynamics.
One can describe a broad swath of multiphysics phenomena with equations of
the form:

 U(
 x, t)), x , t (0, T]
U(x, t) = F(
t

(4.1)

 {U1 , U2 , ..., Un }T is the problem state vector; x and t are coordinates in the
where U
space-time domain  (0, T]; and F {F1 , F2 , ..., Fn }T is a vector-valued operator
that couples the state vector components via a set of governing ordinary-, partial-,
or integro-differential equations. Closing the equation set requires specifying
appropriate boundary and initial conditions:
 U)
 = C(
 x, t), x 
B(

(4.2)

 x, 0) = U0 (x), x 
U(

(4.3)

4.3 Canonical Contexts

 U)
 typically represents linear or nonlinear combinations of U

where  bounds , B(
 species the values of those combinations on .
and its derivatives, and where C
A common step in solving equation (4.1) involves rendering its right-hand side
discrete by projecting the solution onto a nite set of trial-basis functions or by
replacing all spatial differential operators in F by nite difference operators and all
spatial integral operators in F by numerical quadratures. Often, one also integrates
equation (4.1) against a nite set of test functions. Either process can render the
spatial variation of the solution discrete while retaining its continuous dependence
on time. One commonly refers to such schema as semidiscrete approaches. The
resulting equations take the form:
d 
 V(t))

V(t) = R(
t (0, T]
dt

(4.4)

 contains linear and nonlinear discrete

where the right-hand side vector function R
T

operators and where V {V1 , V2 , ..., Vm } might represent m = np samples of the n
 on a p-point grid laid over  .
elements of U
 could represent expansion coefcients, for example, Fourier
Alternatively, V
coefcients, obtained from projecting the solution onto the aforementioned space of
trial functions, such as complex exponentials. In any case, we can now assume that the
boundary conditions have been incorporated into equation (4.4) and no longer need
to be specied separately as in equation (4.2). This allows us to focus on equation
(4.4) as a self-contained, continuous dynamical system. For most of Part II, we ignore
the original spatial dependence in equation (4.1) and use a specic dynamical system
as our prototypical multiphysics model: the Lorenz system (Lorenz 1963):

(v2 v1 )
v

d 1
=
v1 ( v3 ) v2 .
v2

dt
v3
v1 v2 v3

(4.5)

We return to the issue of spatial dependence in Chapters 9 and 12.

The Lorenz equations retain some of the challenges posed by more complicated
equation sets: coupling terms, nonlinearities, and complex behavior. In certain regions of the -- parameter space, the Lorenz equations exhibit chaotic dynamics.
In subsequent chapters, we select the chaos-inducing values = 10, = 28, and
= 8/3. This provides a stringent test for codes in that the extreme sensitivity to
initial conditions characteristic of chaos magnies numerical inaccuracies over time.
One often refers to such extreme sensitivity to minor perturbations as the buttery
effect, wherein a buttery in India aps its wings and causes a hurricane in Florida.
The path by which Lorenz arrived at equation system (4.5) corresponds to a simplied semidiscretization of a set of PDEs governing forced, dissipative uid ow in a
horizontal layer of uniform depth with a xed temperature difference between its top
and bottom. All motion is assumed to be in horizontal planes. Lorenzs discretization
process involved Fourier approximations to the solution of the original PDE. The
dependent variables v1 , v2 , and v3 are the Fourier coefcients retained after truncating the Fourier series to three terms. In particular, v1 measures the strength of
the resulting convective motions, v2 varies proportionately with the overall temperature difference, and v3 varies proportionately with the vertical temperature prole

101

102

Design Patterns Basics

40
30
20
10
0

20

10
0
10
20

20

(x,y,z)

Figure 4.3. Phase-space trajectory of the Lorenz equation system solution.

from linearity. The parameter characterizes the relative magnitudes of the diffusion of momentum to that of thermal energy, whereas determines the stability of
buoyancy-driven motions and relates to the wavelength of the Fourier modes retained in the discretization. When > +1, solutions are unstable at sufciently high
values (> 24.74). Such behavior indicates turbulence in the original ow problem.
Explicit Euler time advancement of the Lorenz equations takes the form:
 
 n+1 = V
 n + t R
 V
n
V
(4.6)
 n+1 are the solution vectors at times tn and tn+1 tn +t, respectively,
 n and V
where V

and where R is the RHS of equation (4.5). Figure 4.3 depicts the behavior of an
explicit Euler time integration of the Lorenz system in phase space that is, in v1 -v2 v3 space. The buttery-shaped geometrical object toward which the initial solution
trajectory evolves is a strange attractor (so called because it possesses a fractional, or
fractal, spatial dimension) characteristic of a variety of forced, nonlinear, dissipative
systems. In the current case, the forcing stems from the energy input required to
maintain the overall thermal gradient.

4.3.2 Quantum Vortex Dynamics in a Superuid

The Lorenz system solution variables depend only on time. Among problems in
which the solution depends also on space, some of the most difcult involve timevarying topologies. In such problems, both the geometry and connectivity of the
domain evolve over time. A tangle of quantum vortices in superuid liquid helium
provides one of the simplest prototypes of such a problem because it can be accurately
modeled with one-dimensional, curvilinear objects subject to deformation, tear, and
reconnection.
Helium occurs naturally in two isotopes: 4 He and 3 He. Liquid 4 He exists in
two phases: He I and He II. He I is an ordinary liquid. He II, a superuid, forms

4.3 Canonical Contexts

below a critical temperature of approximately 2.17 K. Below this temperature, He II

exhibits exotic behavior. This behavior includes the formation of swirling vortices,
the circulation rate of which must be an integer multiple of the quantum of circulation:
h/m = 9.97 cm2 /s, where h is the Planck constant and m is the mass of a helium
atom.
At the core of each quantum vortex lies an evacuated region approximately
1-Angstrom across. When only the macroscopic dynamics of He II are of interest,
these thin cores can be approximated very accurately as 1D laments. They induce a
superuid velocity v(r, t) throughout the surrounding helium described by the BiotSavart law:

(S r) dS

(4.7)
v=
4
S r3
where S = S( , t) traces the vortex lament in 3D space as a function of arc-length
and time t, denotes a vector product, and the integration is performed along all
vortex laments in the simulation. The superuid velocity is thought to arise from
the motion of atoms at the ground energy level. The remaining excited atoms move
with a macroscopic velocity u that can be accurately described by classical mechanics
(see section 12.2.1).
In the absence of any background superuid motion imposed by boundary or
initial conditions, the equation of motion for a point S on a lament is:
dS
= v S (u v)  S [S (u v)]
dt

(4.8)

where S dS/d and where and  are temperature-dependent constants. The rst
term on the RHS is the total velocity induced by other points on the same lament
and other laments. At absolute-zero temperature, with all the atoms in the ground
state, this would be the only term on the RHS. The rst RHS term causes a circular
loop, such as that of Figure 4.4, to propagate in a direction orthogonal to the plane
of the loop. The second and third RHS terms model the effect of mutual friction

To be inserted
To be deleted

S(,t )

S'

Figure 4.4. Quantum vortex lament discretization and adaptive remeshing.

103

104

Design Patterns Basics

Figure 4.5. Quantum vortex lament before (left) and after (right) reconnection. Curved
arrows indicate direction of circulation.

between the superuid component and the excited atoms, which comprise a normal
uid occupying the same volume. For a circular loop, the second term causes the
vortex to shrink as it propagates. For the same loop, the third term, which is typically
the smallest in magnitude, imposes a drag that opposes the induced velocity.
One renders the continuous vortex core discrete by placing a series of mesh
points along each lament as in Figure 4.4. Writing equations (4.7)(4.8) for each
mesh point yields a coupled set of nonlinear, integro-differential equations. The
motions of these mesh points determines the lament motion and shape.
Two processes change the connectivity of the points. First, as the distance
between mesh points changes, new points must be inserted and existing points
removed to balance resolution requirements against computational cost and complexity (Morris 2008). Second, when two laments approach one another, their
interaction tends to draw them closer, resulting in antiparallel vortices (Schwarz
1985, 1988). Koplik and Levine (1993) showed that the laments join and reconnect
whenever two antiparallel vortex laments approach within a few core diameters of
each other, as demonstrated in Figure 4.5.
Broadly speaking, numerical solution of the discretized vortex equation system
breaks down into three pieces. One is interpolating the normal uid velocity u at each
superuid vortex point S. A second is evaluating the Biot-Savart integral. A third is
evaluating the parametric spatial derivative S . Morris (2008) tackled these problems
in a dissertation on superuid turbulence. Appendix A details those aspects of her
numerical methods that appear in the Exercises at the end of Chapter 5.

4.3.3 Burgers Equation: Shock Formation and Dissipation

Spatial dependence enters the quantum vortex context via the line integral (4.7).
More commonly, spatial dependence appears in a model context in the form of spatial
derivatives in a PDE. The Burgers equation exhibits the desired balance between
simplicity and representative complexity for a canonical context:
ut + uux = uxx

(4.9)

which we write in the subscript notation of Table 3.1. Although Burgers (1948) introduced this equation as a simplied model for uid turbulence, in which setting

4.3 Canonical Contexts

u plays a role analogous to velocity and to kinematic viscosity, the equation also
crops up in settings ranging from superconductivity to cosmology (see Canuto et al.
2006).
Despite its apparent simplicity, solving the Burgers equation presents sufciently
vexing challenges to warrant exibility: A numerical method that works in one regime
across a given range of the parameter nu for a given set of initial and boundary conditions might not work well in other regimes. Great value accrues from ensuring the
ability to swap numerical methods within a single solver framework. Furthermore,
the dynamics can change sufciently even within a single solution time interval to
warrant dynamic adaptivity. In particular, the nonlinear second term on the LHS
of equation (4.9) causes solutions to develop steep gradients characteristic of the
nearly discontinuous shock waves that occur in settings as widely disparate as supernovae and rocket plumes. These might necessitate locally adaptive, shock-capturing
schemes in the vicinity of the steep gradients (Berger and Colella 1989), while less
demanding discretizations could be used elsewhere in space and time as the RHS
term dissipates these gradients.
The perils of approximating the Burgers equation arise partly from oscillating
instabilities related to bifurcations in the behavior of the discretized versions of the
equation. For small , solutions near these bifurcations display a rich variety of
behaviors. These include limit cycles (asymptotic approaches periodicity in time)
and strange attractors (Maario et al. 2007).
An additional attraction to including the Burgers equation in our canon stems
from its unusual status as one of the few nonlinear PDEs with known exact solutions.
Hopf (1950) and Cole (1951) demonstrated that transforming the Burgers equation
according to:
u = 2x /
(4.10)
converts it into the linear heat equation:
t = xx .

(4.11)

which admits solutions of the form:

(x, t) =

4 n=

e(x2 n)

2 /(4t)

(4.12)

in the case of periodic boundary conditions. Figure 4.6 plots an initial condition
u(x, 0) = 10 sin(x) along with exact and approximate periodic solutions on the interval
[0, 2). The numerical solution integrates the conservative form:
ut = uxx (u2 /2)x

(4.13)

so called because discretizations of the RHS conserve the integral of u2 /2 over the
domain when = 0. One solution in the gure employs second-order, Runge-Kutta
time advancement and sixth-order accurate Pad scheme for spatial differencing (Moin 2001). Appendix A summarizes these two algorithms, the advantage of
using the high order nite difference scheme is to achieve spectral-like accuracy in
space. Another solution plotted in the same gure, obtained with a second-order
central difference scheme for spatial discretization, is also ploted in Figure 4.6 for

105

Design Patterns Basics

10
t = 0, exact solution
t = 0.1, 2nd order
t = 0.1, 6th order
t = 0.1, exact solution
t = 0.6, 2nd order
t = 0.6, 6th order
t = 0.6, exact solution

8
6
4
2
u

106

0
2
4
6
8

10

3
x

Figure 4.6. Solutions of the Burgers equation with initial condition u = 10 sin x and periodic
boundary conditions.

comparison. With only 16 grid points, the sixth-order Pad scheme leads to more
accurate results than the second-order central difference scheme.
Yes We Canon
Simple pattern demonstrations run the risk of missing important complications
found in actual scientic codes. Complicated ones run the risk of losing the forest
in the trees. We have constructed a small canon of relatively simple problems
that exhibit some of the chief complexities found in more complicated scientic
simulations.

EXERCISES
1. Write an abstract vortex_filament derived type in Fortran with a deferred
remesh binding that implements an abstract interface.
2. Write a pure virtual VortexFilament class derived type in C++ with a pure virtual
remesh member function.
3. List recurring themes you see in buildings you like. Are there analogies you can
make to software architecture?

The Object Pattern

Memory is a crazy woman [who] hoards colored rags and throws away food.
Austin OMalley

5.1 The Problem

Large software development efforts typically require a degree of consistency across
the project to ensure that each developer follows practices consistent with the critical
goals of the project. In high-performance computing (HPC), for example, Amdahls
law (Chapter 1) suggests that scaling up to the tens or hundreds of thousands of processor cores available on leadership-class machines requires that every dusty corner
of the code make efcient use of the available cores. Otherwise, whichever fraction of the code speeds up more slowly with increasing numbers of cores eventually
determines the overall speedup of the code.
Another form of consistency proves useful when one desires some universal way
to reference objects in a project. Doing so facilitates manipulating an object without
knowledge of its identity. The manipulated object could equally well be an instance
of any class in the project.
In HPC, communicating efciently between local memories on distributed processors represents one of the most challenging problems. One might desire to ensure
consistency in communication practices across the project. In these contexts, two
broad requirements drive the desire to impose a degree of uniformity across a design: one stemming from a need for consistent functionality, and the other stemming
from a need for consistent referencing.
Opposing these forces is the desire to avoid overconstraining the design. In
the worst-case scenario, imposing too much uniformity sties creativity and freezes
counterproductive elements into the design. Continuing with the HPC example,
were one to mandate that all communication take place through a shared-memory
protocol, for example, porting the code to a distributed-memory platform would
be difcult or impossible. In a more innocuous scenario, one might force all classes
to implement methods that nominally support a protocol that proves superuous
for those classes that carry out embarrassingly parallel tasks that is, tasks that
107

108

The Object Pattern

involve little or no communication. Writing superuous methods wastes the time of

the writers and potentially sets up unfounded expectations on the part of clients.
Tie Me Updont Tie Me Down
Classes at the base of an inheritance hierarchy can ensure that those logically
further up the hierarchy implement mission-critical capabilities. They can also
do harm when those capabilities prove superuous for certain ADTs. The best
universal base classes contain minimal or no functionality and of sufciently broad
utility that they might be candidates for incorporation into the language.

5.2 The Solution

In a procedural programming project, one might achieve the desired consistency in functionality by writing utilities that perform tasks that are common
within the given application domain. The project leader could require that all
developers use these utilities whenever applicable. In HPC, for example, the Basic Linear Algebra Communication Subprograms (BLACS) serve this role for
ScaLAPACK (Blackford et al. 1997).
In an OOP project, one might construct classes that provide methods that perform common functions on encapsulated data. Continuing with the HPC example,
some methods might communicate data between processors. This communication
might even be hidden from client code. Developers could build atop these classes
by accepting their instances as arguments, by extending them via inheritance or type
extension, or by aggregating them into new classes. The Trilinos1 solver frameworks
Epetra package takes this approach (Heroux et al. 2005). When built atop Trilinos,
most user code never needs to communicate information explicitly, leaving that detail
to Trilinos.
Fullling the expressed desire for consistency in referencing proves somewhat
trickier and potentially language-dependent. Since the application context in the
current chapter does not present a need for universal referencing, we defer the
related discussion to subsequent chapters. Chapter 6 discusses the necessary language
infrastructure: the select type construct in Fortran and the dynamic_cast operator
in C++. Chapter 11 presents an example of using this infrastructure in an actual
open-source software package.
The OBJECT pattern takes its name and inspiration from a native Java construct:
the Object class. In Java, Object is a universal base class that every Java class extends.
By denition then, every class inherits any Object functionality that it does not
explicitly override. This satises the need for common functionality at least insofar
as the minimum functionality desired by the language designers. Furthermore, any
instance of a Java class can always be treated as an Object instance. This satises
the need for a universal reference without need for exception handling.
Hence, the OBJECT pattern denes a base class hierarchy and stipulates that every
class in a given project extends the class at the top of that hierarchy. As with our discussion of a two-part FACADE in section 4.2.3, this pattern potentially maps to multiple
1 http://trilinos.sandia.gov

5.2 The Solution

classes. Several reasons motivate implementing OBJECT as a unidimensional, hierarchical collection of classes rather than a single class. First, when the desired common
functionality involves unrelated methods, collecting them into one class results in
the coincidental cohesion, the weakest and least desirable form (see Section 3.2.1).
A more understandable and maintainable design results from separating concerns.
Second, any deferred bindings in the parent type (base class) hierarchy must
be implemented in all concrete classes in the project. The OBJECT pattern should
therefore be extremely stable. Only deferred bindings expected to survive for the life
of the project should be included. Nonetheless, if some deferred bindings seem likely
to be removed later, placing these deeper in the ancestral hierarchy inoculates the
child classes from subsequent deletion of the relevant ancestor, because the classes
only reference the ultimate descendant that is, ultimate from viewpoint of the
OBJECT hierarchy. Declarations of instances of this ultimate descendant elsewhere in
the package need not be changed when more distant ancestral relatives are removed.
Since efcient use of HPC platforms usually requires a sophisticated understanding of the underlying architecture and the chosen parallel programming paradigm,
we defer further discussion of HPC to Part III. There we exploit the C interoperability features of Fortran 2003 to construct a mixed Fortran/C++, object-oriented,
scalable ADT calculus. That calculus circumvents many of the details required for
more traditional approaches ensuring runtime scalability.
Here we focus on an example involving the bane of every programmer whose
project requires considerable use of dynamic memory allocation: memory leaks. By
introducing allocatable entities, Fortran relieves the programmer from thinking
about leaks in a large fraction of the cases where they could occur in C++. C++ has a
long-standing reputation of difcult memory management relating to exposing lowlevel entities directly to the programmer specically, exposing machine addresses
via pointers. Many partial solutions exist (standard container classes, management
via lifetime of stack allocated objects, reference counting, garbage collectors, etc.),
but the current language standard does not provide any general, portable, or efcient
solution to the problem of managing object lifetimes. Section 5.2.2 presents an invasive reference-counting scheme, but in the absence of a language-provided solution,
this approach will almost certainly prevent a compiler from eliminating temporaries
and a host of other optimizations.
Another approach puts each developer in charge of managing memory but gives
them tools that assist them in doing so. Although the use of allocatable entities
in Fortran obviates the need for developers to explicitly free memory in the vast
majority of circumstances, at least three scenarios warrant giving programmers this
responsibility. First, as recently 2006, three of ve compilers tested exhibited memory leaks resulting from not freeing memory in the manner in which the Fortran
2003 standard obligates them (Rouson et al. 2006). Until widespread availability of
leak-free compilers becomes the norm, developers will often need to adopt defensive postures. Second, as recently as December 2009, only two of seven compilers
surveyed supported the nal subroutines the standard provides for programmers to
instruct the compiler as to how to free memory automatically when a pointer component goes out of scope (Chivers and Sleightholme 2010). Third, programmers might
want to explicitly free memory before the associated entity goes out of scope.

109

110

The Object Pattern

The exibility and power of allocatable entities for automatic dynamic memory management makes the need for pointers (and hence nal subroutines) quite rare
in Fortran. The quantum vortex lament discretization of Chapter 4, however, maps
quite nicely onto a linked list data structure with pointer components. Figure 4.4 naturally evokes notions of translating each mesh point into a node in the list and each
line segment into a pointer or bidirectional pair of pointers. We use this example in
the remainder of this section.

5.2.1 Fortran Implementation

Markus (2003) identied several scenarios wherein memory leaks can occur in
Fortran code in settings that resonate with the central theme of the current book:
evaluating expressions built up from dened operators and dened assignments.
Rouson et al. (2005) and Markus independently proposed nearly identical memory
management algorithms that eliminated leaks in most cases. In those cases, the simplicity of their strategy makes it attractive. Chapter 10 uses it to motivate formal
design constraints.
Stewart (2003) proposed a more complicated approach that eliminates leaks in
several situations not covered by Markus and Rouson and colleagues. He considered
a derived type Rmat containing a pointer component used in a manner analogous to
that shown in Figure 5.1. The rst leak occurs at line 11 in 5.1(b), which implicitly
invokes the type-bound procedure total in 5.1(a). The leaks at lines 1718 appear
similar but have subtle differences discussed in Chapter 10.
The total procedure allocates memory for a result that becomes the rhs argument of the assignment procedure assign. Herein lies the dilemma: Once assign
terminates and control reverts to the main program, rhs goes out of scope that is
main contains no variable name associated with the memory referenced by rhs, the
contents of which have been copied into lhs, which main refers to as c. More importantly, main contains no variable name associated with the component rhs%matrix.
Hence, while the compiler will deallocate rhs, the language semantics do not provide
for deallocation of rhs%matrix. With a standard-compliant Fortran 2003 compiler,
one can remedy this situation most simply by replacing the pointer component with
an allocatable component, obligating the compiler to free the memory. Alternatively, one could dene a nal subroutine that invokes the deallocate intrinsic on
the pointer component. Otherwise, the memorys fate rests with the programmer.
In order to ensure that each developer has access to memory management tools
and implements certain supporting functionality, one might place the tools in a universally extended parent type hierarchy. Figure 5.2 demonstrates this strategy using
the procedures SetTemp, GuardTemp, and CleanTemp adapted from Stewart (2003).
The hierarchy in this example includes only one abstract derived type: hermetic.
The hermetic type satises the aforementioned desire to encourage certain universal practices. If every class extends hermetic, it could also satisfy the desire for
universal referencing. However, it is usually better to separate these two concerns
into different types. Chapter 11 discusses an example in which the the need for universal referencing arises and a second type in the universal parent hierarchy provides
for it.

5.2 The Solution

Following Markus, one marks all function results as temporary. As employed

in ADT calculus by Rouson et al., this covers any intermediate result an operator
allocates and passes to an operator of lower precedence or to a dened assignment. As
implemented by Stewart, SetTemp marks a result temporary by allocating its integer
pointer component if necessary and by setting that components target to 1. The
use of an integer pointer component rather than an integer component facilitates
complying with the Fortran standard requirement that the dummy arguments of
dened operators have the intent(in) attribute. The application of this attribute to
a derived type ows down to the derived type components. In the case of a pointer
component, intent(in) prevents changing with which target a pointer is associated,
but places no restriction on changing the values stored in the target. This allows for
incrementing setting and incrementing the integer as described next.
In Stewarts scheme, each subroutine or function that receives an argument of
type Rmat that could be temporary, it guards it from premature deallocation by
passing it to GuardTemp. If it is temporary, GuardTemp increments the tally stored in
the integer pointer component. Stewart refers to this tally as a depth count due to
his interest in the deeply nested function call trees characteristic of dened-operator
expressions.
The nal step in Stewarts approach requires calling CleanTemp just before exiting each procedure that receives a possibly temporary Rmat argument. If CleanTemp
receives a temporary object, it checks whether the depth count exceeds 1. If so, it
decrements the count. Then if the count it equals 1, CleanTemp frees the memory
associated with the argument. In Stewarts concrete Rmat type, CleanTemp deallocates the memory itself. Our OBJECT implementation in Figure 5.2 denes a hermetic
type. Since this is an abstract type, its only utilities derives from extending it. Furthermore, because the private scoping of its integer pointer component applies at the
module level and because all extended types are implemented in separate modules,
hermetic publishes several accessors, that is, methods that control access to its private state. It also publishes the deferred binding force_finalization. CleanTemp
invokes this binding, which concrete extended types must implement. With a compiler that supports type nalization, force_finalization might simply invoke the
types nal subroutine. Otherwise, it would free the memory itself. We discuss next
an interesting hybrid of these two approaches.
The vortex type of Figure 5.3 demonstrates the beginnings of a solver for
the quantum vortex dynamics context of Section 4.3.2. The vortex type extends
hermetic. Its ring constructor creates a circular linked list of vortex points. Much
like the ring discretization of Figure 4.4, the list connects tail to head.
The vortex implementation also provides a finalize nal subroutine and a
manual_finalizer implementation of the force_finalization binding. Both subroutines exhibit subtle, elegant behaviors associated with Fortran type nalization.
The circular referencing of the linked list returned by ring precludes the simple nullterminated traversal of the linked list common to linear linked list implementations.
Accordingly, the rst step in finalize involves articially converting the list to a
linear form by breaking a link at line 47 in Figure 5.3. This happens immediately after
storing a reference to the node targeted by that link. After these two steps, finalize
passes the linked node to deallocate if the link is associated with a target.

111

112

The Object Pattern

The Fortran standard obligates compilers to nalize an object whenever a programmer deallocates it as line 48 does.2 This makes finalize inherently recursive
as indicated by the recursive attribute on line 43. This artice neatly provides for
automatic traversal and destruction of the entire linked list in the three executable
lines 4648.
The language semantics harbor equally subtle and elegant implications for
manual_finalizer. When one invokes this procedure to free memory in advance
of when a compiler might otherwise do so automatically via the nal subroutine, it
is critical to ensure that the two procedures do not interact in harmful ways. This
issue vanishes with compilers that do not support nal subroutines. For these, the
programmer must remove the final binding at line 15 manually or automatically,
for example via Fortrans conditional compilation facility or Cs preprocessor. In
such cases, manual_finalizer breaks the chain in the same manner as finalizer
and traverses the vortex, deallocating each node until it returns to the broken link.
When the compiler supports nal subroutines, the deallocate command at
line 58 implicitly invokes finalizer, after which the destruction process continues
recursively and automatically, making only one pass through the do while block
that begins at line 56. The automatically invoked nal subroutine and the explicitly
invoked manual destruction procedure peacefully coexist.
Figure 5.4 demonstrates vortex construction, output, and destruction. Exercises
at the end of this chapter guide the reader in writing code for point insertion, removal,
and time advancement via ADT calculus. Line 5 in Figure 5.4 marks the rst Fortran pointer assignment in this text. Were the constructor result produced by the
RHS of that line assigned to a LHS object lacking the pointer attribute, an intrinsic
assignment would copy the RHS into the LHS and then deallocate the RHS. This
would start the aforementioned nalization cascade and destroy the entire ring! The
pointer assignment prevents this.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

Figure 5.1(a)
module Rmat_module
implicit none
type Rmat
real ,dimension(:,:) ,pointer :: matrix=>null()
contains
procedure :: total
procedure :: assign
generic
:: operator(+) => total
generic
:: assignment(=) => assign
end type
contains
type(Rmat) function total(lhs,rhs)
class(Rmat) ,intent(in) :: lhs,rhs
allocate(total%matrix(size(lhs%matrix,1),size(lhs%matrix,2)))
total%matrix = lhs%matrix + rhs%matrix
2 Although the standard does not specify whether nalization or deallocation comes rst, it seems

reasonable to assume compilers will nalize rst to avoid breaking nal subroutines that assume the
object remains allocated upon invocation, and to prevent memory leaks that would result if there
remained no handle for a pointer component still associated with a target.

5.2 The Solution

16
17
18
19
20
21
22

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20

end function
subroutine assign(lhs,rhs)
class(Rmat) ,intent(inout) :: lhs
type(Rmat) ,intent(in) :: rhs
lhs%matrix = rhs%matrix
end subroutine
end module
Figure 5.1(b)
program main
use Rmat_module
implicit none
integer ,parameter :: rows=2,cols=2
type(Rmat) :: a,b,c
allocate(a%matrix(rows,cols))
allocate(b%matrix(rows,cols))
allocate(c%matrix(rows,cols))
a%matrix=1.
b%matrix=2.
c = a + b ! memory leak
call foo(a+b,c)
contains
subroutine foo(x,y)
type(Rmat) ,intent(in) :: x
type(Rmat) ,intent(inout) :: y
y=x
! memory leak
y=x+y
! memory leak
end subroutine
end program
Figure 5.1. Memory leak example adapted from Stewart (2003).

1
2
3
4

Figure 5.2
module hermetic_module
implicit none
private
public :: hermetic ! Expose type and type-bound procedures

5
6
7
8
9
10
11
12
13
14

type ,abstract :: hermetic

private
integer, pointer :: temporary=>null() !Null marks non-temporary data
contains
procedure :: SetTemp
! Mark object as temporary
procedure :: GuardTemp ! Increment the depth count
procedure :: CleanTemp ! Decrement depth count/Free memory if 1
procedure(final_interface) ,deferred :: force_finalization
end type

15
16
17
18
19
20
21

abstract interface
subroutine final_interface(this)
import :: hermetic
class(hermetic) ,intent(inout) :: this
end subroutine
end interface

113

114

The Object Pattern

22
23
24
25
26
27

contains
subroutine SetTemp(this)
class(hermetic) ,intent(inout) :: this
if (.not. associated(this%temporary)) allocate(this%temporary)
this%temporary = 1
end subroutine

28
29
30
31
32

subroutine GuardTemp (this)

class (hermetic) :: this
if (associated(this%temporary)) this%temporary = this%temporary + 1
end subroutine

33
34
35
36
37
38
39
40
41
42
43
44

subroutine CleanTemp(this)
class(hermetic) :: this
if (associated(this%temporary)) then
if (this%temporary > 1) this%temporary = this%temporary - 1
if (this%temporary == 1) then
call this%force_finalization()
deallocate(this%temporary)
end if
end if
end subroutine
end module
Figure 5.2.

1
2
3
4
5

OBJECT

implementation: universal memory management parent.

Figure 5.3
module vortex_module
use hermetic_module ,only : hermetic
implicit none
private
public :: vortex

6
7
8
9
10
11
12
13
14
15
16

integer ,parameter :: space_dim=3

type ,extends(hermetic) :: vortex
private
real ,dimension(space_dim) :: point
type(vortex) ,pointer :: next=>null()
contains
procedure :: force_finalization => manual_finalizer
procedure :: output
final :: finalize
end type

17
18
19
20

interface vortex
procedure ring
end interface

21
22
23
24
25
26
27

contains
function ring(radius,num_points)
real ,intent(in) :: radius
integer ,intent(in) :: num_points
type(vortex) ,pointer :: ring,current
integer :: i

5.2 The Solution

28
29
30
31
32
33
34
35
36
37
38
39
40
41

real ,parameter :: pi = acos(-1.)

real :: delta,theta
delta = 2.*pi*radius/real(num_points) ! arclength increment
allocate(ring)
ring%point = [radius,0.,0.]
current => ring
do i=1,num_points-1
allocate(current%next)
theta = i*delta/radius
current%next%point = [radius*cos(theta),radius*sin(theta),0.]
current => current%next
end do
current%next => ring ! point tail to head
end function

42
43
44
45
46
47
48
49

recursive subroutine finalize (this)

type(vortex), intent(inout) :: this
type(vortex) ,pointer :: next
next => this%next
this%next=>null() ! break the chain so recursion terminates
if (associated(next)) deallocate (next) ! automatic recursion
end subroutine

50
51
52
53
54
55
56
57
58
59
60
61

subroutine manual_finalizer(this)
class(vortex) ,intent(inout) :: this
type(vortex) ,pointer :: current,next
current => this%next
this%next=>null() ! break the chain so loop terminates
do while(associated(current%next))
next => current%next ! start after break in chain
deallocate(current)
current=>next
end do
end subroutine

62
63
64
65
66
67
68
69
70
71
72
73

subroutine output(this)
class(vortex) ,intent(in) ,target :: this
type(vortex) ,pointer :: current
current => this
do
print *,current%point
current => current%next
if (associated(current,this)) exit
end do
end subroutine
end module
Figure 5.3. Quantum vortex linked list with hermetic memory management.

1
2
3
4
5

program main
use vortex_module ,only : vortex
implicit none
type(vortex) ,pointer :: helium
helium => vortex(radius=1.,num_points=8)

115

116

The Object Pattern

6
7
8

call helium%output()
call helium%force_finalization()
end program
Figure 5.4. Vortex creation and destruction.

Hermetic

Vortex
Figure 5.5. The concrete Vortex ADT implements the Hermetic interface.

5.2.2 C++ Style and Tools

This chapter marks the rst substantial C++ examples in this text. We aim to put
Fortran and C++ on roughly equal footing by exploiting primarily those constructs
in one language that have reasonably close counterparts in the other. This strategy
makes the text self-consistent and empowers readers more familiar with one language
to comfortably explore the other.
Our approach requires eschewing some advanced features in both languages
and restricting ourselves to current standards. Our C++ idioms might appear overly
restrictive to advanced programmers who routinely link to libraries outside the
core language. For example, this sections memory management approach could
be greatly enhanced by the BOOST open-source C++ library3 , many parts of which
are expected to enter the upcoming C++ 0x standard.
Although we seek to construct close counterparts to our Fortran implementations, we also seek to write C++ examples that appear natural to C++ programmers.
This gives us freedom to use styles and naming conventions that differ from those
common in Fortran. For example, in class denitions, all (private) data members
names are appended with an underscore (_), which is a common practice in C++ to
indicate the names should only be referenced internally within member functions.
The next subsection demonstrates an instance in which the goals in the prior two
paragraphs collide. There we construct a base class that provides tools for tracking
the number of pointers to dynamically allocated memory. A common way to accomplish this in C++ involves template metaprogramming, a facility with no Fortran
counterpart.

5.2.2.1 Reference Counting

Stewarts depth count scheme corresponds to the widespread reference counting
approach in C++. Absent language support for garbage collection of pointerallocated memory (although BOOSTs shared_ptr construct provides a garbage
collector similar to Fortrans allocatable variables), C++ programmers commonly
3 http://www.boost.org

5.2 The Solution

employ reference counting to obtain functionality similar to garbage collection. Reference counting refers to the strategy of keeping track of the number of references
or pointers that refer to the memory associated with an object, so the memory can be
freed when it is no longer referenced. In practice, reference counting helps prevent
the aforementioned memory leaks in applications that rely frequently on pointers.
The signicant risk of memory leaks associated with raw pointers combined with
the ubiquitous need for pointers in C++ makes a reference-counted pointer (RCP)
class an ideal candidate for incorporation into an OBJECT pattern. Figure 5.6 contains
a RCP implementation based on C++ template classes. Figure 5.6(a) shows a basic
RCP template class pointer<> denition. It wraps raw C++ pointers and thereby
facilitates treating objects as if one were using a C++ pointer directly.

1
2

#ifndef POINTER_H_
#define POINTER_H_

Figure 5.6(a): pointer.h

3
4
5
6

/**
* A simple wrapper for raw pointer.
*/

7
8
9

#include <iostream>
#include <unistd.h>

10
11
12
13
14
15
16

template <typename T>

class pointer
{
public:
pointer() : ptr_(NULL){}
pointer(T *other) : ptr_(other){}

17
18

void assign(T *other) { ptr_ = other; }

19
20
21
22
23
24
25
26
27
28
29
30
31

pointer& operator=(T * other) {

assign(other);
return *this;
}
const pointer& operator=(const T & other) {
assign(const_cast<T*>(&other));
return *this;
}
pointer& operator=(T & other) {
assign(&other);
return *this;
}

32
33
34

T* operator->() { return ptr_; }

const T* operator->() const { return ptr_; }

35
36
37

T& operator*() { return *ptr_; }

const T& operator*() const { return *ptr_; }

38
39

bool operator==(const T *other) const { return ptr_ == other; }

117

118

The Object Pattern

40

bool operator!=(const T *other) const { return ptr_ != other; }

41
42

T* ptr() const { return ptr_; }

43
44
45
46
47

1
2

private:
T *ptr_;
};
#endif

#ifndef REF_H_
#define REF_H_

Figure 5.6(b): Ref.h

3
4
5
6

/**
* A very simple invasive reference counted pointer.
*/

7
8

#include "pointer.h"

9
10
11
12
13

template <typename T>

class Ref : virtual public pointer<T> {
public:
typedef Ref<T> Self_t;

14
15

Ref() {}

16
17
18
19
20

Ref(const Self_t &other) {

assign(other.ptr());
if(ptr() != NULL) ptr()->grab();
}

21
22
23
24
25
26

template <typename Other>

Ref(Ref<Other> other){
assign(other.ptr());
if(ptr() != NULL) ptr()->grab();
}

27
28
29
30
31
32

template <typename Other>

Ref(Other *other) {
assign(other);
if(ptr() != NULL) ptr()->grab();
}

33
34
35
36

Ref() {
if(ptr() != NULL) ptr()->release();
}

37
38
39
40
41
42
43
44

Self_t& operator=(const Self_t &other) {

if(other.ptr() != ptr()) {
if(ptr() != NULL) ptr()->release();
assign( other.ptr());
if(ptr() != NULL) ptr()->grab();
}
return *this;

5.2 The Solution

45

46
47
48
49
50
51
52
53
54
55

Self_t& operator=(T *other) {

if(ptr() != other) {
if(ptr() != NULL) ptr()->release();
assign( other);
if(ptr() != NULL) ptr()->grab();
}
return *this;
}
};

56
57
58
59
60
61

1
2

template <typename Target, typename Source>

Ref<Target> cast(const Ref<Source> &pp) {
return Ref<Target>(dynamic_cast<Target*>(pp.ptr()));
}
#endif

#ifndef REFBASE_H_
#define REFBASE_H_

Figure 5.6(c): RefBase.h

3
4
5
6

/**
* Base for reference counted objects.
*/

7
8

#include "Ref.h"

9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

class RefBase {
public:
RefBase() : cnt_(0) {}
RefBase(const RefBase &) : cnt_(0) {}
virtual RefBase() {}
void grab() throw() {
cnt_++;
}
void release() throw() {
cnt_--;
if(! cnt_) delete this;
}
private:
int cnt_;
};
#endif /* ! REFBASE_H_ */
Figure 5.6. Reference counting in C++: (a) a template pointer base class for the referencecounted class Ref, (b) a template class for the reference-counted pointer Ref, and (c) a
reference-counted object RefBase base class .

We use the RCP class from Figure 5.6(b) and the reference-counted object class of Figure 5.6(c) for the C++ examples throughout this book. The
current implementation grew out of C++ sample code that rst appeared in
Rouson, Adalsteinsson, and Xia (2010). The presentation here provides more

119

120

The Object Pattern

detailed descriptions of the structure and technical approach. In particular, we

attempt to mitigate any confusion that might arise from the template syntax. Templates offer a unique and powerful technique for writing generic code that is
applicable to any data types. Hence, one does not need to repeat the same algorithm for different data types. For excellent tutorials on C++ templates, see
Vandevoorde and Josuttis (2003).
Highlights of the techniques employed in Figure 5.6(a) include the following:
Constructors.

A pointer<T> object can be constructed a few different ways: It can be initialized as a NULL by default or it can be initialized by a pointer. We rely on the
default copy constructor generated by the compiler. Were we to dene a copy
constructor, it would behave identically to the compiler-generated one.
pointer() : ptr_(NULL){}
// default constructor
pointer(T *other) : ptr_(other){} // takes a raw pointer
Assignment operators.

The assignment operators for pointer<> are overloaded to have a few different
forms:
pointer& operator=(T * other);
// RHS is a T*
const pointer& operator=(const T&); // RHS is const T&
pointer& operator=(T & other);
// RHS is T&
As with the constructors, we provide no copy assignment operator and rely on
the compiler to generate the correct behavior.
Member access
The member access operators are dened so that any pointer<T> object is
treated just as a regular pointer. Consider the following declarations of the operators -> and *. We can illustrate their usage with the following example: Given
a declaration pointer<T> x;, one can use pointer syntax such as x->foo() or
(*x).foo(), assuming type T has a member function named foo():
operator ->:
T* operator->();
const T* operator->() const;

// member access; lvalue expr

// member access; rvalue expr

operator *:
T& operator*();
const T& operator*() const;

// indirect; lvalue expr

// indirect; rvalue expr

Pointer comparison.

Since the pointer<> class emulates actual pointer behavior, we also need to
overload the comparison operators == and != so that they are actually comparing pointer addresses. Continuing with the previous example of object x, the

5.2 The Solution

expression x == NULL tests if x contains a NULL pointer. Similarly, x != y returns true if (the address of) the memory associated with x is not the same as
that associated with y:
bool operator==(const T *other) const;
bool operator!=(const T *other) const;
Pointer access.

Providing access to the value of the underlying pointer is straightforward:

T* ptr() const { return ptr_; }
Figure 5.6(b) presents the denition of a template class, Ref<>, that does the
reference counting when the template parameter type is RefBase (see Figure 5.6[c])
or a subclass of RefBase. Class Ref<> is derived from pointer<> so every instance
of Ref<> is also treated as a pointer. In addition, Ref<> provides techniques for
managing reference counts.
Copy constructors.

Copy constructors increase the reference count. Whenever a new instance of

Ref<T> is created, one of the copy constructors is invoked in order to copy the
pointer data. Class Ref<> denes all necessary copy constructors that are used
in our examples. A common theme of these copy constructors is they all call a
method named grab(), which increments the reference count by one:
// copy constructors: call grab()
Ref(const Self_t &other);
template <typename Other> Ref(Ref<Other> other);
template <typename Other> Ref(Other *other);
Destructor.

The destructor, dened as Ref<>, decreases the reference count when a RCP
goes out of scope. As shown in Figure 5.6(b), the destructor invokes the method
release() to decrement the reference count by one. If the reference count
reaches zero, then all pointers have reached the end of their lifetimes and the
memory to which the pointer points is freed. The RefBase class denes the
grab() and release()implementations as discussed after the next bullet:
Ref();

// destructor; call release()

Assignment operators.

The assignment operators manage the reference counts for both the LHS and
RHS. As shown below, these operators take an approach that differs slightly
from the copy constructors. One can envisage the process of assigning a Ref<>
pointer comprising disassociating the LHS pointer followed by the reassociating
it with the RHS pointer. The implementations of the assignment operators reect
this process: release the LHS pointer rst then associate (the address of) the LHS

121

122

The Object Pattern

pointer to the RHS pointer, and nally grab() the LHS pointer to update the
reference count on the pointer:
// assignment operators:
Self_t& operator=(const Self_t&); // RHS is a Ref<T>
Self_t& operator=(T *other);
// RHS is T*
A helper function named cast() is also dened for Ref<>. As implied by its
name, this function typecasts between two Ref<> pointers with different parameter
types using C++s dynamic_cast. The function is used in Chapter 68 when type cast
is actually needed.
The class denition of RefBase, as shown in Figure 5.6(c), provides a base class
for any object that needs reference counting. This class implements the method
grab() and release(). The actual number of references to this object is stored by
the attribute cnt_. A careful reader might have noticed the absence of keyword
virtual in denitions of grab() and release(). This is intentional. It prevents
subclasses of RefBase from overriding these methods and potentially breaking the
reference counting scheme.
The combination of Ref<> and its parameter class RefBase closely corresponds to
the hermetic type dened in Fortran. We also use Ref<> and RefBase in Chapters 68
to emulate Fortrans rank-one allocatable arrays.
5.2.2.2 Output Format for Floating Point Quantities
The I/O formatting differences between C++ and Fortran warrant a format conversion tool to facilitate direct comparison between the two languages output. Since all
of our Fortran examples use list-directed output for oating point arrays, we developed a helper class to handle the C++ oat output format. Figure 5.7(a) denes this
class fmt. It stores two parameters, w_ and p_, for controlling the output width and
precision (the number of signicant digits in the output eld). Users can choose a
desired width and precision when constructing a fmt object. By default, the output
width is 12 characters wide and with 8 signicant digits, matching roughly that in
Fortran for a single oat4 .
The internal member v_, which is of C++ STLs vecotr<float> type, stores the
actual value of the output item. Figure 5.7(b) gives the detailed type denitions of
real_t and crd_t. To simulate the list-directed output format for a Fortran real
(C++ float) array, such as that produced by print *, X, the C++ operator <<
is overloaded, as shown from line 18 to 29 in fmt.h. Hence, the C++ analogue to
the aforementioned Fortran print statement is std::cout << fmt(X);.

5.2.3 C++ Implementation of Vortex

Figures 5.8(a)(b) show the C++ vortex class denition header and implementation
les, respectively. Since memory management is the central theme for the vortex
4 Contrary to Fortrans user-controlled formatting, which uses an edit descriptor such as f12.8, list-

directed output formatting is controlled by compilers, which choose appropriate width and precisions
depending on the data type of the output item list.

5.2 The Solution

ring, the vortex class requires reference counting. This is readily seen by the fact
that class vortex extends RefBase. Similar to the Fortran denition, the C++ vortex
stores the vortex coordinates using an array of oats. Member variable next_, dened
as type Ref<vortex>, replaces a raw C++ pointer in the linked list node structure.
Since next_ does the reference counting on its own, it is not necessary to dene any
functionality for the destructor as is evident in the code at lines 89 in Figure 5.8(b)5 .
In the class denition of vortex, we include a typedef iterator to enable
node traversal for insertion, deletion, update, or query. Since an iterator normally
requires no reference counting when iterating through the vortex ring, we use type
pointer<Ref<vortext>>. The constructor:
vortex(real_t radius, int num_points);
creates a vortex ring with the input radius and number of points, providing the same
functionality as ring() in Figure 5.3.
The method destroy() initiates the vortex ring destruction. It may seem odd
that vortex has an empty destructor while an additional function call is needed to
initiate its destruction. Its necessity derives from the rings circular structure: Its tail
points to its head. Upon creation, there are two references to the head: one held
by the external user and another that is the vortex rings tail. Hence, the number of
references to the head does not reach zero when the external user no longer needs
the ring.

1
2
3
4
5
6
7
8
9
10
11
12
13

Figure 5.7(a): fmt.h

#ifndef _H_FMT_
#define _H_FMT_
#include "globals.h"
#include <iostream>
#include <iomanip>
// The fmt(...) helper class helps hide the mess that in <iomanip>
struct fmt {
explicit fmt(real_t value, int width=12, int prec=8) :
v_(1, value), w_(width), p_(prec)
{}
explicit fmt(crd_t value, int width=12, int prec=8) :
v_(value), w_(width), p_(prec)
{}

14
15
16
17
18
19
20
21
22
23
24

const crd_t v_;

const int w_, p_;
};
inline std::ostream& operator<<(std::ostream &os, const fmt &v) {
// Store format flags for the stream.
std::ios_base::fmtflags flags = os.flags();
// Force our own weird format.
for(crd_t::const_iterator it = v.v_.begin(); it != v.v_.end(); ++it) {
os << " " <<std::setw(v.w_) <<std::setprecision(v.p_) <<std::fixed
<< *it;
5 Its always a good practice to dene a virtual destructor, even if it is empty, in anticipation of the class

being extended later by subclasses.

123

124

The Object Pattern

25
26
27
28
29
30

1
2

}
// Restore original format flags.
os.flags(flags);
return os;

}
#endif //! _H_FMT_

#ifndef _H_GLOBALS_
#define _H_GLOBALS_

Figure 5.7(b): globals.h

3
4
5
6

#include "Ref.h"
#include "RefBase.h"
#include <vector>

7
8
9

typedef float real_t;

typedef std::vector<real_t> crd_t;

10
11

#endif //!_H_GLOBALS
Figure 5.7. Output tools: (a) C++ emulation of Fortrans list-directed output and (b) global
header le dening real_t and crd_t.

1
2

#ifndef _VORTEX_H_
#define _VORTEX_H_

Figure 5.8(a): vortex.h

3
4
5
6

#include "RefBase.h"
#include "globals.h"
#include <vector>

7
8
9
10
11
12
13
14

class vortex : virtual public RefBase

{
public:
typedef Ref<vortex> ptr_t;
typedef pointer<ptr_t> iterator;
vortex();
vortex(real_t radius, int num_points);

15

virtual vortex();

16
17

void destroy();
virtual void output() const;

18
19
20
21
22
23
24

};

private:
crd_t
ptr_t

25
26

#endif

point_;
next_;

5.2 The Solution

1
2
3

#include <cmath>
#include "vortex.h"
#include "fmt.h"

Figure 5.8(b): vortex.cpp

4
5
6

vortex::vortex()
{}

7
8
9

vortex::vortex()
{}

10
11
12
13
14
15
16

void vortex::destroy()
{
// the following line start destruction of the vortex ring
// breaks the ring and triggers the chain action
this->next_ = NULL;
}

17
18
19
20
21
22

vortex::vortex(real_t radius, int num_points)

{
const real_t pi = 3.1415926f;
real_t delta;
iterator current;

23
24

delta = 2.0f*pi*radius/real_t(num_points);

25
26
27
28

this->point_.push_back(radius);
this->point_.push_back(0.0f);
this->point_.push_back(0.0f);

29
30

this->next_ = new vortex();

31
32

current = this->next_;

33
34
35
36
37

for (int i = 0; i < num_points-1; ++i)

{
crd_t temp_point;
real_t theta;

38

theta = real_t(i+1) * delta/radius;

39
40

temp_point.push_back(radius*cos(theta));
temp_point.push_back(radius*sin(theta));
temp_point.push_back(0.0);

41
42
43
44

(*current)->point_ = temp_point;

45
46

(*current)->next_ = new vortex();

47
48
49
50

current = (*current)->next_;

51
52
53
54

*current = this; // this will cause the last memory allocation

// returned by the new operation in the last
// iteration to be automatically deleted

125

126

The Object Pattern

55

56
57
58
59
60

void vortex::output() const

{
iterator current;
crd_t tmp;

61

tmp = this->point_;

62
63

std::cout << fmt(tmp) << std::endl;

64
65

current = this->next_;

66
67

do
{

68
69
70
71

tmp = (*current)->point_;
std::cout << fmt(tmp) << std::endl;

72
73
74
75

1
2

current = (*current)->next_;
} while (*current != this);

#include <iostream>
#include "vortex.h"

Figure 5.8(c): main.cpp

3
4
5
6

int main ()
{
typedef vortex::ptr_t ptr_t;

ptr_t helium = ptr_t(new vortex(1.0, 8));

8
9

helium->output();
helium->destroy();

10
11
12
13
14

return 0;

15

Figure 5.8. Quantum vortex class denition.

In the aforementioned scenario, the reference-counting scheme would only lead

to ring destruction for a ring that is no longer circular, that is, if something breaks the
ring. The function destroy() does just that, as shown at line 15 in Figure 5.8(b). By
breaking the ring at the rst link, it effectively sets the reference count for the second
vortex point to zero. This initiates the destruction of the second point, which in turns
initiates the destruction of the third point, and so forth. The destruction process thus
cascades throughout the entire ring until it reaches the head, at which point only the
external user holds a reference to the head. Figure 5.8(c) provides a simple program
demonstrating a vortex rings creation and destruction.

5.4 Related Patterns

5.3 The Consequences

Any design pattern involves trade-offs between its benets and its drawbacks. The
OBJECT pattern benets a software package by ensuring the availability of important functionality. The project leadership can ensure access to that functionality by
mandating that all classes in the project extend the OBJECT hierarchy.
Although not exploited in the current chapter, the second benet of the OBJECT
pattern derives from the availability of a universal way to reference objects in the
project. Since all classes extend the OBJECT class hierarchy, any object in the project
could be referenced by an entity declared as an instance of the object class at the
base of the hierarchy. Chapter 11 demonstrates the utility of such a capability.
The aforementioned benets can be double-edged swords. The rst benet might
provide a false sense of security that breaks down when developers make no use of
the functionality provided by the base hierarchy. The second benet might lead to a
lot of type guarding in Fortran or type-switching in C++ if liberal passing of the base
type confronts frequent use of the additional functionality provided in a child type.
Finally, universal mandates might also be viewed as universal handcuffs. Judicious use of OBJECT implies minimalistic construction. A base hierarchy might have
the best chance of feeling nonconstricting when the functionality it provides is not
domain-specic and simply meets general programming needs. Needs addressed by
a well-designed OBJECT are likely to be low level in the sense that they could be candidates for inclusion in the language itself. This certainly is the case with garbage
collection, which Fortran includes for allocatable objects and C++ provides for
some STL vector objects.

5.4 Related Patterns

Since the intention is for all classes in a package to extend OBJECT, it is in some
sense the largest of the patterns discussed in this book and therefore relates to every other. In the context of solving differential equations, the OBJECT hierarchy will
likely be the immediate ancestor of an integrand object, an ABSTRACT CALCULUS pattern implementation that facilitates integration of differential equations. Chapter 6
explores this scenario.
In some settings (and possibly in less polite company), we facetiously refer to
the abstract class at the base of an OBJECT hierarchy as a GOD pattern instance, the
notion being that a GOD is the ultimate progenitor of all parents. Each remaining
abstract class in an OBJECT would then be DEMIGOD. The advantage of this terminology accrues from any clarity added by differentiating between the classes and roles
because OBJECT often represents a multiabstraction pattern.
Finally, the linked list data structure chosen to implement this chapters quantum
vortex codes naturally inspires thoughts of the GoF ITERATOR pattern. An ITERATOR
implementation facilitates traversing an ordered collection of objects. A typical implementation invokes methods an object provides for returning the beginning, next,
and ending elements while remaining agnostic with respect to the underlying infrastructure, which might equally well be a linked list, an array, or any ordered
collection of objects with a dened beginning and ending. Therein lies the rub: A
circular linked list has no unambiguous beginning or end. Although one might create

127

128

The Object Pattern

articial beginning and ending points by breaking a link as Section 5.2 demonstrates
in the context of object destruction, doing so merely to traverse the list would be
overkill at best and catastrophic at worst, given the possibility of starting the destruction cascade. A second approach might involve storing references to arbitrarily
chosen beginning and ending objects. We opted to avoid these approaches in favor
of simplicity.

EXERCISES
1. Draw a UML sequence diagram for a three-point ring undergoing the object
destruction cascade described in Section 5.2.
2. Write a remeshable_vortex implementation that extends the vortex ADT,
adding an insert point insertion procedure and a remove point removal
procedure.
3. Add the requisite operators in the vortex type to facilitate time advancement
via ADT calculus based on the local induction approximation.

The Abstract Calculus Pattern

All professions are conspiracies against the laity.

George Bernard Shaw

6.1 The Problem

The context of ABSTRACT CALCULUS is the construction of numerical software that
approximates various differential and integral forms. Two pairs of conicting forces
arise in this context. In the rst pair, the low-level nature of the mathematical constructs provided by mainstream programming languages constrains the design of
most scientic programs. A desire for syntax and semantics that naturally represents the much richer mathematical language of scientists and engineers opposes
this constraint.
The C++ language contains native scalar and one-dimensional array variables.
The C++ STL extends these in vectors with nice properties such as automatic memory
management, including sizing and resizing upon assignment. Fortran 2003 provides
similar capabilities with its multidimensional allocatable array construct. It also
provides numerous useful intrinsic procedures for determining array properties including size, shape, maximum element, and minimum element, as well as intrinsic
procedures and operators for combining arrays into sums, matrix vector products,
and other derived information. It is common in scientic and engineering work to
build up from these native constructs a set of array classes with a variety of additionally useful methods (Barton and Nackman 1994; Heroux et al. 2005). Nonetheless,
the resulting objects model very low-level mathematical entities in the sense that
one typically arrives at these entities after fairly involved derivations from, and approximations to, much higher-level constructs. In physics-based simulations, these
higher-level constructs include scalar, vector, and tensor elds along with the various
operations these admit: gradients, divergences, curls, surface and volume integrals,
and others. The path from these high-level abstractions to simple arrays is littered
with assumptions and approximations that make the code less general.
At an even higher level of abstraction, one uses the available mathematical
constructs to model physical constructs. These include uid and solid continuua, particles, electrical and thermal currents, and a host of other materials and phenomena.
129

130

The Abstract Calculus Pattern

One might imagine generalizing the various mathematical operators to operate on a

whole uid, particle, or conductor such that the time derivative of the physical entity
implies the time derivative of all of the mathematical entities that describe its state.
The second pair of opposing forces arises in the context of code reuse. With the
proliferation of scientic simulations as research-and-design tools and the concomitant proliferation of codes, one rarely writes custom applications from scratch. More
commonly, one leverages existing capabilities by coupling one disciplines software
applications to those in related disciplines or by building atop highly optimized libraries that provide specialized services such as nonlinear iteration, sparse matrix
inversion, and stiff system integration. As discussed in Sections 1.21.3, linking to
conventional, procedural libraries increases the data dependencies that bedevil developers during debugging (cf. Press et al. 2002). This imposes an opposing desire to
preserve data privacy and, if possible, to avoid passing data altogether.

6.2 The Solution

ABSTRACT CALCULUS resolves the conicts from the previous section by writing abstract specications for software constructs that model high-level mathematical
constructs. These specications comprise stateless, abstract classes that support a
set of arithmetic, differential, and integral operators. As such, they simply delineate the allowable operations and operands of a calculus, leaving implementors to
construct concrete, descendant types and dene the algorithms that carry out the
operations.
Clients can use the abstract class to write abstract expressions that perform sophisticated mathematical calculations without exposing the concrete type involved
in the calculation. At runtime, such a procedure can accept any concrete descendant
and invoke the requisite methods on the received object.
Consider the integration of the Lorenz equation system (4.5) over some time
interval. Suppose we encapsulate the equation system in a Lorenz ADT. Rather
than passing a Lorenz objects state vector into a conventional time integration
procedure, we might consider dening an abstract Integrand class that species
the operations required to integrate objects via ADT calculus. The integration
algorithm could then be written to operate on an Integrand and serve this same
purpose for any classes that extend Integrand. In this context, ABSTRACT CALCULUS also goes by the name SEMI-DISCRETE pattern in reference to its use for solving
the systems of ODEs that result from semidiscrete numerical approximations to
PDEs (Rouson, Adalsteinsson, and Xia 2010).
Since we aim to write an abstract integration procedure without knowledge of
the classes to be integrated, the procedure lives logically outside any class denition.
It might be packaged with, though not a part of, the Integrand abstract class. UML
uses the package concept to represent collections of elements, including classes, interfaces, collaborations, use cases, and diagrams. UML packages may be composed of
other UML elements, including other packages. Packaging also forms a namespace:
a list of unique identiers for the elements contained in the package.
Figure 6.1 describes a Lorenz system integrator design. Figure 6.1(a) employs the
UML package graphical element. A tab showing the package name Integrator adorns
the package. The Integrator contents, designated as public, include the Integrand

6.2 The Solution

(a)
Integrand
+ add(Integrand,Integrand) : Integrand
+ multiply(Integrand,real) : Integrand
+ t(Integrand) : Integrand
(b)
Integrand


Lorenz
(c)
Figure 6.1. Lorenz system integration: (a) integration package (Integrator), (b) abstract
Integrand interface, and (c) class diagram.

class and the Integration procedure. Figure 6.1(b) depicts the Integrand class interface. It contains the addition, multiplication, and time differentiation operations
required to implement Runge-Kutta marching schemes, in which context authors
often refer to the time differentiation operation as function evaluation because it
 in equation (4.4).
involves evaluating the RHS function in an ODE system, namely R
UML depicts abstract operation names such as add and multiply in italics and abstract
class names such as Integrand in bold italics.

6.2.1 Fortran Implementation

In Fortran, the module construct provides the namespace encapsulation mechanism.
Abstract classes correspond to derived types with the abstract attribute. An abstract
derived type can have deferred type-bound procedure bindings that conform to an
abstract interface. Figure 6.2 exploits these language features to implement the
design described in Figure 6.1.
Figure 6.2(a) declares a lorenz object, instantiates it by invoking its likenamed constructor, and steps it forward in time by looping over calls to integrate.
Figure 6.2(b) encapsulates an integrand abstract type and a polymorphic integrate
procedure. The integrand denition includes deferred bindings for addition of two

131

132

The Abstract Calculus Pattern

integrand objects, multiplication of an integrand by a real scalar, time differentiation of an integrand, and the assignment of one integrand to another. The
integrate procedure accepts a polymorphic object along with a real, scalar time
step and uses the aforementioned operators to overwrite the object with a version
of itself one time step in the future. It arrives at the future time step instance via the
explicit Euler algorithm expressed in ADT calculus. Figure 6.2(c) denes the lorenz
type, which implements a constructor along with each of the deferred bindings and
an output method.
The Lorenz system code in Figure 6.2(c) implements an ADT calculus much like
that presented in Chapter 3. Whereas ADT calculus decouples the client code (the
code that integrates lorenz over time) from the lorenz implemenation, ABSTRACT
CALCULUS goes one step further: It decouples the client code from depending even on
the lorenz interface at least insofar as the public types and type-bound procedure
interfaces comprise a concrete interface. Instead, the integration code depends only
on an abstract interface specication to which the concrete interface conforms.

1
2
3

Figure 6.2(a)
program main
use lorenz_module ,only : lorenz,integrate
implicit none ! Prevent implicit typing

4
5
6
7
8
9
10
11
12

type(lorenz)
:: attractor
integer
:: step ! time step counter
integer ,parameter :: num_steps=2000, &
space_dimension=3 ! phase space dimension
real
,parameter :: sigma=10.,rho=28.,beta=8./3.,&
dt=0.01 ! Lorenz parameters and time step size
real
,parameter ,dimension(space_dimension) &
:: initial_condition=(/1.,1.,1./)

13
14
15
16
17
18
19
20

1
2
3
4

attractor = lorenz(initial_condition,sigma,rho,beta)
print *,attractor%output()
do step=1,num_steps
call integrate(attractor,dt)
print *,attractor%output()
end do
end program
Figure 6.2(b)
module integrand_module
implicit none
! Prevent implicit typing
private
! Hide everything by default
public :: integrate ! expose time integration procedure

5
6
7
8
9
10
11
12

type ,abstract ,public :: integrand

contains
procedure(time_derivative
) ,deferred
procedure( symmetric_operator ) ,deferred
procedure( symmetric_assignment) ,deferred
procedure(asymmetric_operator ) ,deferred

::
::
::
::

t
add
assign
multiply

6.2 The Solution

13
14
15
16

generic :: operator(+) => add ! Map operators to proceures

generic :: operator(*) => multiply
generic :: assignment(=) => assign
end type

17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40

abstract interface
function time_derivative(this) result(dState_dt)
import :: integrand
class(integrand) ,intent(in) :: this
class(integrand) ,allocatable :: dState_dt
end function time_derivative
function symmetric_operator(lhs,rhs) result(operator_result)
import :: integrand
class(integrand) ,intent(in) :: lhs,rhs
class(integrand) ,allocatable :: operator_result
end function symmetric_operator
function asymmetric_operator(lhs,rhs) result(operator_result)
import :: integrand
class(integrand) ,intent(in) :: lhs
class(integrand) ,allocatable :: operator_result
real
,intent(in) :: rhs
end function asymmetric_operator
subroutine symmetric_assignment(lhs,rhs)
import :: integrand
class(integrand) ,intent(in)
:: rhs
class(integrand) ,intent(inout) :: lhs
end subroutine symmetric_assignment
end interface

41
42
43
44
45
46
47
48

1
2
3

contains
subroutine integrate(model,dt)
class(integrand) :: model
real ,intent(in) :: dt
! time step size
model = model + model%t()*dt ! Explicit Euler formula
end subroutine
end module integrand_module
Figure 6.2(c)
module lorenz_module
use integrand_module ,only : integrand,integrate
implicit none

4
5
6
7

private
! Hide everything by default
public :: integrate ! Expose time integration procedure
public :: lorenz

8
9
10
11
12
13
14
15
16
17

type ,extends(integrand) :: lorenz

private
real ,dimension(:) ,allocatable :: state
! solution vector
real :: sigma ,rho ,beta
! Lorenz parameters
contains
procedure ,public :: t
=> dLorenz_dt
procedure ,public :: add
=> add_lorenz
procedure ,public :: multiply => multiply_lorenz
procedure ,public :: assign
=> assign_lorenz

133

134

The Abstract Calculus Pattern

18
19

procedure ,public :: output

end type

20
21
22
23

interface lorenz
procedure constructor
end interface

24
25

contains

26
27
28
29
30
31
32
33
34
35

type(lorenz) function constructor(initial_state,s,r,b)

real ,dimension(:) ,intent(in) :: initial_state
real
,intent(in) :: s ,r ,b ! passed values for
! sigma, rho and beta
constructor%state=initial_state
constructor%sigma=s
constructor%rho=r
constructor%beta=b
end function

36
37
38
39
40
41

function output(this) result(coordinates)

class(lorenz)
,intent(in) :: this
real ,dimension(:) ,allocatable :: coordinates
coordinates = this%state
end function output

42
43
44
45
46
47

! time derivative: encapsulates Lorenz equations

function dLorenz_dt(this) result(dState_dt)
class(lorenz)
,intent(in) :: this
class(integrand) ,allocatable :: dState_dt
type(lorenz)
,allocatable :: delta

48
49
50
51
52
53
54
55
56
57
58
59
60
61
62

allocate(delta)
allocate(delta%state(size(this%state)))
! 1st lorenz equation
delta%state(1)=this%sigma*( this%state(2) -this%state(1))
! 2nd lorenz equation
delta%state(2)=this%state(1)*(this%rho-this%state(3))-this%state(2)
! 3rd lorenz equation
delta%state(3)=this%state(1)*this%state(2)-this%beta*this%state(3)
! hold Lorenz parameters constant over time
delta%sigma=0.
delta%rho=0.
delta%beta=0.
call move_alloc (delta, dState_dt)
end function

63
64
65
66
67
68

function add_Lorenz(lhs,rhs) result(sum) ! add two Lorenz objects

class(lorenz)
,intent(in) :: lhs
class(integrand) ,intent(in) :: rhs
class(integrand) ,allocatable :: sum
type(lorenz)
,allocatable :: local_sum

69
70
71
72

allocate (lorenz :: local_sum)

select type(rhs)
class is (lorenz)

6.2 The Solution

73
74
75
76
77
78
79
80
81

local_sum%state = lhs%state + rhs%state

local_sum%sigma = lhs%sigma + rhs%sigma
local_sum%rho
= lhs%rho
+ rhs%rho
local_sum%beta = lhs%beta + rhs%beta
class default
stop add_Lorenz: rhs argument type not supported
end select
call move_alloc(local_sum, sum)
end function

82
83
84
85
86
87
88

! multiply a Lorenz object by a real scalar

function multiply_Lorenz(lhs,rhs) result(product)
class(lorenz) ,intent(in) :: lhs
real
,intent(in) :: rhs
class(integrand) ,allocatable :: product
type(lorenz)
,allocatable :: local_product

89
90
91
92
93
94
95
96

allocate (local_product)
local_product%state = lhs%state*rhs
local_product%sigma = lhs%sigma*rhs
local_product%rho
= lhs%rho *rhs
local_product%beta = lhs%beta *rhs
call move_alloc(local_product, product)
end function

97
98
99
100
101

! assign one lorenz object to another

subroutine assign_lorenz(lhs,rhs)
class(lorenz)
,intent(inout) :: lhs
class(integrand) ,intent(in)
:: rhs

102
103
104
105
106
107
108
109
110
111
112
113

select type(rhs)
class is (lorenz)
lhs%state = rhs%state
lhs%sigma = rhs%sigma
lhs%rho
= rhs%rho
lhs%beta = rhs%beta
class default
stop assign_lorenz: rhs argument type not supported
end select
end subroutine
end module lorenz_module
Figure 6.2. Fortran implementation of ABSTRACT CALCULUS applied to a Lorenz system:
(a) Integration test for the Lorenz system; (b) Integrand class; (c) Lorenz class.

Figure 6.2(c) represents the rst occurrence of the select type construct and
move_alloc intrinsic procedure in this text. A conuence of circumstances likely to
occur in most abstract calculi necessitates these language facilities. First, the deferred
binding add in the abstract type integrand in Figure 6.2(b) species an abstract
interface symmetric_operator that forces procedures implementing the deferred
binding to accept a dummy argument declared as class(integrand). This restriction
cannot apply to the passed-object dummy argument lhs, which must be declared as
class(lorenz) in order for the compiler to differentiate this implementation of the

135

136

The Abstract Calculus Pattern

deferred binding from other derived types implementations. The restriction applies
to the other argument rhs.
Inside the select type construct, the code accesses lorenz components not
included in the integrand abstract type specication. Since this code would fail for
any extensions of integrand that do not contain these components, the language
semantics require guarding against this possibility. Although the type determination
happens at runtime, the class is(lorenz) type guard statement ensures that the
guarded code only executes when rhs is a lorenz object or an instance of a type that
extends lorenz and therefore contains the required components.
Were add_Lorenz to compute its result (sum) directly, an additional (nested)
type guard statement would be required to ensure that the result (delcared as
class(integrand)) has the requisite components. The desired result type is known,
so this level of type guarding feels somewhat superuous even though the deferred
binding requires it. We therefore eliminate the need for it by declaring and computing a local_sum of type lorenz and transferring its allocation to the actual result
sum via the move_alloc intrinsic procedure. After the move_alloc invocation at the
end of add_Lorenz, sum will be associated with the memory originally associated
with local_sum, whereas the latter will be unallocated. This move_alloc invocation
ensures that the chosen method for avoiding extra type guarding does not impose
additional dynamic memory utilization. For this reason, the move_alloc intrinsic
will also play a signicant role in memory recycling in Chapter 10.
In Fortran, the select type construct provides the only means for declared runtime type conversions because the language contains no explicit type casting facility.
Similar to a select case construct, a select type allows a programmer to specify
a list of possible blocks of code to execute based on the possible dynamic types of
the object at runtime. Each constituent block is bound with a statement, which can
be type is or class is, specifying a unique data type. A special type guard statement, class default, also provides an error-handling facility in case all specied
types fail to match the dynamic type of the object1 . The Fortran 2003 standard also
guarantees that, at most, one of the constituent blocks executes. If none of the blocks
is selected and no class default statement is provided, the program simply exits
the construct quietly and continues to subsequent lines. Although the select type
construct guards against crashes, it remains the programmers responsibility to insert
a class default case to detect and handle the scenario in which no block is selected.
In C++, the corresponding technique involves attempting a dynamic_cast to
some useful type. If the ultimate type proves incompatible with the type being
cast, however, dynamic_cast either returns NULL for casts involving a pointer or
throws an exception for casts involving a reference. Handling such exceptions necessitates a conditional test (e.g., an if statement) or a exception-handling code (i.e., a
try-catch block). This design approach suffers from fragility: It relies heavily on an
assumption that all programmers will use safe practices. Neglecting the conditional
or the exception handling can lead to catastrophic results wherein the code crashes
or an incorrect result goes undetected until some later (and potentially distant) point
in the code.
1 Although C++ programmers commonly decry the lack of exception handling in Fortran, this scenario

exemplies the general rule that exception handling is less necessary in Fortran than in C++.

6.2 The Solution

Even though both language constructs rely on runtime type information (RTTI)
and both have pitfalls, a select type, in general, is safer than a dynamic_cast in
that a select type only performs a downcast along the class hierarchy tree based
upon the actual dynamic type (the runtime type) of the object or pointer. Downcasts
are the most common situations arising from need for explicit type changes (casts)
so that the programmer can access additional data members and functions of an
extended type. In C++, dynamic_cast operators can be used for downcasts as well
as upcasts. A common casting error occurs when one casts to a type on a different
branch of an inheritance hierarchy2 . The distinction between the two constructs is
also apparent from the language design point of view: C++s dynamic_cast relies
solely on programmers knowledge of the actual runtime type to which an object
is cast, whereas Fortrans select type asks programmers to supply a list of possible dynamic types with corresponding execution blocks to be selected. Therefore,
dynamic_cast is more exible to use at the cost of placing a greater number of
responsibilities on programmers.

6.2.2 C++ Implementation

Figure 6.3 provides a C++ implementation of ABSTRACT CALCULUS, following
the same structure as the Fortran code in Figure 6.2. One signicant difference is the inclusion of exception handling in the form of try-catch blocks
in the main.cpp in Figure 6.3(a). With exception handling, runtime errors in
integrate(attractor, dt) cause the program control to transfer to error-handling
code blocks. The code blocks that catch and handle the error can be local to the routine in which the error occurred or somewhere further up the call tree in code that
invoked the routine containing the try. Fortran lacks a corresponding capability.

1
2
3

#include "lorenz.h"
#include "fmt.h"
#include <iostream>

Figure 6.3(a): main.cpp

4
5
6
7

int main () {
using namespace std;
typedef lorenz::ptr_t ptr_t;

8
9
10
11

const int
num_steps=2000, space_dimension=3;
const float sigma=10, rho=28, beta=8.0/3.0, dt=0.01;
const crd_t initial_condition(space_dimension, 1.0);

12
13
14
15

ptr_t

attractor = ptr_t(new lorenz(initial_condition,

sigma,rho,beta));
const crd_t &output
= attractor->output();

16
17
18

try {
std::cout << fmt(output, 12, 10) << "\n";
2 An upcast followed by a downcast can end up with cross-branch casting.

137

138

The Abstract Calculus Pattern

19
20
21
22
23
24
25
26
27
28
29
30
31
32
33

1
2

for (int step = 1; step <= num_steps; ++step) {

integrate (attractor, dt);
std::cout << fmt(output, 12, 10) << "\n";
}
} catch(std::exception &e) {
std::cerr << "Error exit following exception of type " << e.what()
<< "\n";
return EXIT_FAILURE;
} catch(...) {
std::cerr << "Error exit following an unknown exception type\n";
return EXIT_FAILURE;
}
return EXIT_SUCCESS;

Figure 6.3(b): integrand.h

#ifndef __H_integrand__
#define __H_integrand__ 1

3
4

#include "globals.h"

5
6
7
8

class integrand : virtual public RefBase {

public:
typedef Ref<integrand> ptr_t;

9
10

virtual integrand();

11
12
13
14

virtual ptr_t d_dt() const = 0;

virtual void operator+=(ptr_t) = 0;
virtual ptr_t operator*(real_t) const = 0;

15
16
17
18
19

protected:
integrand(const integrand&);
integrand();
};

20
21

void integrate (integrand::ptr_t, real_t);

22
23

#endif

#include "integrand.h"

Figure 6.3(c): integrand.cpp

2
3

typedef integrand::ptr_t ptr_t;

4
5
6

integrand::integrand() : RefBase() {
}

7
8
9

integrand::integrand(const integrand&) : RefBase(){

}

10
11
12
13

integrand::integrand() {
}

6.2 The Solution

14
15
16

1
2

void integrate (ptr_t model, real_t dt) {

*model += *(model->d_dt()) * dt;
}

#ifndef __H_LORENZ__
#define __H_LORENZ__ 1

Figure 6.3(d): lorenz.h

3
4

#include "integrand.h"

5
6
7
8
9
10
11

class lorenz : public integrand {

public:
typedef Ref<lorenz> ptr_t;
lorenz ();
lorenz (const crd_t, real_t sigma, real_t rho, real_t beta);
// Default copy and assignment operators are just fine for this type.

12
13
14
15
16

public:
integrand::ptr_t
void
integrand::ptr_t

d_dt() const;
operator+=(integrand::ptr_t other);
operator*(float val) const;

17
18

const crd_t&

output() const;

19
20

virtual lorenz();

21
22
23
24
25

private:
crd_t state_;
// solution vector.
float sigma_, rho_, beta_;
};

26
27

1
2

#endif

#include <iostream>
#include <exception>

Figure 6.3(e): lorenz.cpp

3
4

#include "lorenz.h"

5
6

using namespace std;

7
8
9
10

struct LorenzError : public std::exception {

virtual LorenzError() throw() {}
};

11
12
13
14

// default constructor
lorenz::lorenz ()
{}

15
16
17
18
19

// constructor using each element

lorenz::lorenz (const crd_t initial_state, real_t s, real_t r, real_t b)
: state_(initial_state), sigma_(s), rho_(r), beta_(b)
{}

20
21

const crd_t& lorenz::output() const {

139

140

The Abstract Calculus Pattern

22
23

return state_;

24
25
26
27
28
29
30
31
32
33
34

integrand::ptr_t lorenz:: d_dt() const

{
ptr_t result = ptr_t(new lorenz);
result->state_.resize(3);
result->state_.at(0) = sigma_*(state_.at(1) - state_.at(0));
result->state_.at(1) = state_.at(0)*(rho_ - state_.at(2))
- state_.at(1);
result->state_.at(2) = state_.at(0)*state_.at(1) - beta_*state_.at(2);
return result;
}

35
36
37
38
39
40
41
42
43
44
45
46

void lorenz::operator+=(integrand::ptr_t rhs) {

ptr_t other = cast<lorenz>(rhs);
if(other == NULL) {
std::cerr << "lorenz::operator+=: Failed dynamic cast\n";
throw LorenzError();
}
if(other->state_.size() != this->state_.size()) {
std::cerr << "lorenz::operator+=: Non-identical dimensions.\n";
throw LorenzError();
}

47
48
49
50
51

for(size_t i = 0; i < state_.size(); ++i) {

state_.at(i) += other->state_.at(i);
}

52
53
54
55
56
57
58
59
60

integrand::ptr_t lorenz::operator*(real_t rhs) const

{
ptr_t result = ptr_t(new lorenz(*this));
for(size_t i = 0; i < result->state_.size(); ++i) {
result->state_.at(i) *= rhs;
}
return result;
}

61
62
63

lorenz::lorenz()
{}
Figure 6.3. C++ implementation of an abstract calculus for the Lorenz system: (a) Integration
test, (b) Integrand declaration, (c) Integrand implementation, (d) Lorenz declaration, and (e)
Lorenz implementation.

6.3 The Consequences

In ABSTRACT CALCULUS, obvious benets accrue from the simplicity, clarity, and generality of the interface. Only a few operators must be dened to construct any new
dynamical system. Their purpose very closely mirrors the corresponding operators
in the mathematical statement of the system and the time marching algorithm. Often
the internals of each operator relies on sufciently simple calculations that they can

6.4 Related Patterns

delegate their work to highly optimized versions of linear algebra libraries such as
BLAS (Blackford et al. 2002) and LAPACK (Barker et al. 2001) or scalable parallel libraries such as ScaLAPACK (Blackford et al. 1997), PETSc (Balay et al. 2007),
or Trilinos (Heroux et al. 2005). We detail the process of building atop Trilinos in
Chapter 11.
A common thread running through many design patterns is their fostering of
loose couplings between abstractions. Unlike procedural time-integration libraries
and even many object-oriented ones, integrate() never sees the data it integrates.
It therefore relies on no assumptions about the datas layout in memory, its type, or
its precision. The programmer retains the freedom to restructure the data or change
its type or precision with absolutely no impact on the time-integration code.
Sometimes, one developers benet is anothers drawback. For example, if implemented naively, one potential drawback of ABSTRACT CALCULUS lies in making it
more challenging for the compiler to optimize. Section 3.3 delineated several of the
corresponding challenges for ADT calculus and potential workarounds. Naturally,
these same potential pitfalls and solutions apply to ABSTRACT CALCULUS, because this
pattern simply creates a unied interface to an ADT calculus supported by any classes
that implement the interface.
The benets of the ABSTRACT CALCULUS over type-specic ADT calculus stem
from the uniformity the pattern imposes across the elements of a package and the
resulting ability to vary the dynamic type of a mathematical entity at execution time.
The cost paid for this is the runtime overhead involved in discovering that type on the
y and, more importantly, the seeming impossibility of performing the static analysis
of the code required for the more advanced optimization strategies particularly interprocedural optimizations. This returns again to the theme of coarse-grained data
structure design that facilitates large numbers of computations inside each operator, tied to a sound understanding of theoretical operation-count predictions and a
disciplined approach to empirical proling.
Ultimately the highest virtue to which ABSTRACT CALCULUS aspires harkens back
to Timeless Way (Section 4.2.1), which attempts to breathe life into designs by releasing ones innate understanding of the application domain. For those trained in
scientic simulation, the underlying mathematics eventually feels innate. ABSTRACT
CALCULUS takes the conversations that historically occurred at the blackboard to
the screen. This is a dialogue spoken in pattern languages. Although each person
possesses a pattern language unique to their perspective and life experiences (some
prefer vector notation, others tensor notation), providing an elementary set of common operations enables users to extend the classes so developed in ways that reect
their predilections. In doing so, it draws back the curtain and attracts software users
into a dialogue previously limited to expert developers.

6.4 Related Patterns

ABSTRACT CALCULUS and PUPPETEER, the two domain-specic patterns in this text,
have close counterparts in the GoF patterns. Most design patterns discussions list
such pattern relationships under the heading Also Known As. Outside of the
scientic simulation domain, a pattern with the same structural form and purpose

141

142

The Abstract Calculus Pattern

as ABSTRACT CALCULUS would be known as the GoF FACADE pattern. Much as the
facade described in Section 4.2.3 presents a simplied and unied interface to two
complicated subsystems, so does ABSTRACT CALCULUS greatly simplify and unify the
top-level interface to a multiphysics solver. As formulated, this pattern supports
writing expressions on a single class of objects at a time without reference to, or a
mechanism for, combining classes.
Chapter 8 outlines how to aggregate multiple classes into one entity, a PUPPETEER,
for writing ABSTRACT CALCULUS expressions. From one viewpoint, a PUPPETEER serves
merely as the concrete side of the double-faceted FACADE structure described in
section 4.2.3. It could therefore be treated as part of an ABSTRACT CALCULUS. However,
it plays an additional role that warrants treating it as a separate pattern. Chapter 8
explains further.
The ordering of our patterns chapters draws inspiration from A Pattern Language (Alexander et al. 1975). Just as they started broadest patterns for towns and
regions and proceeded to smaller patterns that complete the larger ones, ABSTRACT
CALCULUS sets functionality requirements that propagate downward throughout the
remainder of the design. In one sense, the remainder of Part II concerns laying out
patterns necessary to complete an ABSTRACT CALCULUS in a complex, multiphysics
setting.
Were we presenting a complete pattern language in the style of
Alexander et al. (1977), we would express the relationships between patterns in a
graph with nodes representing patterns and edges drawn between the nodes to indicate which patterns complete which other patterns. Instead, we focus Part II on a
core set of patterns that would likely all be involved in completing a solver.

EXERCISES
1. Write an abstract type or pure virtual class that denes the operators and methods
required to time advance the heat conduction equation used throughout Part I of
this text.
2. Write a concrete ADT that implements the interface dened in Exercise 1 and
solves the heat conduction equations. Explore varying approaches to code reuse:
(a) aggregate an instance of one of the heat conduction ADTs from Part I and delegate calculations to that object, and (b) modify one of the earlier heat conduction
ADTs so that it extends the interface dened in Exercise 1 and directly associates
any deferred type-bound procedures or pure virtual functions with those in the
modied ADT. Discuss the benets and drawbacks of each approach in terms of
code construction costs and maintainability.

The Strategy and Surrogate Patterns

However beautiful the strategy, you should occasionally look at the

results.
Winston Churchill

7.1 The Problem

This chapter introduces the GoF STRATEGY pattern along with a Fortran-specic,
enabling pattern: SURROGATE. In scientic programming, one nds context for a
STRATEGY when the choice of numerical algorithms must evolve dynamically. Multiphysics modeling, for example, typically implies multi-numerics modeling. As the
physics changes, so must the numerical methods.
A problem arises when the software does not separate its expression of the
physics from its expression of the discrete algorithms. In the Lorenz system, for
example, equation (4.5) expresses the physics, whereas equation (4.6) expresses
the discretization. Consider a concrete ADT that solves the same equations as the
abstractions presented in Section 6.2 and extends the integrand ADT but requires
a new time integration algorithm. The extended type must overload the integrate
name. This would be a simple task when changing one algorithm in one particular
ADT. However, when numerous algorithms exist, one faces the dilemma of either
putting all possible algorithms in the parent type or leaving to extended types the
task of each implementing their own algorithms. Section 7.3 explains the adverse
impact the rst option has on code maintainability. The second option could lead to
redundant (and possibly inconsistent) implementations.
The Lorenz equation solver described in Chapter 6 uses explicit Euler time
advancement. That solver updates the solution vector at each time step without
explicitly storing a corresponding time coordinate. Such an approach might be appropriate when one requires only a representative set of points in the problem phase
space to calculate geometrical features of the solution such as the fractal dimension
of its strange attractor. By contrast, if one desires temporal details of the solution
trajectory, then it might be useful to create an extended type that stores a time stamp
and uses a marching algorithm with higher-order accuracy in time such as 2nd -order
143

144

The Strategy and Surrogate Patterns

Runge-Kutta (RK2). RK2 advances the solution in time according to:

  t
 =V
n +R
 V
n
V
2
 
n +R
 V
  t
 n+1 = V
V

(7.1)
(7.2)

which has been written in a predictor/corrector form involving two substeps but can
also easily be combined into a single step.
 in an object and time advancing that object via
Encapsulating the state vector V
ABSTRACT CALCULUS makes sense if the overloaded and dened operations in that
calculus have reasonable mathematical interpretations when applied to each state
vector component. More specically, the overloaded forms of equations (7.1)(7.2)
have meaningful interpretations if each component of the state vector satises a
differential equation. Time is a state vector component in the current context, so we
augment the governing equations (4.5) with a fourth equation:
d
=1
dt

(7.3)

where is the objects time stamp. The challenge is to extend the lorenz type by
adding the new component , adding its governing equation (7.3), and adding the
ability to select an integration strategy dynamically at runtime.
Obviously the driving force in this context is the need for exibility. An opposing force is a desire for simplicity. Placing the responsibility for implementing
new numerical quadrature algorithm in the hands of each integrand descendant
requires little effort on the part of the integrand designer and is thus the simplest
approach from a design standpoint. Conversely, requiring the integrand designer to
implement each new algorithm requires greater effort on the part of the integrand
designer, but still retains the simplicity of having all the algorithms in one abstraction
and under the control of that abstractions developers.
The next section discusses how STRATEGY resolves the conicting forces in the
current design context. The Fortran implementation of STRATEGY provides the context for the SURROGATE pattern. We therefore discuss the problem a SURROGATE
solves in the Fortran implementation section 7.2.1.

7.2 The Solution

The GoF resolved this problem with the STRATEGY pattern. They stated the essence
of this pattern as follows:
Dene a family of algorithms, encapsulate each one, and make them interchangeable.
Strategy lets the algorithm vary independently from clients that use it.

A STRATEGY severs the link between algorithms and data. Data objects delegate
operations to Strategy classes that apply appropriate algorithms to the data.

7.2 The Solution

Our STRATEGY implementation denes a timed_lorenz type that extends the

lorenz type. As before, the lorenz concrete type extends the integrand abstract
type. In addition to the deferred bindings that integrand stipulates must be implemented by its extended types, it now contains a reference to an abstract derived type,
strategy, to which integrand delegates the responsibility to provide a type-bound
integrate() procedure. The strategy denes only the interface (via deferred binding) for the time-integration method, leaving its own extended types to provide actual
quadrature schemes. In applying the strategy pattern, one passes a reference to a
lorenz dynamical system as an actual argument to the integrate() method of the
strategy object. Again, the program syntax mirrors the mathematical syntax:
class(surrogate) ,intent(inout) :: this
real ,intent(in) :: dt
...
this_half = this + this%t()*(0.5*dt)
this = this + this_half%t()*dt
We explain the role of the new surrogate type next.
Figure 7.1 depicts a UML class model of our Fortran STRATEGY implementation,
including an empty abstract surrogate type, instances of which substitute for all references to integrand objects and their extended types when referenced inside the
strategy hierarchy. Declaring all such instances with the Fortran class keyword
defers to runtime the resolution of their ultimate dynamic type, which can be any
descendant of the surrogate type. The role played by the surrogate provides an
example of a Fortran-specic design pattern. A SURROGATE circumvents Fortrans
prohibition against circular references, wherein one modules reference to a second
module via a use statement precludes referencing the rst module in the second via
a reciprocal use. In C++, one avoids such circular references by using forward references. Forward references allow one ADT to declare another and then to manipulate
references to the declared ADT without knowing its denition.

surrogate

strategy

integrand

lorenz

timed_lorenz

euler_integrate

runge_kutta2_integrate

Figure 7.1. Fortran class model for applying the STRATEGY and SURROGATE patterns to extend
the Lorenz model.

145

146

The Strategy and Surrogate Patterns

7.2.1 Fortran Implementation

Figure 7.2 provides STRATEGY and SURROGATE pattern demonstrations in Fortran. The
gure lists the codes top-down, starting with main in Figure 7.2(a) and progressing
through the UML class diagram of Figure 7.1 from the surrogate in Figure 7.2(b),
down through the integrand in Figure 7.2(c), and the strategy in Figure 7.2(d).
From there, the listing proceeds down the left-hand branch of the class diagram
through the lorenz implementation of Figure 7.2(e) and the timed_lorenz implementation of Figure 7.2(f). Figures 7.2(g) and 7.2(h) traverse the right branch,
comprising the explicit_euler and runge_kutta_2nd implementations.

1
2
3
4
5
6

Figure 7.2(a): main.f03

program main
use lorenz_module ,only : lorenz
use timed_lorenz_module ,only : timed_lorenz
use explicit_euler_module ,only : explicit_euler
use runge_kutta_2nd_module ,only : runge_kutta_2nd
implicit none ! Prevent implicit typing

7
8
9
10
11
12
13

type(explicit_euler)
type(runge_kutta_2nd)
type(lorenz)
type(timed_lorenz)
integer
integer ,parameter

::
::
::
::
::
::

lorenz_integrator
!Integration strategy
timed_lorenz_integrator !Integration strategy
attractor ! Lorenz equation/state abstraction
timed_attractor ! Time-stamped abstraction
step
! Time step counter
num_steps=2000,space_dimension=3

14
15
16
17
18

! Lorenz parameters and step size

real
,parameter
:: sigma=10.,rho=28.,beta=8./3.,dt=0.01
real
,parameter ,dimension(space_dimension) &
:: initial=(/1.,1.,1./)

19
20
21
22
23
24

! Initialize and choose strategy

attractor = &
lorenz(initial,sigma,rho,beta,lorenz_integrator)
print *,lorenz attractor:
print *,attractor%output()

25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41

! Run explicit Euler at increased resolution for comparison to RK2

do step=1,4*num_steps
call attractor%integrate(dt/4.)
print *,attractor%output()
end do
! Re-initialize, choose new strategy
timed_attractor = &
timed_lorenz(initial,sigma,rho,beta,timed_lorenz_integrator)
print *,
print *,timed_lorenz attractor:
print *,timed_attractor%output()
do step=1,num_steps
call timed_attractor%integrate(dt)
print *,timed_attractor%output()
end do
end program main

7.2 The Solution

1
2
3
4

module surrogate_module
implicit none
private
public :: surrogate

Figure 7.2(b): surrogate.f03

! Prevent implicit typing
! Hide everything by default

5
6
7
8
9

!
!
!
!

This stateless type serves only for purposes of extension by other

types. In such a role, it can serve as a substitute for the child
type when that type is inaccessible because of Fortrans
prohibition against circular references.

10
11
12
13

1
2
3
4
5

type ,abstract :: surrogate

end type
end module
Figure 7.2(c): integrand.f03
module integrand_module
use surrogate_module ,only : surrogate ! Integrand parent
use strategy_module ,only : strategy
! Integration strategy parent
implicit none ! Prevent implicit typing
private
! Hide everything by default

6
7
8
9
10
11
12
13
14
15
16

type ,abstract ,public, extends(surrogate) :: integrand

private
class(strategy), allocatable :: quadrature
contains
procedure, non_overridable :: integrate
! Time integrator
procedure, non_overridable :: set_quadrature
procedure, non_overridable :: get_quadrature
procedure(time_derivative) ,deferred :: t ! Time derivative that
! evaluates evolution
! equations

17
18
19
20

procedure(symmetric_operator) ,deferred :: add

procedure(asymmetric_operator),deferred :: multiply
procedure(symmetric_assignment) ,deferred :: assign

21
22
23
24
25
26

! Map operators to corresponding procedures

generic
:: operator(+) => add
generic
:: operator(*) => multiply
generic
:: assignment(=) => assign
end type

27
28
29
30
31
32
33
34
35
36
37
38
39

abstract interface
function time_derivative(this) result(dState_dt)
import :: integrand
class(integrand) ,intent(in) :: this
class(integrand) ,allocatable :: dState_dt
end function time_derivative
function symmetric_operator(lhs,rhs) result(operator_result)
import :: integrand
class(integrand) ,intent(in) :: lhs,rhs
class(integrand) ,allocatable :: operator_result
end function symmetric_operator
function asymmetric_operator(lhs,rhs) result(operator_result)

147

148

The Strategy and Surrogate Patterns

40
41
42
43
44
45
46
47
48
49
50

import :: integrand
class(integrand) ,intent(in) :: lhs
class(integrand) ,allocatable :: operator_result
real
,intent(in) :: rhs
end function asymmetric_operator
subroutine symmetric_assignment(lhs,rhs)
import :: integrand
class(integrand) ,intent(in)
:: rhs
class(integrand) ,intent(inout) :: lhs
end subroutine symmetric_assignment
end interface

51
52

contains

53
54
55
56

subroutine set_quadrature (this, s)

class(integrand), intent(inout) :: this
class(strategy), intent(in) :: s

57
58

if (allocated(this%quadrature)) deallocate (this%quadrature)

59
60
61

allocate (this%quadrature, source=s)

end subroutine

62
63
64
65

function get_quadrature (this) result (this_strategy)

class(integrand), intent(in) :: this
class(strategy), allocatable :: this_strategy

66
67
68

allocate (this_strategy, source=this%quadrature)

end function

69
70
71
72
73
74
75
76
77
78
79

1
2
3

subroutine integrate(model,dt)
class(integrand) :: model
! integrand
real ,intent(in)
:: dt
! time step size
if (allocated(model%quadrature)) then
call model%quadrature%integrate(model,dt)
else
stop integrate: no integration procedure available.
end if
end subroutine
end module integrand_module

Figure 7.2(d): strategy.f03

module strategy_module
! Substitute for integrand (avoiding circular references)
use surrogate_module ,only : surrogate

4
5
6

implicit none
private

! Prevent implicit typing

! Hide everything by default

7
8
9
10
11
12

! Abstract time integration strategy

type, abstract ,public :: strategy
contains
! Abstract integration procedure interface
procedure(integrator_interface), nopass, deferred :: integrate

7.2 The Solution

13

end type strategy

14
15
16
17
18
19
20
21
22

1
2
3
4
5

abstract interface
subroutine integrator_interface(this,dt)
import :: surrogate
class(surrogate) ,intent(inout) :: this ! integrand
real
,intent(in)
:: dt
! time step size
end subroutine
end interface
end module

Figure 7.2(e): lorenz.f03

module lorenz_module
use strategy_module ,only : strategy
! time integration strategy
use integrand_module ,only : integrand ! Abstract integrand
implicit none ! Prevent implicit typing
private
! Hide everything by default

6
7
8

public :: integrand ! Expose integrand type

public :: lorenz
! Expose lorenz type and constructor

9
10
11
12
13
14
15
16
17
18
19
20

type ,extends(integrand) ,public :: lorenz

private
real ,dimension(:) ,allocatable :: state
! solution vector
real :: sigma ,rho ,beta
! Lorenz parameters
contains
procedure ,public :: t => dlorenz_dt ! Time deriv. (eval. RHS)
procedure ,public :: add
=> add_lorenz
procedure ,public :: multiply => multiply_lorenz
procedure ,public :: assign
=> assign_lorenz
procedure ,public :: output
! Accessor: return state
end type

21
22
23
24

interface lorenz
procedure constructor
end interface

25
26

contains

27
28
29
30
31
32
33
34
35
36
37
38
39

! Constructor: allocate and initialize

type(lorenz) function constructor(initial_state,s,r,b,this_strategy)
real ,dimension(:) ,intent(in) :: initial_state
real
,intent(in) :: s ,r ,b
class(strategy)
,intent(in) :: this_strategy
! constructor%state is automatically allocated by the assignment
constructor%state=initial_state
constructor%sigma=s
constructor%rho=r
constructor%beta=b
call constructor%set_quadrature(this_strategy)
end function

40
41
42

! Time derivative (specifies evolution equations)

function dLorenz_dt(this) result(dState_dt)

149

150

The Strategy and Surrogate Patterns

43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62

class(lorenz)
,intent(in) :: this
class(integrand) ,allocatable :: dState_dt
type(lorenz)
,allocatable :: local_dState_dt
allocate(local_dState_dt)
call local_dState_dt%set_quadrature(this%get_quadrature())
allocate(local_dState_dt%state(size(this%state)))
! 1st Lorenz equation
local_dState_dt%state(1) = this%sigma*(this%state(2)-this%state(1))
! 2nd Lorenz equation
local_dState_dt%state(2) = this%state(1)*(this%rho-this%state(3)) &
-this%state(2)
! 3rd Lorenz equation
local_dState_dt%state(3) = this%state(1)*this%state(2)
&
-this%beta*this%state(3)
local_dState_dt%sigma
= 0.
local_dState_dt%rho
= 0.
local_dState_dt%beta
= 0.
! transfer the allocation from local_dState_dt to dState_dt
call move_alloc(local_dState_dt,dState_dt)
end function

63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84

! Add two instances of type lorenz

function add_lorenz(lhs,rhs) result(sum)
class(lorenz)
,intent(in) :: lhs
class(integrand) ,intent(in) :: rhs
class(integrand) ,allocatable :: sum
type(lorenz)
,allocatable :: local_sum
select type(rhs)
class is (lorenz)
allocate(local_sum)
call local_sum%set_quadrature(lhs%get_quadrature())
! local_sum%state is automatically allocated by assignment
local_sum%state = lhs%state + rhs%state
local_sum%sigma = lhs%sigma + rhs%sigma
local_sum%rho
= lhs%rho
+ rhs%rho
local_sum%beta = lhs%beta + rhs%beta
class default
stop assig_lorenz: unsupported class
end select
! no additional allocation needed by using move_alloc
call move_alloc(local_sum,sum)
end function

85
86
87
88
89
90
91
92
93
94
95
96
97

! Multiply an instance of lorenz by a real scalar

function multiply_lorenz(lhs,rhs) result(product)
class(lorenz)
,intent(in) :: lhs
real
,intent(in) :: rhs
class(integrand) ,allocatable :: product
type(lorenz)
,allocatable :: local_product
allocate(local_product)
call local_product%set_quadrature(lhs%get_quadrature())
! local_product%state is automatically allocated by assignment
local_product%state = lhs%state* rhs
local_product%sigma = lhs%sigma* rhs
local_product%rho
= lhs%rho * rhs

7.2 The Solution

98
99
100
101

local_product%beta = lhs%beta * rhs

! avoid unnecessary memory allocation by using move_alloc
call move_alloc(local_product,product)
end function

102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118

! Assign one instance to another

subroutine assign_lorenz(lhs,rhs)
class(lorenz)
,intent(inout) :: lhs
class(integrand) ,intent(in)
:: rhs
select type(rhs)
class is (lorenz)
! let assignment automatically allocate lhs%state
lhs%state = rhs%state
lhs%sigma = rhs%sigma
lhs%rho
= rhs%rho
lhs%beta = rhs%beta
call lhs%set_quadrature(rhs%get_quadrature())
class default
stop assign_lorenz: unsupported class
end select
end subroutine

119
120
121
122
123
124
125
126
127

1
2
3
4
5
6
7

! Accessor: return state

function output(this) result(coordinates)
class(lorenz)
,intent(in) :: this
real ,dimension(:) ,allocatable :: coordinates
! assignment allocates coordinates automatically
coordinates = this%state
end function output
end module lorenz_module

Figure 7.2(f): timed_lorenz.f03

module timed_lorenz_module
use lorenz_module ,only :lorenz,integrand ! Parent and grandparent
use strategy_module,only :strategy
! Time integration strategy
implicit none
! Prevent implicit typing
private
! Hide everything by default
public :: integrand
! Expose abstract integrand and integrator
public :: timed_lorenz

8
9
10
11
12
13
14
15
16
17
18

type ,extends(lorenz) :: timed_lorenz

private
real :: time
! time stamp
contains
procedure ,public :: t => dTimed_lorenz_dt ! time deriv. (eval. RHS)
procedure ,public :: add
=> add_timed_lorenz
procedure ,public :: multiply => multiply_timed_lorenz
procedure ,public :: assign
=> assign_timed_lorenz
procedure ,public :: output
! accessor: return state
end type timed_lorenz

19
20
21
22

interface timed_lorenz
procedure constructor
end interface

151

152

The Strategy and Surrogate Patterns

23
24

contains

25
26
27
28
29
30
31
32
33
34
35

! Constructor: allocate and initialize state

type(timed_lorenz) function constructor(initial,s,r,b,this_strategy)
real ,dimension(:) ,intent(in) :: initial
real
,intent(in) :: s ,r ,b ! Lorenz parameters:
! sigma, rho and beta
class(strategy),intent(in) :: this_strategy ! marching algorithm
constructor%lorenz = lorenz &
(initial_state=initial,s=s,r=r,b=b,this_strategy=this_strategy)
constructor%time
= 0.
end function

36
37
38
39
40
41

! time derivative (expresses evolution equations)

function dTimed_lorenz_dt(this) result(dState_dt)
class(timed_lorenz),intent(in) :: this
class(integrand)
,allocatable :: dState_dt
type(timed_lorenz) ,allocatable :: local_dState_dt

42
43
44
45
46
47
48

allocate(local_dState_dt)
local_dState_dt%time
= 1.
local_dState_dt%lorenz = this%lorenz%t()
! avoid unnecessary memory allocation
call move_alloc(local_dState_dt,dState_dt)
end function

! dt/dt = 1.
! delegate to parent

49
50
51
52
53
54
55

! add two instances of timed_lorenz

function add_timed_lorenz(lhs,rhs) result(sum)
class(timed_lorenz)
,intent(in) :: lhs
class(integrand)
,intent(in) :: rhs
class(integrand)
,allocatable :: sum
type(timed_lorenz)
,allocatable :: local_sum

56
57
58
59
60
61
62
63
64

select type(rhs)
class is (timed_lorenz)
allocate(local_sum)
local_sum%time
= lhs%time
+ rhs%time
local_sum%lorenz = lhs%lorenz + rhs%lorenz
class default
stop add_timed_lorenz: type not supported
end select

65
66
67
68

! avoid unnecessary memory allocation

call move_alloc(local_sum,sum)
end function

69
70
71
72
73
74
75

! multiply one instance of timed_lorenz by a real scalar

function multiply_timed_lorenz(lhs,rhs) result(product)
class(timed_lorenz)
,intent(in) :: lhs
real
,intent(in) :: rhs
class(integrand)
,allocatable :: product
type(timed_lorenz)
,allocatable :: local_product

76
77

allocate(local_product)

7.2 The Solution

78
79

local_product%time
= lhs%time * rhs
local_product%lorenz = lhs%lorenz* rhs

80
81
82
83

! transfer allocation from local_product to result product

call move_alloc(local_product,product)
end function

84
85
86
87
88
89
90
91
92
93
94
95
96

! assign one instance of timed_lorenz to another

subroutine assign_timed_lorenz(lhs,rhs)
class(timed_lorenz)
,intent(inout) :: lhs
class(integrand)
,intent(in)
:: rhs
select type(rhs)
class is (timed_lorenz)
lhs%time
= rhs%time
lhs%lorenz = rhs%lorenz
class default
stop assign_timed_lorenz: type not supported
end select
end subroutine

97
98
99
100
101

! return state
function output(this) result(coordinates)
class(timed_lorenz) ,intent(in) :: this
real ,dimension(:) ,allocatable :: coordinates

102
103
104
105
106

1
2
3
4

! assignment automatically allocates coordinates

coordinates = [ this%time, this%lorenz%output() ]
end function
end module timed_lorenz_module

Figure 7.2(g): explicit_euler.f03

module explicit_euler_module
use surrogate_module ,only : surrogate ! time integration strategy
use strategy_module ,only : strategy
! time integration strategy
use integrand_module ,only : integrand ! abstract integrand

5
6
7

implicit none ! Prevent implicit typing

private
! Hide everything by default

8
9
10
11
12
13

! 1st-order explicit time integrator

type, extends(strategy) ,public :: explicit_euler
contains
procedure, nopass :: integrate
end type

14
15

contains

16
17
18
19
20
21
22
23

subroutine integrate(this,dt) ! Time integrator

class(surrogate) ,intent(inout) :: this ! integrand
real
,intent(in)
:: dt
! time step size
select type (this)
class is (integrand)
this = this + this%t()*dt ! Explicit Euler formula
class default

153

154

The Strategy and Surrogate Patterns

24
25
26
27

1
2
3
4

stop integrate: unsupported class.

end select
end subroutine
end module
Figure 7.2(h): runge_kutta_2nd.f03
module runge_kutta_2nd_module
use surrogate_module,only : surrogate
use strategy_module ,only : strategy
! time integration strategy
use integrand_module,only : integrand ! abstract integrand

5
6
7

implicit none
private

! Prevent implicit typing

! Hide everything by default

8
9
10
11
12
13

! 2nd-order Runge-Kutta time integration

type, extends(strategy) ,public :: runge_kutta_2nd
contains
procedure, nopass :: integrate
! integration procedure
end type

14
15

contains

16
17
18
19
20
21
22

! Time integrator
subroutine integrate(this,dt)
class(surrogate) ,intent(inout) :: this
! integrand
real
,intent(in)
:: dt
! time step size
class(integrand) ,allocatable
:: this_half ! function evaluation
! at interval t+dt/2.

23
24
25
26
27
28
29
30
31
32
33

select type (this)

class is (integrand)
allocate(this_half,source=this)
this_half = this + this%t()*(0.5*dt)
this
= this + this_half%t()*dt
class default
stop integrate: unsupported class
end select
end subroutine
end module

! predictor step
! corrector step

Figure 7.2. Fortran strategy pattern implementation for Lorenz and timed Lorenz systems:
(a) integration test of Lorenz and time-stamped Lorenz systems, (b) abstract surrogate,
(d) abstract strategy, (c) abstract integrand, (e) Lorenz ADT, (f) time-stamped Lorenz ADT,
(g) explicit Euler strategy, and (h) RK2 strategy.

One construct appearing for the rst time in this text is the non_overridable
attribute applied to the integrate procedure binding at line 11 in Figure 7.2(c). This
attribute ensures that types that extend integrand retain its implementation of that
binding. This ensures consistency in the sense that extended types will not be able to
overload the integrate name. Each will instead have access to the same collection
of integration methods implemented in a separate inheritance hierarchy.
Another construct making its rst appearance is the nopass attribute appearing
on line 12 in Figures 7.2(d), (g), and (h). It instructs the compiler to not pass as an

7.2 The Solution

argument the object on which the type-bound procedure is invoked. Instead, the
user must pass the argument (i.e., the integrand) explicitly. This simply forces the
procedure invocation syntax, say, integrate(x,dt), to mirror the corresponding
mathematical syntax xdt. Although one must prepend x% to the procedure invocation, this could be avoided by dening a module procedure that in turn invokes the
type-bound procedure.
The previously used intrinsic procedure allocate appears in a new form:
sourced allocation. This form appears rst at line 60 in Figure 7.2(c):
allocate (this%quadrature, source=s)
where the new source argument ensures that this%quadrature takes on the value
and dynamic type of s, which must be of a concrete strategy type. Another sourced
allocation occurs at line 26 in Figure 7.2(h):
allocate(this_half,source=this)
after which the immediately subsequent line overwrites this_half and thereby
negates the utility of copying the value of this. Fortran 2008 provides a new molded
allocation of the form:
allocate(this_half,mold=this)
which copies only the dynamic type of this, not its value.
Beginning with this chapter, the software problems being solved exhibit sufcient complexity that the code length exceeds the length of the related text. This
increases the expectation that the codes will be intelligible in a self-contained way.
We therefore increase the frequency with which we document our intent in comments. We also continue to attempt to write code that comments itself in the manner
described in Chapter 1. Also, the complexity of the examples in this chapter begin
to show the real power and benet of UML, as Figure 7.1 succinctly summarizes the
structural relationships between the different parts of the code in Figure 7.2.
Our strategies for dealing with code length evolve as the trend toward longer
codes continues unabated in subsequent chapters. In Chapter 8, we shift the code
listings to the ends of the relevant sections to preserve continuity in the text. In
Chapters 11 and 12, where we discuss multipackage, open-source projects with code
listings in the tens of thousands of lines, we shift to an even greater reliance on UML
supplemented by illustrative code snippets.

7.2.2 C++ Implementation

Figure 7.3 provides C++ STRATEGY code corresponding to the Fortran code listed in
Figure 7.2. As shown at line 7 in Figure 7.3(b), the strategy class forward references
the integrand class in its class denition, thus obviating the need for a SURROGATE
C++ class. In addition, because strategy extends class RefBase, it becomes eligible
as a parameter type for the reference counted pointer, Ref<>. Class integrand, as
shown at line 28 in Figure 7.3(c), uses a such reference counted strategy for its data
member quadrature_. This marks the rst use of a reference counted pointer in our
C++ examples that imitates the behaviors of a polymorphic allocatable component in

155

156

The Strategy and Surrogate Patterns

a derived type denition in Fortran. The corresponding Fortran declaration is found

at line 9 of Figure 7.2(c).

1
2
3
4
5

#include
#include
#include
#include
#include

Figure 7.3(a): main.cpp

"timed_lorenz.h"
"explicit_euler.h"
"runge_kutta_2nd.h"
"fmt.h"
<iostream>

6
7
8
9

int main ()
{
typedef lorenz::ptr_t ptr_t;

10

static const int num_steps=2000;

const real_t sigma=10., rho=28., beta=8./3., dt=0.01;
crd_t initial_condition(3, 1.0);

11
12
13
14

Ref<lorenz> attractor = new lorenz(initial_condition, sigma, rho, beta,

new explicit_euler);
std::cout << " lorenz attractor:\n"
<< fmt(attractor->coordinate(), 12, 9) << std::endl;
for (int step = 0; step < 4*num_steps; ++step)
{
attractor->integrate(0.25*dt);
std::cout << fmt(attractor->coordinate(), 12, 9) << std::endl;
}

15
16
17
18
19
20
21
22
23
24

Ref<timed_lorenz> timed_attractor
= new timed_lorenz(initial_condition, sigma, rho, beta,
new runge_kutta_2nd);
std::cout << "\n timed_lorenz attractor:\n"
<< fmt(timed_attractor->get_time(), 12, 9) << " "
<< fmt(timed_attractor->coordinate(), 12, 9) << std::endl;
for (int i = 0; i < num_steps; ++i)
{
timed_attractor->integrate(dt);
std::cout << fmt(timed_attractor->get_time(), 12, 9) << " "
<< fmt(timed_attractor->coordinate(), 12, 9) << std::endl;
}

25
26
27
28
29
30
31
32
33
34
35
36
37
38
39

1
2

return 0;

#ifndef _H_STRATEGY_
#define _H_STRATEGY_

Figure 7.3(b): strategy.h

3
4
5

#include "globals.h"
#include "RefBase.h"

6
7
8
9

class integrand;
class strategy : public RefBase {
public:

7.2 The Solution

10
11

typedef Ref<strategy> ptr_t;

typedef Ref<integrand> model_t;

12
13
14
15

virtual strategy() {}
virtual void integrate (model_t this_obj, real_t dt) const = 0;
};

16
17

1
2

#endif //!_H_STRATEGY_

#ifndef _H_INTEGRAND_
#define _H_INTEGRAND_

Figure 7.3(c): integrand.h

3
4
5
6

#include "strategy.h"
#include "RefBase.h"
#include "globals.h"

7
8
9
10
11
12

class integrand : virtual public RefBase

{
public:
typedef Ref<integrand> ptr_t;
typedef strategy::ptr_t strategy_t;

13
14
15
16

integrand(strategy_t);
integrand(const integrand&);
virtual integrand();

17
18
19
20

void set_strategy (strategy_t);

strategy_t get_strategy () const;
void integrate (real_t);

21
22
23
24
25

virtual
virtual
virtual
virtual

ptr_t
ptr_t
ptr_t
ptr_t

clone() const = 0;
d_dt() const = 0;
operator+=(ptr_t) = 0;
operator*=(real_t) = 0;

26
27
28
29

private:
strategy_t quadrature_;
};

30
31

#include "model_ops.h"

32
33

1
2

#endif //!_H_integrand_

#ifndef _H_MODEL_OPS_
#define _H_MODEL_OPS_

Figure 7.3(d): model_ops.h

3
4
5
6
7
8
9
10

inline integrand::ptr_t
operator+(integrand::ptr_t a, integrand::ptr_t b)
{
integrand::ptr_t tmp = a->clone();
*tmp += b;
return tmp;
}

157

158

The Strategy and Surrogate Patterns

11
12
13
14
15
16
17
18

inline integrand::ptr_t operator*(integrand::ptr_t a,

real_t b)
{
integrand::ptr_t tmp = a->clone();
*tmp *= b;
return tmp;
}

19
20

1
2

#endif //! _H_MODEL_OPS_

#include "integrand.h"
#include <exception>

Figure 7.3(e): integrand.cpp

3
4
5
6
7

struct integrand_error : public std::exception

{
virtual integrand_error() throw() {}
};

8
9
10

typedef integrand::ptr_t ptr_t;

typedef integrand::strategy_t strategy_t;

11
12
13
14

integrand::integrand(strategy_t quad) :
RefBase(), quadrature_(quad)
{}

15
16
17
18

integrand::integrand(const integrand& other) :

RefBase(), quadrature_(other.quadrature_)
{}

19
20
21

integrand::integrand()
{}

22
23
24
25
26

void integrand::set_strategy (strategy_t quad)

{
quadrature_ = quad;
}

27
28
29
30
31

strategy_t integrand::get_strategy () const

{
return quadrature_;
}

32
33
34
35
36

1
2

void integrand::integrate (real_t dt)

{
quadrature_->integrate(this, dt);
}

#ifndef _H_LORENZ_
#define _H_LORENZ_

3
4
5

#include "integrand.h"

Figure 7.3(f): lorenz.h

7.2 The Solution

6
7
8
9
10

class lorenz : public integrand

{
public:
typedef integrand::ptr_t ptr_t;
typedef integrand::strategy_t strategy_t;

11
12
13
14

lorenz(const crd_t&, real_t sigma, real_t rho, real_t beta,

strategy_t);
virtual lorenz();

15
16
17
18
19

virtual
virtual
virtual
virtual

ptr_t
ptr_t
ptr_t
ptr_t

clone() const;
d_dt() const;
operator+=(ptr_t);
operator*=(real_t);

20
21
22
23
24
25

void set_coordinate(const crd_t&);

const crd_t& coordinate() const;
real_t sigma() const;
real_t rho() const;
real_t beta() const;

26
27
28
29
30

private:
crd_t state_;
real_t sigma_, rho_, beta_;
};

31
32

1
2
3

#endif // !_H_LORENZ_

#include "lorenz.h"
#include "fmt.h"
#include <exception>

Figure 7.3(g): lorenz.cpp

4
5
6
7
8

struct lorenz_error : public std::exception

{
virtual lorenz_error() throw() {}
};

9
10

typedef lorenz::ptr_t ptr_t;

11
12

typedef lorenz::strategy_t strategy_t;

13
14
15
16
17

lorenz::lorenz(const crd_t& ste, real_t s, real_t r, real_t b,

strategy_t str) :
integrand(str), state_(ste), sigma_(s), rho_(r), beta_(b)
{}

18
19
20

lorenz::lorenz()
{}

21
22
23
24
25
26

ptr_t lorenz::clone() const

{
return ptr_t(new lorenz(*this));
}

159

160

The Strategy and Surrogate Patterns

27
28
29
30
31
32
33
34
35

ptr_t lorenz::d_dt() const

{
crd_t new_state(3);
new_state.at(0) = sigma_ * (state_.at(1) - state_.at(0));
new_state.at(1) = state_.at(0) * (rho_ - state_.at(2)) - state_.at(1);
new_state.at(2) = state_.at(0) * state_.at(1) - beta_ * state_.at(2);
return ptr_t(new lorenz(new_state, sigma_, rho_, beta_,
get_strategy()));
}

36
37
38
39
40
41
42
43
44
45
46
47
48
49

ptr_t lorenz::operator+=(ptr_t inval)

{
Ref<lorenz> other = cast<lorenz>(inval);
if((other == NULL) || (state_.size() != other->state_.size())) {
std::cerr << "lorenz::operator+=: Invalid input argument\n";
throw lorenz_error();
}
size_t size = state_.size();
for(size_t i = 0; i < size; ++i) {
state_.at(i) += other->state_.at(i);
}
return ptr_t(this);
}

50
51
52
53
54
55
56
57
58

ptr_t lorenz::operator*=(real_t val)

{
size_t size = state_.size();
for(size_t i = 0; i < size; ++i) {
state_.at(i) *= val;
}
return ptr_t(this);
}

59
60
61
62
63

void lorenz::set_coordinate(const crd_t& state)

{
state_ = state;
}

64
65
66
67
68

const crd_t& lorenz::coordinate() const

{
return state_;
}

69
70
71
72
73

real_t lorenz::sigma() const

{
return sigma_;
}

74
75
76
77
78

real_t lorenz::rho() const

{
return rho_;
}

79
80
81

real_t lorenz::beta() const

{

7.2 The Solution

82
83

1
2

return beta_;

Figure 7.3(h): timed_lorenz.h

#ifndef _H_TIMED_LORENZ_
#define _H_TIMED_LORENZ_

3
4
5

#include "strategy.h"
#include "lorenz.h"

6
7
8
9
10
11

class timed_lorenz : public lorenz {

public:
typedef lorenz::ptr_t ptr_t;
typedef lorenz::strategy_t strategy_t;

12
13
14
15
16
17
18
19

timed_lorenz(const crd_t&, real_t sigma, real_t rho, real_t beta,

strategy_t, double t_init = 0);
virtual timed_lorenz();
virtual ptr_t clone() const;
virtual ptr_t d_dt() const;
virtual ptr_t operator+=(ptr_t);
virtual ptr_t operator*=(real_t);

20
21
22
23
24
25

void
set_time (real_t);
real_t get_time() const;
private:
real_t time_;
};

26
27

1
2

#endif //!_H_TIMED_LORENZ_
Figure 7.3(i): timed_lorenz.cpp
#include "timed_lorenz.h"
#include <exception>

3
4
5
6
7

struct timed_lorenz_error : public std::exception

{
virtual timed_lorenz_error() throw() {}
};

8
9
10

typedef timed_lorenz::ptr_t ptr_t;

typedef timed_lorenz::strategy_t strategy_t;

11
12
13
14
15

timed_lorenz::timed_lorenz(const crd_t& ste, real_t s, real_t r,

real_t b, strategy_t strat, double t_init) :
lorenz(ste, s, r, b, strat), time_(t_init)
{}

16
17
18

timed_lorenz::timed_lorenz()
{}

19
20
21
22

ptr_t timed_lorenz::clone() const

{
return ptr_t(new timed_lorenz(*this));

161

162

The Strategy and Surrogate Patterns

23

24
25
26
27
28
29
30
31

ptr_t timed_lorenz::d_dt() const

{
Ref<lorenz> parent = cast<lorenz>(lorenz::d_dt());
return ptr_t(new timed_lorenz(parent->coordinate(), parent->sigma(),
parent->rho(), parent->beta(),
parent->get_strategy(), 1.0));
}

32
33
34
35
36
37
38
39
40
41
42
43

ptr_t timed_lorenz::operator+=(ptr_t inval)

{
Ref<timed_lorenz> other = cast<timed_lorenz>(inval);
if(other == NULL) {
std::cerr << "timed_lorenz::operator+=: Invalid input type\n";
throw timed_lorenz_error();
}
lorenz::operator+=(other);
time_ += other->time_;
return ptr_t(this);
}

44
45
46
47
48
49
50

ptr_t timed_lorenz::operator*=(real_t val)

{
lorenz::operator*=(val);
time_ *= val;
return ptr_t(this);
}

51
52
53
54
55

void timed_lorenz::set_time (real_t t)

{
time_ = t;
}

56
57
58
59
60

1
2

real_t timed_lorenz::get_time() const

{
return time_;
}
Figure 7.3(j): explicit_euler.h
#ifndef _H_EXPLICIT_EULER_
#define _H_EXPLICIT_EULER_

3
4
5

#include "strategy.h"
#include "integrand.h"

6
7
8
9
10
11
12

class explicit_euler : public strategy

{
public:
virtual explicit_euler();
virtual void integrate (model_t this_obj, real_t dt) const;
};

13
14

#endif //!_H_EXPLICIT_EULER_

7.2 The Solution

1
2
3

Figure 7.3(k): explicit_euler.cpp

#include "explicit_euler.h"
#include "integrand.h"
#include <exception>

4
5
6

explicit_euler::explicit_euler()
{}

7
8
9
10

1
2

void explicit_euler::integrate (model_t this_obj, real_t dt) const {

*this_obj += this_obj->d_dt() * dt;
}
Figure 7.3(l): runge_kutta_2nd.h
#ifndef _H_RUNGE_KUTTA_2ND_
#define _H_RUNGE_KUTTA_2ND_

3
4
5

#include "strategy.h"
#include "integrand.h"

6
7
8
9
10
11

class runge_kutta_2nd : public strategy {

public:
virtual runge_kutta_2nd();
virtual void integrate(model_t this_obj, real_t dt) const;
};

12
13

1
2

#endif

#include <iostream>
#include <exception>

Figure 7.3(m): runge_kutta_2nd.cpp

3
4
5

#include "integrand.h"
#include "runge_kutta_2nd.h"

6
7

using namespace std;

8
9
10

runge_kutta_2nd::runge_kutta_2nd()
{}

11
12
13
14
15
16

void runge_kutta_2nd::integrate (model_t this_obj, real_t dt) const

{
model_t this_half = this_obj + this_obj->d_dt() * (0.5*dt);//predictor
*this_obj += this_half->d_dt() * dt; // corrector
}
Figure 7.3. C++ strategy implementation for Lorenz and timed Lorenz systems: (a) Lorenz
and time-stamped Lorenz integration tests, (b) abstract strategy class denition using forward
references to an abstract integrand class, (c) operator denitions (+ and *) for integrands, (d)
abstract integrand denition, (e) abstract integrand implementation, (f) Lorenz denition, (g)
Lorenz implementation, (h) time-stamped Lorenz denition, (i) time-stamped Lorenz implementation, (j) explicit Euler strategy denition, (k) explicit Euler strategy implementation,
(l) RK2 strategy denition, and (m) RK2 strategy integration.

163

164

The Strategy and Surrogate Patterns

7.3 The Consequences

Because STRATEGY uses aggregation instead of inheritance, it allows better decoupling of the classes that focus on data (the context) from those that focus on algorithms
(behaviors). One obvious advantage is that strategies can provide varying implementations (or algorithms) for the same behavior, thus giving users more freedom to
choose at runtime based on the problem at hand. Compared to patterns that employ
inheritance as the means for maintaining algorithms (e.g., the INTERPRETER pattern of
the GoF), a STRATEGY is much easier to understand and extend. Since each concrete
strategy class implements one particular algorithm for the context, the use of this
pattern also encourages programmers to avoid lumping many different algorithms
into one class that often lead to unmanageable code.
This pattern is commonly used in applications where a family of related algorithms or behaviors exists for the context (data). Examples given by the GoF
include the register allocation schemes and instruction scheduling policies used in
the compiler optimization code in the Register Transfer Language (RTL) systems.
The application of strategy patterns yielded great exibility for the optimizer in
targeting different machine architectures.
As with every design pattern, STRATEGY pattern also has drawbacks. One potential disadvantage is that in applying the pattern, a user must be aware of the
differences among algorithms in order to select the appropriate one. This sometimes
becomes a burden as programmers must acquire knowledge on various algorithms.
Another potential shortcoming of strategies lies in the possible communication
overhead between data objects and algorithm objects. Although this overhead can
normally be minimized by careful design of the strategy interfaces, a nave implementer of the strategy pattern may nevertheless attempt to design the interfaces with
many unnecessary parameters to pass between data and algorithms. Thus achieving
a balance between data-algorithm coupling and communication overhead should
always be one of the goals for designing a good STRATEGY.

7.4 Related Patterns

In the context of writing differential equation solvers, the STRATEGY pattern relates
closely to ABSTRACT CALCULUS. From differentiation and integration to addition and
subtraction, virtually every operation performed in an ABSTRACT CALCULUS serves as
an umbrella for a collection of strategies. In numerical analysis, approximations to
derivatives and integrals come in many varieties with the varying orders of accuracy,
stencil widths, phase errors, stability criterion, and symmetries. Any situation in
which the ability to vary these properties provides fertile ground for a STRATEGY to
take root. In this sense, a STRATEGY helps complete an ABSTRACT CALCULUS.
In Fortran, SURROGATE and STRATEGY are inextricably bound. The former completes the latter, the relationship being sufciently close that one can consider a
SURROGATE to be part of a STRATEGY. The uses for a SURROGATE, however, extend
to all situations in which the class relationships in a design suggest the need for
circular references. As evidenced by the frequency with which the question of how
to facilitate such references appears in the comp.lang.fortran Usenet newsgroup.1
1 http://groups.google.com/group/comp.lang.fortran

7.4 Related Patterns

In a multiphysics simulation context, one needs the next chapters PUPPETEER

pattern to complete a STRATEGY. The PUPPETEER hides the single-physics submodels
by aggregating their ADTs. It provides a single ADT on which the STRATEGY can
operate without knowledge of the individual sub-models.

EXERCISES
1. Runge-Kutta-Fehlberg (RKF) algorithms provide one context for dynamically
varying a time integration method. By alternating between two schemes with
different orders of accuracy, RKF methods facilitate time step size control based
on error estimation. For example, a 4th-order-accurate Runge-Kutta-Fehlberg
method (RKF45) uses a 5th-order-accurate step for error estimation. In advancing
a differential equation dy/dt = f (t, y) from a time ti over a time step hi to a time
ti+1 ti + hi , RKF45 constructs both approximations from one set of function
evaluations (Morris 2008):
K1 = hf (ti , yi )


1
1
K2 = hf ti + h, yi + K1
4
4


3
3
9
K3 = hf ti + h, yi + K1 + K2
8
32
32


12
1932
7200
7296
K4 = hf ti + h, yi +
K1
K2 +
K3
13
2197
2197
2197


439
3680
845
K5 = hf ti + h, yi +
K1 8K2 +
K3
K4
216
513
4104


1
8
3544
1849
11
K6 = hf ti + h, yi K1 + 2K2
K3
K4 K 5
2
27
2565
4104
40
which yield the 4th - and 5th -order-accurate schemes
25
1408
2197
1
K1 +
K3 +
K4 K5
216
2565
4101
5
16
6656
28561
9
2
yRK5
K1 +
K3 +
K4 K5 + K5
i+1 = yi +
135
12825
56430
50
55

yRK4
i+1 = yi +

RK5
Estimating the error as e yRK4
i+1 yi+1 , one adjusts the time step size just before
each step according to:


hi =

 0.20

Shi1  etol
 e e 0.25
tol
Shi1  e 

if e etol
if e > etol

where S is a safety factor just below unity and etol is the error tolerance.
Design a STRATEGY for solving the Lorenz system (4.5) using the RKF45 algorithm.
Express your design in a UML class diagram. Given S = and etol =, write an
implementation of your design.

165

166

The Strategy and Surrogate Patterns

2. In many situations, nite difference approximations perform best when the stencil
employed to approximate derivatives at a given point is skewed toward the direction from which information is propagating. Labeling a given direction in one
dimension forward and the opposite direction backward, it is desirable to be
able to switch between the forward and backward differencing schemes described
in Section A.5.1. Design a STRATEGY for switching between these two schemes at
runtime. Express your design in a UML class diagram. Write an implementation
of your design.

The Puppeteer Pattern

Never be afraid to try something new. Remember, amateurs built the ark.
Professionals built the Titanic.
Miss Piggy

8.1 The Problem

While the ABSTRACT CALCULUS and STRATEGY patterns apply to the integration of
a single physics abstraction, our chief concern lies in linking multiple abstractions.
This poses at least two signicant software design problems. The rst involves how
to facilitate interabstraction communication. The GoF addressed interabstraction
communication with the MEDIATOR pattern. When N objects interact, a software architect can reduce the N(N 1)+ associations between the objects to 2N associations
by employing a MEDIATOR.
The MEDIATOR association count stems from the requirements that the Mediator know each communicating party and those parties know the Mediator. For
example, in a MEDIATOR implementation presented by Gamma et al. (1995), the
sender passes a reference to itself to the MEDIATOR. The sender must be aware of
the MEDIATOR in order to know where to send the message. Likewise, the MEDIATOR
must be aware of the sender in order to invoke methods on the sender via the passed
reference. Figure 8.1 illustrates the associations in an atmospheric boundary layer
model, wherein the air, ground, and cloud ADTs might solve equation sets for the
airow, ground transpiration, and discrete droplet motion, respectively.
A second and conceptually more challenging problem concerns how one assembles quantities of an inherently global nature that is, information that can only
be determined with simultaneous knowledge of the implementation details of each
of the single-physics abstractions. This cuts to the heart of the question of whether
ABSTRACT CALCULUS can be considered a general software design philosophy.
We illustrate this second dilemma with an important and common mathematical
problem: implicit time advancement of nonlinear equation systems. Of particular
concern is the desire to calculate cross-coupling terms without violating abstractions by exposing their data. Consider marching the ordinary differential equation
167

168

The Puppeteer Pattern

Air

Cloud

Mediator

Ground

Figure 8.1. Associations in the Mediator pattern.

system (4.4) forward in time according to the trapezoidal rule:



t    n     n+1 
n+1
n


=V +
V
R V +R V
2

(8.1)

 contains
 n+1 on the RHS makes iteration necessary when R
where the presence of V
nonlinearities. One generally poses the problem in terms of nding the roots of a
residual vector f formed by gathering the terms in equation (8.1):

  


 V
 n+1
 n+1 V
 n + t R
 n+1 ) V
 V
n +R
(8.2)
f(V
2
 n is known from a previous time step.
where V
The difculty arises in nding the roots of f using Jacobian-based iteration methods. Consider Newtons method. Dening ym as the mth iterative approximation to
 n+1 , Newtons method can be expressed as:
V
Jym f(ym )

(8.3)

ym+1 ym + ym

Jij

 n+1
V

fi
fi
t Ri
=
= ij
n+1
yj y=ym Vj
2
 n+1 =ym
Vj n+1
V

(8.4)


 n+1 =ym
V

(8.5)

where J is the Jacobian, Ri is the RHS of the ith governing equation, and ij is the
Kronecker delta. Equation (8.3) represents a linear algebraic system. Equation (8.4)
represents vector addition.
 are partitioned such that different elements are
Now imagine that f, y, and R

hidden behind interfaces to different abstractions. In particular, let us partition R

T
 a c g , where the partitions separate the elements of R
 corresuch that R
sponding to the air, cloud, and ground ADTs, respectively. For present purposes,
each partition can be thought of as a 1D vector containing the RHS of one component equation from the Lorenz system (4.5). With this notation, one can rewrite the
Jacobian in equation (8.3) as:1
J I

t (a, c, g )
2 (
, ,
  )

(8.6)

1 In equation (8.4) and elsewhere, we abbreviate a common notation for Jacobians in which the list of

f components appears in the numerator and the list of y components appears in the denominator.
We do not mean for the arrows to connote vectors that transform as rst-order tensors.

8.2 The Solution

where  ,  , and  are the air, cloud, and ground state vectors, respectively, so

T
y    , and where I is the identity matrix.
The second dilemma presents itself in the need to calculate cross terms such
as a/  . The question is, How does the air ADT differentiate its components
 with respect to the state vector partition  when that partition lies hidden
of R
inside the cloud abstraction? Even if one solves this puzzle, a more perplexing one
remains: Where does J live? Since the precise form of equations (8.5)(8.6) varies
with the choice of time integration algorithms, the natural place to construct J is
inside the time integration procedure. In an ABSTRACT CALCULUS implementation,
however, that procedure maintains a blissful ignorance of the governing equations
inside the dynamical system it integrates as exemplied by the integrate procedures
in Chapter 6.
It appears the conicting forces driving the software problems in this book climax
in the current chapters context, potentially forcing the violation of abstractions.
The desire to maintain a modular architecture with local, private data confronts
the need to handle intermodule dependencies. These dependencies force rigidly
specied communication and complicate the calculation of inherently global entities.
A Showdown at the OOP Corral
The ABSTRACT CALCULUS edice threatens to collapse under the weight of the
problems posed here. Modularity and data privacy present signicant impediments to coupling independently developed abstractions without exposing their
implementation details. Of prime concern are how to facilitate interabstraction
communication and how to assemble inherently global quantities.

8.2 The Solution

A PUPPETEER encapsulates references to each time-integrable system involved in
the simulation. When one of the integrate procedures of Chapter 6, for example,
receives a PUPPETEER argument, that PUPPETEER controls the behavior of all other
integrable systems in the simulation. It does so by delegating operations and serving
as an intermediary for communications.
The PUPPETEER designer also ensures consistency in the communication format.
She might do so by reducing, or requiring the puppets to reduce, each datum communicated to a type intrinsic to the language. In doing so, the PUPPETEER would obviate
the need to expose information about the data structures each ADT employs. When
an application calls for more nuanced data structures than the language supports
for example, compressed storage of sparse matrices the PUPPETEER designer might
instead enforce that all puppets encapsulate message information in a data structure
tailored to the given application domain. In distributed computing applications, for
example, the Trilinos Epetra objects discussed in Chapters 5 and 11 is one likely
candidate for encapsulating data that all ADTs would be required to accept.
A PUPPETEER exploits the regularity and predictability of interabstraction communications dened formally in the terms that couple the governing equations. The
developer of each ADT requests the requisite coupling terms by placing them in the
argument lists of the procedures that dene each dynamical system. The PUPPETEER

169

170

The Puppeteer Pattern

Puppeteer

Ground

Air

Cloud

Figure 8.2. Aggregation associations in the Puppeteer pattern.

developer fullls these requests by calling accessors made public by the other dynamical systems in the simulation. As depicted in Figure 8.2, the PUPPETEER uses object
aggregation to reduce the aforementioned number of interabstraction associations
from 2N = 6 to N = 3: the PUPPETEER knows a datums sender and recipient, but they
need not know the PUPPETEER.
The dialogue the PUPPETEER mediates in the atmospheric boundary layer model
can be paraphrased as follows:
Puppeteer to Air: Please pass me a vector containing the partial derivatives of each
of your governing equations RHS with respect to each element in your state
vector.
Air to Puppeteer: Here is the requested vector ( a/  ). You can tell the dimension
of my state by the size of this vector.
Puppeteer to Air: I also know that your governing equations depend on the Cloud
state vector because your interface requests information that I retrieved from a
Cloud. Please pass me a vector analogous to the previous one but differentiated
with respect to the Cloud state information I passed to you.
Puppeteer note to self : Since I did not pass any information to the Air object from
the Ground object, I will set the cross-terms corresponding to a/ g to zero.
Ill determine the number of zero elements by multiplying the size of a/  by
the size of g /  after I receive the latter from my Ground puppet.
The PUPPETEER then holds analogous dialogues with its Cloud and Ground pup V
 to integrate. The
pets, after which the PUPPETEER passes an array containing R/
latter procedure uses the passed array to form J, which it then passes to a PUPPETEER
for use in inverting the matrix system (equation 8.3). Most importantly, integrate
does so without violating the PUPPETEERs data privacy, and the PUPPETEER responds
without violating the privacy of its puppets. The PUPPETEERs construction is based
solely on information from the public interfaces of each puppet. Figure 8.3 details
 V
 in a UML sequence diagram. Note the coordinate()
the construction of R/
method calls return solution variables. The next three calls return diagonal blocks
 V.
 The subsequent six calls return off-diagonal blocks. The nal two steps
of R/
 V.

allocate and ll R/
8.2.1 Fortran Implementation
Figure 8.4 provides the denition of the atmosphere type, which acts as a PUPPETEER in our prototype atmospheric boundary-layer simulation involving the Lorenz

8.2 The Solution

this:atmosphere

air_puppet:air

cloud_puppet:cloud

ground_puppet:ground

coordinate()
air's state vector
coordinate()
cloud's state vector
coordinate()
ground's state vector
d_dAir(cloud_coord)
dAir_dAir
d_dCloud(air_coord, ground_coord)
dCloud_dCloud
d_dGround(air_coord, cloud_coord)
dGround_dGround
d_dy(cloud_coord)
dAir_dCloud
dAir_dGround
d_dx (air_coord, ground_coord)
dCloud dAir
d_dz (air_coord, gound_coord)
d_Cloud_dGround
d_dx (air_coord, cloud_coord)
dGround_dAir
d_dy (air_coord, cloud_coord)
dGround_dCloud
allocate

dRHS_dState
:real

assembly result

Figure 8.3. A UML sequence diagram for the Puppeteers Jacobian contribution.

equations partitioned in the manner described in the previous section. Here the
atmosphere contains references to all its puppets, which are of type air, cloud, and
ground. The puppets data can be constructed independently after the atmosphere
object is created.
Lines 1618 in Figure 8.4 show the composition of the atmosphere type. The
atmosphere object relays all information using Fortran allocatable scalar data.

171

172

The Puppeteer Pattern

Although the puppets are instantiated before the puppeteer, they are destroyed
when the puppeteer is destroyed, for example when its name goes out of scope. This
happens automatically at the hands for the compiler.
The function dRHS_dV (from line 72 to line 153) illustrates the construction of a
 V,
 corresponding to the UML sequence diagram given
Jacobian contribution, R/
in Figure 8.3. Figure 8.5 details the denitions for all the puppets in the atmosphere
system. Puppet air, cloud, and ground track the rst, second, and third state variables
in the Lorenz system, respectively.

1
2
3
4
5
6
7
8

Figure 8.4: atmosphere.f03

module atmosphere_module
use air_module ,only : air ! puppet for 1st Lorenz eq. and
! corresonding state variable
use cloud_module ,only : cloud ! puppet for 2nd Lorenz eq. and
! corresonding state variable
use ground_module ,only : ground ! puppet for 3rd Lorenz eq. and
! corresonding state variable
use global_parameters_module ,only:debugging !print call tree if true

9
10
11
12

!implicit none
! Prevent implicit typing
private
! Hide everything by default
public :: atmosphere

13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33

type atmosphere ! Puppeteer

private
type(air)
,allocatable :: air_puppet
type(cloud) ,allocatable :: cloud_puppet
type(ground) ,allocatable :: ground_puppet
contains
procedure :: t ! time derivative
procedure :: dRHS_dV ! Jacobian contribution (dR/dV)
procedure :: state_vector ! return atmosphere solution vector
procedure :: add ! add two atmospheres
procedure :: subtract ! subtract one atmosphere from another
procedure :: multiply ! multiply an atmosphere by a real scalar
procedure ,pass(rhs) :: inverseTimes !abstract Gaussian elimination
procedure :: assign ! assign one atmosphere to another
generic :: operator(+) => add
generic :: operator(-) => subtract
generic :: operator(*) => multiply
generic :: assignment(=) => assign
generic :: operator(.inverseTimes.) => inverseTimes
end type atmosphere

34
35
36
37

interface atmosphere
procedure constructor
end interface

38
39

contains

40
41
42
43
44

! constructor for an atmosphere object

type(atmosphere) function constructor &
(air_target,cloud_target,ground_target)
type(air)
,allocatable ,intent(inout) :: air_target

8.2 The Solution

45
46
47
48
49
50
51
52
53

type(cloud) ,allocatable ,intent(inout) :: cloud_target

type(ground) ,allocatable ,intent(inout) :: ground_target
if (debugging) print *,
atmosphere%construct(): start
! transfer allocations from puppets to Puppeteer
call move_alloc(air_target, constructor%air_puppet)
call move_alloc(ground_target, constructor%ground_puppet)
call move_alloc(cloud_target, constructor%cloud_puppet)
if (debugging) print *,
atmosphere%construct(): end
end function

54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69

! time derivative (evolution equations)

function t(this) result(dState_dt)
class(atmosphere) ,intent(in) :: this
class(atmosphere) ,allocatable :: dState_dt
type(atmosphere) ,allocatable :: delta
if (debugging) print *,
atmosphere%t(): start
allocate(delta)
delta%air_puppet= this%air_puppet%t(this%cloud_puppet%coordinate())
delta%cloud_puppet = this%cloud_puppet%t( &
this%air_puppet%coordinate(),this%ground_puppet%coordinate())
delta%ground_puppet = this%ground_puppet%t( &
this%air_puppet%coordinate(),this%cloud_puppet%coordinate())
call move_alloc(delta, dState_dt)
if (debugging) print *,
atmosphere%t(): end
end function

70
71
72
73

! atmosphere contribution to Jacobian

function dRHS_dV(this) result(dRHS_dState)
class(atmosphere)
,intent(in) :: this

74
75
76
77
78
79

! Sub-blocks of dR/dV array

real ,dimension(:,:) ,allocatable ::
&
dAir_dAir ,dAir_dCloud ,dAir_dGround,
&
dCloud_dAir ,dCloud_dCloud ,dCloud_dGround,&
dGround_dAir ,dGround_dCloud ,dGround_dGround

80
81

real ,dimension(:,:) ,allocatable :: dRHS_dState

82
83
84

real ,dimension(:)
,allocatable ::
&
air_coordinate, cloud_coordinate, ground_coordinate

85
86
87

integer :: air_eqs ,air_vars,cloud_eqs ,cloud_vars , &

ground_eqs ,ground_vars ,i ,j ,rows ,cols

88
89

if (debugging) print *,atmosphere%dRHS_dV(): start

90
91
92
93

! Calculate matrices holding partial derivative of puppet

! evolution equation right-hand sides with respect to the
! dependent variables of each puppet.

94
95
96
97

air_coordinate = this%air_puppet%coordinate()
cloud_coordinate= this%cloud_puppet%coordinate()
ground_coordinate=this%ground_puppet%coordinate()

98
99

! compute diagonal block submatrix

173

174

The Puppeteer Pattern

100
101
102
103
104
105
106
107
108
109
110

dAir_dAir
dCloud_dCloud

= this%air_puppet%d_dAir(cloud_coordinate)
= this%cloud_puppet%d_dCloud(air_coordinate,&
ground_coordinate)
dGround_dGround = this%ground_puppet%d_dGround(air_coordinate,&
cloud_coordinate)
air_eqs
= size(dAir_dAir,1)
! submatrix rows
air_vars
= size(dAir_dAir,2)
! submatrix columns
cloud_eqs
= size(dCloud_dCloud,1)
! submatrix rows
cloud_vars = size(dCloud_dCloud,2)
! submatrix columns
ground_eqs = size(dGround_dGround,1) ! submatrix rows
ground_vars = size(dGround_dGround,2) ! submatrix columns

111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126

! compute off-diagonal values

dAir_dCloud
= this%air_puppet%d_dy(cloud_coordinate)
dAir_dGround = reshape(source=(/(0.,i=1,air_eqs*ground_vars)/)&
,shape=(/air_eqs,ground_vars/))
dCloud_dAir
= this%cloud_puppet%d_dx(air_coordinate,
&
ground_coordinate)
dCloud_dGround = this%cloud_puppet%d_dz(air_coordinate,
&
ground_coordinate)
dGround_dAir
= this%ground_puppet%d_dx(air_coordinate,
&
cloud_coordinate)
dGround_dCloud = this%ground_puppet%d_dy(air_coordinate,
&
cloud_coordinate)
rows=air_eqs+cloud_eqs+ground_eqs
cols=air_vars+cloud_vars+ground_vars
allocate(dRHS_dState(rows,cols))

127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154

! Begin result assembly

dRHS_dState(1:air_eqs, 1:air_vars)
dAir_dAir
dRHS_dState(1:air_eqs, air_vars+1:air_vars+cloud_vars)
dAir_dCloud
dRHS_dState(1:air_eqs, air_vars+cloud_vars+1:cols)
dAir_dGround
dRHS_dState(air_eqs+1:air_eqs+cloud_eqs, 1:air_vars)
dCloud_dAir
dRHS_dState(air_eqs+1:air_eqs+cloud_eqs,
air_vars+1:air_vars+cloud_vars)
dCloud_dCloud
dRHS_dState(air_eqs+1:air_eqs+cloud_eqs,
air_vars+cloud_vars+1:cols)
dCloud_dGround
dRHS_dState(air_eqs+cloud_eqs+1:rows, 1:air_vars)
dGround_dAir
dRHS_dState(air_eqs+cloud_eqs+1:rows,
air_vars+1:air_vars+cloud_vars)
dGround_dCloud
dRHS_dState(air_eqs+cloud_eqs+1:rows,
air_vars+cloud_vars+1:cols)
dGround_dGround
! result assembly finishes
if (debugging) print *,
atmosphere%dRHS_dV(): end
end function dRHS_dV

&

&

&

&

&
&

&
&

&

&
&

&
&

8.2 The Solution

155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175

! assemble and return solution vector

function state_vector(this) result(phase_space)
class(atmosphere) ,intent(in) :: this
real ,dimension(:) ,allocatable :: state
real ,dimension(:) ,allocatable :: x,y,z,phase_space
integer
:: x_start,y_start,z_start
integer
:: x_end ,y_end ,z_end
x = this%air_puppet%coordinate()
x_start=1
x_end=x_start+size(x)-1
y = this%cloud_puppet%coordinate()
y_start=x_end+1
y_end=y_start+size(y)-1
z = this%ground_puppet%coordinate()
z_start=y_end+1
z_end=z_start+size(z)-1
allocate(phase_space(size(x)+size(y)+size(z)))
phase_space(x_start:x_end) = x
phase_space(y_start:y_end) = y
phase_space(z_start:z_end) = z
end function

176
177
178
179
180
181
182
183
184
185
186
187

function add(lhs,rhs) result(sum)

class(atmosphere) ,intent(in) :: lhs
class(atmosphere) ,intent(in) :: rhs
class(atmosphere) ,allocatable :: sum
if (debugging) print *,
atmosphere%add(): start
allocate(sum)
sum%air_puppet
= lhs%air_puppet
+ rhs%air_puppet
sum%cloud_puppet = lhs%cloud_puppet + rhs%cloud_puppet
sum%ground_puppet = lhs%ground_puppet + rhs%ground_puppet
if (debugging) print *,
atmosphere%add(): end
end function

188
189
190
191
192
193
194
195
196
197
198
199

function subtract(lhs,rhs) result(difference)

class(atmosphere)
,intent(in) :: lhs
class(atmosphere) ,intent(in) :: rhs
class(atmosphere) ,allocatable :: difference
if (debugging) print *, atmosphere%subtract(): start
allocate(difference)
difference%air_puppet = lhs%air_puppet - rhs%air_puppet
difference%cloud_puppet = lhs%cloud_puppet - rhs%cloud_puppet
difference%ground_puppet = lhs%ground_puppet - rhs%ground_puppet
if (debugging) print *,
atmosphere%subtract(): end
end function

200
201
202
203
204
205
206
207
208
209

! Solve linear system Ax=b by Gaussian elimination

function inverseTimes(lhs,rhs) result(product)
class(atmosphere) ,intent(in) :: rhs
class(atmosphere) ,allocatable :: product
real ,dimension(:,:) ,allocatable ,intent(in) :: lhs
real ,dimension(:)
,allocatable :: x,b
real ,dimension(:,:) ,allocatable :: A
real ,parameter :: pivot_tolerance=1.0E-02
integer :: row,col,n,p ! p=pivot row/col

175

176

The Puppeteer Pattern

210

real :: factor

211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237

n=size(lhs,1)
b = rhs%state_vector()
if (n /= size(lhs,2) .or. n /= size(b)) &
stop atmosphere.f03: ill-posed matrix problem in inverseTimes()
allocate(x(n))
A = lhs
do p=1,n-1
! Forward elimination
if (abs(A(p,p))<pivot_tolerance) &
stop invert: use an algorithm with pivoting
do row=p+1,n
factor=A(row,p)/A(p,p)
forall(col=p:n)
A(row,col) = A(row,col) - A(p,col)*factor
end forall
b(row) = b(row) - b(p)*factor
end do
end do
x(n) = b(n)/A(n,n) ! Back substitution
do row=n-1,1,-1
x(row) = (b(row) - sum(A(row,row+1:n)*x(row+1:n)))/A(row,row)
end do
allocate(product,source=rhs)
product%air_puppet = air(x(1),x(2))
call product%cloud_puppet%set_coordinate(x(3))
call product%ground_puppet%set_coordinate(x(4))
end function

238
239
240
241
242
243
244
245
246
247
248
249
250

! multiply atmosphere object by a real scalar

function multiply(lhs,rhs) result(product)
class(atmosphere) ,intent(in) :: lhs
real
,intent(in) :: rhs
class(atmosphere) ,allocatable :: product
if (debugging) print *,
atmosphere%multiply(): start
allocate(product)
product%air_puppet
= lhs%air_puppet
* rhs
product%cloud_puppet = lhs%cloud_puppet * rhs
product%ground_puppet = lhs%ground_puppet * rhs
if (debugging) print *,
atmosphere%multiply(): end
end function

251
252
253
254
255
256
257
258
259
260
261
262

! assign one atmosphere object to another

subroutine assign(lhs,rhs)
class(atmosphere) ,intent(inout) :: lhs
class(atmosphere) ,intent(in)
:: rhs
if (debugging) print *,
atmosphere%assign(): start
lhs%air_puppet
= rhs%air_puppet
! automatic allocation
lhs%cloud_puppet = rhs%cloud_puppet ! automatic allocation
lhs%ground_puppet = rhs%ground_puppet ! automatic allocation
if (debugging) print *,
atmosphere%assign(): end
end subroutine
end module atmosphere_module
Figure 8.4. Fortran implementation of an atmosphere puppeteer.

8.2 The Solution

1
2
3
4
5

Figure 8.5(a): air.f03

module air_module
use global_parameters_module ,only : debugging !print call treeif true
implicit none ! Prevent implicit typing
private
! Hide everything by default
public :: air

6
7
8
9

! Number of evolution equations/variables exposed to the outside world

! (via Jacobian sub-block shapes):
integer ,parameter :: num_eqs=2,num_vars=2

10
11
12
13
14
15
16
17
18
19

!
!
!
!
!
!
!
!
!

This type tracks the evolution of the first state variable in the
Lorenz system according to the first Lorenz equation. It also
tracks the corresponding paramater (sigma) according to the
differential equation d(sigma)/dt=0. For illustrative purposes, this
implementation exposes the number of state variables (2) to the
puppeteer without providing direct access to them or exposing
anything about their layout, storage location or identifiers (x and
sigma). Their existence is apparent in the rank (2) of the matrix
d_dAir() returns as its diagonal Jacobian element contribution.

20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36

type air
private
real :: x,sigma !1st Lorenz equation solution variable and parameter
contains
procedure :: coordinate! accessor: return phase-space coordinates
procedure :: t ! time derivative ( evaluate RHS of Lorenz equation)
procedure :: d_dAir ! contribution to diagonal Jacobian element
procedure :: d_dy
! contribution to off-diagonal Jacobian element
procedure ,private :: add ! add two instances
procedure ,private :: subtract ! subtract one instance from another
procedure ,private :: multiply ! multiply instance by a real scalar
! map defined operators to corresponding procedures
generic
:: operator(+) => add
generic
:: operator(-) => subtract
generic
:: operator(*) => multiply
end type

37
38
39
40

interface air
procedure constructor
end interface

41
42
43
44
45
46
47
48
49
50
51

contains
! constructor: allocate and initialize components
type(air) function constructor(x_initial,s)
real
,intent(in) :: x_initial
real
,intent(in) :: s
if (debugging) print *,
air%construct: start
constructor%x = x_initial
constructor%sigma = s
if (debugging) print *,
air%construct: end
end function

52
53
54

! accessor: returns phase-space coordinates

function coordinate(this) result(return_x)

177

178

The Puppeteer Pattern

55
56
57
58

class(air)
,intent(in) :: this
real ,dimension(:) ,allocatable :: return_x
return_x = [ this%x ,this%sigma ]
end function

59
60
61
62
63
64
65
66
67
68
69

! time derivative (evolution equations)

function t(this,y) result(dx_dt)
class(air)
,intent(in) :: this
real ,dimension(:) ,intent(in) :: y
type(air)
:: dx_dt
if (debugging) print *,
air%t: start
dx_dt%x=this%sigma*(y(1)-this%x)
dx_dt%sigma=0.
if (debugging) print *,
air%t: end
end function

70
71
72
73
74
75
76
77
78
79
80
81
82
83
84

! contribution to diagonal Jacobian element

function d_dAir(this,y) result(dRHS_dx)
class(air) ,intent(in)
:: this
real
,dimension(:,:) ,allocatable :: dRHS_dx
real
,dimension(:) ,allocatable :: y
if (debugging) print *,
air%d_dAir: start
!allocate(dRHS_dx(num_eqs,num_vars))
!dRHS_dx = [ d{sigma*(y-x)}/dx
d{sigma*(y-x)}/dsigma ]
!
[ d{0}/dx
d{0}/dsigma
]
if (size(y) /= 1) stop d_dAir: invalid y size
dRHS_dx = reshape(source=(/-this%sigma,0.,y(1)-this%x,0./),
shape=(/num_eqs,num_vars/))
if (debugging) print *,
air%d_dAir: end
end function

85
86
87
88
89
90

! contribution to off-diagonal Jacobian element

function d_dy(this,y) result(dRHS_dy)
class(air) ,intent(in)
:: this
real
,dimension(:,:) ,allocatable :: dRHS_dy
real
,dimension(:)
,allocatable :: y

91
92
93
94
95
96
97
98
99

if (debugging) print *,
air%d_dy:
allocate(dRHS_dy(num_eqs,size(y)))
!dRHS_dy = [ d{sigma*(y(1)-x(1))}/dy(1)
!
[ d{0}/dy(1)
dRHS_dy = 0.
dRHS_dy(1,1) = this%sigma
if (debugging) print *,
air%d_dy:
end function

start
0
0

... 0]
... 0]

end

100
101
102
103

function add(lhs,rhs) result(sum) ! add two instances

class(air) ,intent(in) :: lhs,rhs
type(air)
:: sum

104
105
106
107
108
109

if (debugging) print *,
air%add: start
sum%x
= lhs%x
+ rhs%x
sum%sigma = lhs%sigma + rhs%sigma
if (debugging) print *,
air%add: end
end function

&

8.2 The Solution

110
111
112
113
114
115
116
117
118
119

! subtract one instance from another

function subtract(lhs,rhs) result(difference)
class(air) ,intent(in) :: lhs,rhs
type(air)
:: difference
if (debugging) print *,
air%subtract: start
difference%x
= lhs%x
- rhs%x
difference%sigma = lhs%sigma - rhs%sigma
if (debugging) print *,
air%subtract: end
end function

120
121
122
123
124
125
126
127
128
129
130
131

1
2

! multiply an instance by a real scalar

function multiply(lhs,rhs) result(product)
class(air) ,intent(in) :: lhs
real
,intent(in) :: rhs
type(air)
:: product
if (debugging) print *,
air%multiply: start
product%x
= lhs%x
*rhs
product%sigma = lhs%sigma*rhs
if (debugging) print *,
air%multiply: end
end function
end module air_module
Figure 8.5(b): cloud.f03
module cloud_module
use global_parameters_module ,only :debugging !print call tree if true

3
4
5
6

implicit none ! Prevent implicit typing

private
! Hide everything by default
public :: cloud

7
8
9
10
11
12
13
14
15
16
17

!
!
!
!
!
!
!
!
!
!

This type tracks the evolution of the second state variable in the
Lorenz system according to the second Lorenz equation. It also
tracks the corresponding paramater (rho) according to the
differential equation d(rho)/dt=0. For illustrative purposes, this
implementation does not expose the number of state variables (2) to
the puppeteer because no iteration is required and the need for
arithmetic operations on rho is therefore an internal concern. The
rank of the matrix d_dCloud() returns is thus 1 to reflect the only
variable on which the puppeteer needs to iterate when handling
nonlinear couplings in implicit solvers.

18
19
20
21

! Number of evolution equations/variables exposed to the outside

! world (via Jacobian sub-block shapes):
integer ,parameter :: num_eqs=1,num_vars=1

22
23
24
25
26
27
28
29
30
31

type cloud
private
real :: y,rho ! 2nd Lorenz equation solution variable and parameter
contains
procedure :: set_coordinate ! accessor: set phase-space coordinate
procedure :: coordinate! accessor: return phase-space coordinate
procedure :: t
! time derivative
procedure :: d_dCloud ! contribution to diagonal Jacobian element
procedure :: d_dx ! contribution to off-diagonal Jacobian element

179

180

The Puppeteer Pattern

32
33
34
35
36
37
38
39
40

procedure :: d_dz ! contribution to off-diagonal Jacobian element

procedure ,private :: add
! add two instances
procedure ,private :: subtract ! subtract one instance from another
procedure ,private :: multiply ! multiply instance by a real scalar
! map defined operators to corresponding procedures
generic
,public :: operator(+) => add
generic
,public :: operator(-) => subtract
generic
,public :: operator(*) => multiply
end type cloud

41
42
43
44

interface cloud
procedure constructor
end interface

45
46
47
48
49
50
51
52
53
54
55

contains
! constructor: allocate and initialize components
type(cloud) function constructor(y_initial,r)
real ,intent(in) :: y_initial
real ,intent(in) :: r
if (debugging) print *,
cloud: start
constructor%y = y_initial
constructor%rho = r
if (debugging) print *,
cloud: end
end function

56
57
58
59
60
61
62

! accessor: set component

subroutine set_coordinate(this,y_update)
class(cloud) ,intent(inout) :: this
real ,intent(in) :: y_update
this%y = y_update
end subroutine

63
64
65
66
67
68
69

! accessor (returns phase-space coordinate)

function coordinate(this) result(return_y)
class(cloud)
,intent(in) :: this
real ,dimension(:) ,allocatable :: return_y
return_y = [this%y]
end function

70
71
72
73
74
75
76
77
78
79
80
81

! contribution to diagonal Jacobian element

function d_dCloud(this,x_ignored,z_ignored) result(dRHS_dy)
class(cloud)
,intent(in) :: this
real ,dimension(:)
,allocatable ,intent(in) :: x_ignored,z_ignored
real ,dimension(:,:) ,allocatable
:: dRHS_dy
if (debugging) print *,
cloud%d_dCloud: start
allocate(dRHS_dy(num_eqs,num_vars))
!dRHS_dy(1) = [ d{x(1)*(rho-z(1))-y(1)}/dy(1) ]
dRHS_dy(1,1) = -1.
if (debugging) print *,
cloud%d_dCloud: end
end function

82
83
84
85
86

! contribution to off-diagonal Jacobian element

function d_dx(this,x,z) result(dRHS_dx)
class(cloud)
,intent(in) :: this
real ,dimension(:)
,allocatable ,intent(in) :: x,z

8.2 The Solution

87
88
89
90
91
92
93
94

real ,dimension(:,:) ,allocatable

:: dRHS_dx
if (debugging) print *,
cloud%d_dx: start
allocate(dRHS_dx(num_eqs,size(x)))
!dRHS_dx = [ d{x(1)*(rho-z(1))-y}/dx(1) 0 ... 0]
dRHS_dx = 0.
dRHS_dx(1,1) = this%rho-z(1)
if (debugging) print *,
cloud%d_dx: end
end function

95
96
97
98
99
100
101
102
103
104
105
106
107

! contribution to off-diagonal Jacobian element

function d_dz(this,x,z) result(dRHS_dz)
class(cloud)
,intent(in) :: this
real ,dimension(:)
,allocatable ,intent(in) :: x,z
real ,dimension(:,:) ,allocatable
:: dRHS_dz
if (debugging) print *,
cloud%d_dz: start
allocate(dRHS_dz(num_eqs,size(z)))
!dRHS_dz = [ d{x(1)*(rho-z(1))-y(1)}/dz(1) 0 ... 0]
dRHS_dz = 0.
dRHS_dz(1,1) = -x(1)
if (debugging) print *,
cloud%d_dz: end
end function

108
109
110
111
112
113
114
115
116
117
118

! time derivative (evolution equations)

function t(this,x,z) result(dy_dt)
class(cloud)
,intent(in) :: this
real ,dimension(:) ,intent(in) :: x,z
type(cloud)
:: dy_dt
if (debugging) print *,
cloud%t: start
dy_dt%y = x(1)*(this%rho-z(1))-this%y
dy_dt%rho = 0.
if (debugging) print *,
cloud%t: end
end function

119
120
121
122
123
124
125
126
127

function add(lhs,rhs) result(sum) ! add two instances

class(cloud) ,intent(in) :: lhs,rhs
type(cloud)
:: sum
if (debugging) print *,
cloud%add: start
sum%y
= lhs%y
+ rhs%y
sum%rho = lhs%rho + rhs%rho
if (debugging) print *,
cloud%add: end
end function

128
129
130
131
132
133
134
135
136
137

! subtract one instance from another

function subtract(lhs,rhs) result(difference)
class(cloud) ,intent(in) :: lhs,rhs
type(cloud)
:: difference
if (debugging) print *,
cloud%subtract: start
difference%y
= lhs%y
- rhs%y
difference%rho = lhs%rho - rhs%rho
if (debugging) print *,
cloud%subtract: end
end function

138
139
140
141

! multiply an instance by a real scalar

function multiply(lhs,rhs) result(product)
class(cloud) ,intent(in) :: lhs

181

182

The Puppeteer Pattern

142
143
144
145
146
147
148
149

1
2

real
,intent(in) :: rhs
type(cloud)
:: product
if (debugging) print *,
cloud%multiply: start
product%y
= lhs%y* rhs
product%rho = lhs%rho* rhs
if (debugging) print *,
cloud%multiply: end
end function
end module cloud_module
Figure 8.5(c): ground.f03
module ground_module
use global_parameters_module ,only :debugging !print call tree if true

3
4
5
6

implicit none ! Prevent implicit typing

private
! Hide everything by default
public :: ground

7
8
9
10
11
12
13
14
15
16
17

!
!
!
!
!
!
!
!
!
!

This type tracks the evolution of the third state variable in the
Lorenz system according to the third Lorenz equation. It also
tracks the corresponding paramater (beta) according to the
differential equation d(beta)/dt=0. For illustrative purposes,
this implementation does not expose the number of state variables(2)
to the puppeteer because no iteration is required and the need for
arithmetic operations on beta is therefore an internal concern. The
rank of the matrix d_dGround() returns is thus 1 to reflect the
only variable on which the puppeteer needs to iterate when handling
nonlinear couplings in implicit solvers.

18
19
20
21

! Number of evolution equations/variables exposed to the outside

! world (via Jacobian sub-block shapes):
integer ,parameter :: num_eqs=1,num_vars=1

22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40

type ground
private
real :: z,beta ! 3rd Lorenz equation solution variable and parameter
contains
procedure :: set_coordinate ! accessor set phase-space coordinate
procedure :: coordinate ! accessor: return phase-space coordinate
procedure :: t
! time derivative
procedure :: d_dGround ! contribution to diagonal Jacobian element
procedure :: d_dx
! contribution to off-diagonal Jacobian element
procedure :: d_dy
! contribution to off-diagonal Jacobian element
procedure ,private :: add ! add two instances
procedure ,private :: subtract ! subtract one instance from another
procedure ,private :: multiply ! multiply instance by a real scalar
! map defined operators to corresponding procedures
generic ,public :: operator(+) => add
generic ,public :: operator(-) => subtract
generic ,public :: operator(*) => multiply
end type

41
42
43
44
45

interface ground
procedure constructor
end interface

8.2 The Solution

46

contains

47
48
49
50
51
52
53
54
55
56

! constructor: allocate and initialize components

type(ground) function constructor(z_initial,b)
real ,intent(in) :: z_initial
real ,intent(in) :: b
if (debugging) print *,
ground%construct: start
constructor%z = z_initial
constructor%beta = b
if (debugging) print *,
ground%construct: end
end function

57
58
59
60
61
62
63

! accessor (returns phase-space coordinate)

function coordinate(this) result(return_z)
class(ground)
,intent(in) :: this
real ,dimension(:) ,allocatable :: return_z
return_z = [ this%z ]
end function

64
65
66
67
68
69
70

! accessor (returns phase-space coordinate)

subroutine set_coordinate(this,z_update)
class(ground) ,intent(inout) :: this
real ,intent(in) :: z_update
this%z = z_update
end subroutine

71
72
73
74
75
76
77
78
79
80
81

! time derivative (evolution equations)

function t(this,x,y) result(dz_dt)
class(ground)
,intent(in) :: this
real ,dimension(:) ,intent(in) :: x,y
type(ground)
:: dz_dt
if (debugging) print *,
ground%t: start
dz_dt%z = x(1)*y(1) - this%beta*this%z
dz_dt%beta = 0.
if (debugging) print *,
ground%t: end
end function

82
83
84
85
86
87
88
89
90
91
92
93

! contribution to diagonal Jacobian element

function d_dGround(this,x_ignored,y_ignored) result(dRHS_dz)
class(ground)
,intent(in) :: this
real ,dimension(:)
,allocatable ,intent(in) :: x_ignored,y_ignored
real ,dimension(:,:) ,allocatable
:: dRHS_dz
if (debugging) print *,
ground%d_dGround: start
!dRHS_dz = [ d{x(1)*y(1) - beta*z}/dz(1) ]
allocate(dRHS_dz(num_eqs,num_vars))
dRHS_dz(1,1) = -this%beta
if (debugging) print *,
ground%d_dGround: end
end function

94
95
96
97
98
99
100

! contribution to off-diagonal Jacobian element

function d_dx(this,x,y) result(dRHS_dx)
class(ground)
,intent(in) :: this
real ,dimension(:)
,allocatable ,intent(in) :: x,y
real ,dimension(:,:) ,allocatable
:: dRHS_dx
if (debugging) print *,
ground%d_dx: start

183

184

The Puppeteer Pattern

101
102
103
104
105
106

allocate(dRHS_dx(num_eqs,size(x)))
!dRHS_dz = [ d{x(1)*y(1) - beta*z(1)}/dx(1)
0 ...
dRHS_dx=0.
dRHS_dx(1,1) = y(1)
if (debugging) print *,
ground%d_dx: end
end function

0 ]

107
108
109
110
111
112
113
114
115
116
117
118
119

! contribution to off-diagonal Jacobian element

function d_dy(this,x,y) result(dRHS_dy)
class(ground)
,intent(in) :: this
real ,dimension(:)
,allocatable ,intent(in) :: x,y
real ,dimension(:,:) ,allocatable
:: dRHS_dy
if (debugging) print *,
ground%d_dy: start
allocate(dRHS_dy(num_eqs,size(y)))
!dRHS_dz = [ d{x(1)*y(1) - beta*z(1)}/dy(1)
0 ... 0 ]
dRHS_dy = 0.
dRHS_dy(1,1) = x(1)
if (debugging) print *,
ground%d_dy: end
end function

120
121
122
123
124
125
126
127
128

function add(lhs,rhs) result(sum) ! add two instances

class(ground) ,intent(in) :: lhs,rhs
type(ground)
:: sum
if (debugging) print *,
ground%add: start
sum%z
= lhs%z
+ rhs%z
sum%beta = lhs%beta + rhs%beta
if (debugging) print *,
ground%add: end
end function

129
130
131
132
133
134
135
136
137
138

! subtract one instance from another

function subtract(lhs,rhs) result(difference)
class(ground) ,intent(in) :: lhs,rhs
type(ground)
:: difference
if (debugging) print *,
ground%subtract: start
difference%z
= lhs%z
- rhs%z
difference%beta = lhs%beta - rhs%beta
if (debugging) print *,
ground%subtract: end
end function

139
140
141
142
143
144
145
146
147
148
149
150

! multiply an instance by a real scalar

function multiply(lhs,rhs) result(product)
class(ground) ,intent(in) :: lhs
real
,intent(in) :: rhs
type(ground)
:: product
if (debugging) print *,
ground%multiply: start
product%z
= lhs%z
* rhs
product%beta = lhs%beta* rhs
if (debugging) print *,
ground%multiply: end
end function
end module ground_module
Figure 8.5. Fortran 2003 implementation of puppets for an atmosphere system. (a) The air
puppet class. It tracks the rst-state variable in the Lorenz system. (b) The cloud puppet class.
It tracks the second-state variable in the Lorenz system. (c) The ground puppet class. It tracks
the third-state variable in the Lorenz system.

8.2 The Solution

Figure 8.6(a) illustrates the process of a puppeteer (the atmosphere) working

with its puppets (the air, cloud, and ground) for an atmosphere system. Figure 8.6
(b) provides a global debug ag that can be used to turn on addition debugging
information when running the application.

8.2.2 A C++ Tool: 2D Allocatable Arrays

The use of Fortran two-dimensional allocatable real array is simulated using a mat_t
class in C++. The denition of this class is shown in Figure 8.7. Internally the actual data, denoted by data member data_, is stored using an STLs vector of
type float. Two additional attributes, r_ and c_, are used to represent the shape
(row and column) of a two dimensional array. Some member functions emulate
the behaviors of an allocatable array in Fortran. For instance, the auto-allocation
and auto-reallocation of a 2D allocatable arrays in an assignment are achieved by
resize() and clear_resize(), respectively. 2
One syntactic distinction we note here is the difference in denoting a twodimensional array and its element ordering by subscripts between Fortran and C++.
Both languages represent a 2D array as a sequence of scalars arranged in a rectangular pattern; for example, it can be viewed as a matrix composed of rows and columns.
However, the same subscript values normally result in completely different element
indices in the array by the two languages. One well-understood difference is the index
base: C++ is zero-based indexing (subscript starts from zero), whereas by default,
Fortran uses one-based indexing (subscript value starts from one) although Fortran
allows one to use arbitrary value as lower bound.

1
2
3
4
5
6

Figure 8.6(a): main.f03

program main
use air_module ,only : air
use cloud_module ,only :cloud
use ground_module ,only : ground
use atmosphere_module ,only : atmosphere
use global_parameters_module ,only :debugging !print call tree if true

7
8

implicit none ! Prevent implicit typing

9
10
11
12
13

!
!
!
!

This code integrates the Lorenz equations over time using separate
abstractions for equation and hiding the coupling of those
abstractions inside an abstraction that follows the Puppeteer design
pattern of Rouson, Adalsteinsson and Xia (ACM TOMS 37:1, 2010).

14
15
16
17
18
19

type(air)
,allocatable
type(cloud)
,allocatable
type(ground) ,allocatable
type(atmosphere)
integer

::
::
::
::
::

sky
!puppet for
puff !puppet for
earth !puppet for
boundary_layer !
step
!

1st Lorenz equation

2nd Lorenz equation
3rd Lorenz equation
Puppeteer
time step

2 In class denition for mat_t, we only provide a minimum set of denitions of functions and operators

that are actually used throughout the example. It is by no means to be used as a realistic class
denition for two-dimensional allocatable array. Readers are encouraged to extend this class so it
becomes more complete.

185

186

The Puppeteer Pattern

20
21
22

integer ,parameter
real
,parameter
real

23
24

real

,parameter

:: num_steps=1000 ! total time steps

:: x=1.,y=1.,z=1. ! initial conditions
:: t
! time coordinate
! Lorenz parameters
:: sigma=10.,rho=28.,beta=8./3.,dt=.02

25
26
27
28
29
30

if (debugging) print *,main: start

allocate (sky, puff, earth)
sky = air(x,sigma)
puff = cloud(y,rho)
earth = ground(z,beta)

31
32
33
34

! transfer allocations into puppeteer

boundary_layer = atmosphere(sky,puff,earth)
! all puppets have now been deallocated

35
36
37
38
39
40
41
42
43
44
45

t=0.
write(*,(f10.4),advance=no) t
print *,boundary_layer%state_vector()
do step=1,num_steps
call integrate(boundary_layer,dt)
t = t + dt
write(*,(f10.4),advance=no) t
print *,boundary_layer%state_vector()
end do
if (debugging) print *,main: end

46
47
48
49
50
51
52
53
54
55

contains
! abstract Trapezoidal rule integration
subroutine integrate(integrand,dt)
type(atmosphere) ,intent(inout) :: integrand
real
,intent(in)
:: dt
type(atmosphere) ,allocatable
:: integrand_estimate,residual
integer ,parameter :: num_iterations=5
integer :: newton_iteration,num_equations,num_statevars,i,j
real ,dimension(:,:) ,allocatable :: dRHS_dState,jacobian,identity

56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74

if (debugging) print *, integrate: start

allocate(integrand_estimate, source=integrand)
allocate(residual)
do newton_iteration=1,num_iterations
dRHS_dState
= integrand_estimate%dRHS_dV()
num_equations = size(dRHS_dState,1)
num_statevars = size(dRHS_dState,2)
if (num_equations /= num_statevars) &
stop integrate: ill-posed problem.
identity = reshape( &
source=(/((0.,i=1,num_equations),j=1,num_statevars)/), &
shape=(/num_equations,num_statevars/) )
forall(i=1:num_equations) identity(i,i)=1.
jacobian = identity - 0.5*dt*dRHS_dState
residual = integrand_estimate - &
( integrand + (integrand%t() + integrand_estimate%t())*(0.5*dt))
integrand_estimate = integrand_estimate - &
(jacobian .inverseTimes. residual)

8.2 The Solution

75
76
77
78
79

1
2
3
4
5

end do
integrand = integrand_estimate
if (debugging) print *, integrate: end
end subroutine
end program main
Figure 8.6(b): global_parameters.f03
module global_parameters_module
! parameter that is used to control the debugging. The call tree
! information is printed if value of debugging is true
logical ,parameter :: debugging=.false.
end module
Figure 8.6. Fortran 2003 implementation of puppeteer pattern applied to an atmosphere
system. (a) Time-integration test of Lorenz system using puppeteer pattern. (b) A global
debugging ag.

Another difference is the element ordering with regard to dimensions. Fortrans

element sequence ordering (traversal rule) is column-based, whereas C++ is rowbased. This second difference can sometimes be problematic in applications mixing
the two languages. Consider the following declaration of a 3 3 matrix x,
Fortran:
real
C++:
float

x(3,3)

! Fortran declaration

x[3][3]; // C++ declaration

In Fortran, x(3,1) refers to the third element of x, and x(1,3) refers to seventh element
of x. The corresponding C++ expressions are x[0][2] and x[2][0], respectively.
In designing mat_ class, we choose zero-based notation because it is the most
natural way to express array subscripts to C++ programmers. We also adhere to Fortrans convention in element ordering. The computation of element order is shown
in the denitions of operator() in gure 8.7(b) as follows:
mat_t::value_type mat_t::operator()(int r, int c) const {
// error checking for invalid subscript range; code omitted
return data_.at(c*r_ + r);
}
Thus for the aforementioned 3 3 matrix, mat_(2, 0) corresponds to Fortrans x(3,
1), and mat_(0, 2) corresponds to x(1, 3).

1
2

#ifndef _H_MAT_
#define _H_MAT_

3
4
5
6

#include "globals.h"
#include <iostream>
#include <iomanip>

Figure 8.7(a): mat.h

187

188

The Puppeteer Pattern

7
8
9
10
11

class mat_t {
public:
typedef crd_t::value_type value_type;
typedef crd_t::reference reference;

12
13
14
15
16
17
18
19
20
21
22
23
24
25

mat_t();
mat_t(int rows, int cols);
void clear();
void resize(int rows, int cols);
void clear_resize(int rows, int cols, value_type value = 0);
void identity(int rows);
int rows() const;
int cols() const;
value_type operator()(int r, int c) const;
reference operator()(int r, int c);
void set_submat(int r, int c, const mat_t &other);
mat_t& operator-=(const mat_t&);
mat_t& operator*=(real_t);

26
27
28
29
30

private:
int r_, c_;
crd_t data_;
};

31
32
33
34
35
36

inline mat_t operator-(const mat_t &a, const mat_t &b) {

mat_t retval(a);
retval -= b;
return retval;
}

37
38
39
40
41
42

inline mat_t operator*(real_t value, const mat_t &matrix) {

mat_t retval(matrix);
retval *= value;
return retval;
}

43
44
45
46

struct dim_t {
const int eqs;
const int vars;

47
48
49
50
51

dim_t(int eqcnt, int varcnt) :

eqs(eqcnt), vars(varcnt)
{}
};

52
53
54
55
56
57
58
59
60
61

inline std::ostream& operator<<(std::ostream &os, const mat_t &mat) {

std::ios_base::fmtflags flags = os.flags();
for(int r = 0; r < mat.rows(); ++r) {
os << "[";
for(int c = 0; c < mat.cols(); ++c) {
os << " " <<std::setw(12) <<std::setprecision(8)
<<std::fixed <<mat(r,c);
}
os << "]\n";

8.2 The Solution

62
63
64
65

}
os.flags(flags);
return os;

66
67

1
2
3

#endif // !_H_MAT_

#include "mat.h"
#include <exception>
#include <iostream>

Figure 8.7(b): mat.cpp

4
5
6
7

struct matrix_error : public std::exception {

virtual matrix_error() throw() {}
};

8
9
10
11

mat_t::mat_t() :
r_(0), c_(0)
{}

12
13
14
15

mat_t::mat_t(int rows, int cols) {

this->resize(rows, cols);
}

16
17
18
19
20

void mat_t::clear() {
r_ = c_ = 0;
data_.clear();
}

21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36

void mat_t::resize(int rows, int cols) {

if(rows < 0 || cols < 0) {
std::cerr << "mat_t::resize: Rows and columns must be >= 0.\n";
throw matrix_error();
}
if(! data_.empty()) {
// Copy data over.
}
else {
// common case.
data_.resize(rows*cols);
r_ = rows;
c_ = cols;
}
}

37
38
39
40
41
42
43
44
45
46
47

void mat_t::clear_resize(int rows, int cols, value_type value) {

if(rows < 0 || cols < 0) {
std::cerr << "mat_t::clear_resize: Rows and columns must be >= 0\n";
throw matrix_error();
}
data_.resize(rows*cols);
r_ = rows;
c_ = cols;
std::fill(data_.begin(), data_.end(), value);
}

189

190

The Puppeteer Pattern

48
49
50
51
52
53
54

void mat_t::identity(int size) {

this->clear_resize(size, size, 0);
for(int i = 0; i < size; ++i) {
this->operator()(i,i) = 1;
}
}

55
56
57
58

int mat_t::rows() const {

return r_;
}

59
60
61
62

int mat_t::cols() const {

return c_;
}

63
64
65
66
67
68
69
70
71

mat_t::value_type mat_t::operator()(int r, int c) const {

if(r < 0 || r >= r_ || c < 0 || c >= c_) {
std::cerr << "mat_t::operator(): Invalid index (" << r << ", " << c
<< "). Bounds are (" << r_ << ", " << c << ")\n";
throw matrix_error();
}
return data_.at(c*r_ + r);
}

72
73
74
75
76
77
78
79
80

mat_t::reference mat_t::operator()(int r, int c) {

if(r < 0 || r >= r_ || c < 0 || c >= c_) {
std::cerr << "mat_t::operator(): Invalid index (" << r << ", " << c
<< "). Bounds are (" << r_ << ", " << c << ")\n";
throw matrix_error();
}
return data_.at(c*r_ + r);
}

81
82
83
84
85
86
87
88

void mat_t::set_submat(int startrow, int startcol, const mat_t &other) {

for(int r = 0; r < other.rows(); ++r) {
for(int c = 0; c < other.cols(); ++c) {
this->operator()(r+startrow, c+startcol) = other(r,c);
}
}
}

89
90
91
92
93
94
95
96
97
98
99
100

mat_t& mat_t::operator-=(const mat_t &other) {

if(this->rows() != other.rows() || this->cols() != other.cols()) {
std::cerr <<
"mat_t::operator-=: Matrices must be of identical size.\n";
throw matrix_error();
}
const size_t size = data_.size();
for(size_t i = 0; i < size; ++i)
data_[i] -= other.data_[i];
return *this;
}

101
102

mat_t& mat_t::operator*=(real_t value) {

8.2 The Solution

103
104
105
106

for(crd_t::iterator it = data_.begin(); it != data_.end(); ++it)

*it *= value;
return *this;

Figure 8.7. A mat_t class in C++ that simulates the functionality of 2D allocatable real arrays in
Fortran. (a) Class denition for mat_t, which simulates Fortrans two-dimensional allocatable
array of type real. (b) Implementation of class mat_t.

8.2.3 C++ Implementation

Figure 8.8 shows the C++ denition for the atmosphere class, the puppeteer in a
weather system. In Figure 8.9, we give C++ implementations for air, cloud, and
ground classes. These are the puppets of the atmosphere system. Figure 8.10 shows
the C++ puppeteer corresponding to Fortran puppeteer given in Figure 8.6. Again
the similarity to the Fortran counterpart is obvious.

1
2

#ifndef _H_ATMOSPHERE_
#define _H_ATMOSPHERE_

Figure 8.8(a): atmosphere.h

3
4
5
6
7

#include
#include
#include
#include

"integrand.h"
"air.h"
"cloud.h"
"ground.h"

8
9
10
11
12

class atmosphere : public integrand {

public:
typedef integrand::ptr_t ptr_t;
typedef Ref<atmosphere> self_t;

13
14
15

atmosphere(const air&, const cloud&, const ground&);

virtual atmosphere();

16
17
18
19
20
21
22

// The following methods do dynamic allocation.

virtual ptr_t d_dt() const;
virtual void dRHS_dV(mat_t&) const;
virtual ptr_t clone() const;
virtual ptr_t inverse_times(const mat_t&) const;
virtual crd_t state_vector() const;

23
24
25
26
27

// The following methods are destructive updates.

virtual ptr_t operator+=(ptr_t);
virtual ptr_t operator-=(ptr_t);
virtual ptr_t operator*=(real_t);

28
29
30
31
32
33

private:
air air_;
cloud cloud_;
ground ground_;
};

34
35

#endif //!_H_ATMOSPHERE_

191

192

The Puppeteer Pattern

1
2
3

#include "atmosphere.h"
#include <exception>
#include <cmath>

Figure 8.8(b): atmosphere.cpp

4
5
6
7

struct atmosphere_error : public std::exception {

virtual atmosphere_error() throw() {}
};

8
9

typedef atmosphere::ptr_t ptr_t;

10
11
12
13

atmosphere::atmosphere(const air &a, const cloud &c, const ground &g) :

air_(a), cloud_(c), ground_(g)
{}

14
15
16

atmosphere::atmosphere() {
}

17
18
19
20
21
22
23

ptr_t atmosphere::d_dt() const {

return
ptr_t(new atmosphere(air_.d_dt(cloud_.coordinate()),
cloud_.d_dt(air_.coordinate(),ground_.coordinate()),
ground_.d_dt(air_.coordinate(),cloud_.coordinate())));
}

24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54

void atmosphere::dRHS_dV(mat_t &result) const {

// Figure out the required dimensions.
dim_t adim = air_.dimensions();
dim_t cdim = cloud_.dimensions();
dim_t gdim = ground_.dimensions();
// Resize the result matrix.
result.clear_resize(adim.eqs + cdim.eqs + gdim.eqs,
adim.vars + cdim.vars + gdim.vars);
if(result.rows() != result.cols()) {
std::cerr << "atmosphere::dRHS_dV: Ill-formed problem: total of "
<< result.rows() << " equations and " << result.cols()
<< " variables\n";
throw atmosphere_error();
}
// dAir/dAir
result.set_submat(0, 0, air_.d_dAir(cloud_.coordinate()));
// dAir/dCloud
result.set_submat(0, adim.vars, air_.d_dy(cloud_.coordinate()));
// dAir/dGround is all zero -- skipping that one.
// dCloud/dAir
result.set_submat(adim.eqs, 0,
cloud_.d_dx(air_.coordinate(),ground_.coordinate()));
// dCloud/dCloud
result.set_submat(adim.eqs, adim.vars,
cloud_.d_dCloud(air_.coordinate(),ground_.coordinate()));
// dCloud/dGround
result.set_submat(adim.eqs, adim.vars+cdim.vars,
cloud_.d_dz(air_.coordinate(),ground_.coordinate()));
// dGround/dAir
result.set_submat(adim.eqs+cdim.eqs, 0,

8.2 The Solution

55
56
57
58
59
60
61
62

ground_.d_dx(air_.coordinate(),cloud_.coordinate()));
// dGround/dCloud
result.set_submat(adim.eqs+cdim.eqs, adim.vars,
ground_.d_dy(air_.coordinate(),cloud_.coordinate()));
// dGround/dGround
result.set_submat(adim.eqs+cdim.eqs, adim.vars+cdim.vars,
ground_.d_dGround(air_.coordinate(),cloud_.coordinate()));

63
64
65
66

ptr_t atmosphere::clone() const {

return ptr_t(new atmosphere(*this));
}

67
68
69
70
71
72
73
74

crd_t atmosphere::state_vector() const {

const crd_t &cc = cloud_.coordinate(), &gc = ground_.coordinate();
crd_t state_space = air_.coordinate();
state_space.insert(state_space.end(), cc.begin(), cc.end());
state_space.insert(state_space.end(), gc.begin(), gc.end());
return state_space;
}

75
76
77
78

ptr_t atmosphere::inverse_times(const mat_t &lhs) const {

static const real_t pivot_tolerance = 1e-2;

79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109

const int n = lhs.rows();

crd_t b = this->state_vector();
if((n != lhs.cols()) || (size_t(n) != b.size())) {
std::cerr <<"integrand::inverse_times: ill-posed matrix problem\n";
throw atmosphere_error();
}
crd_t x(n);
mat_t A(lhs);
for(int p = 0; p < n-1; ++p) { // forward elimination
if(fabs(A(p,p)) < pivot_tolerance) {
std::cerr << "integrand::inverse_times: "
<< "use an algorithm with pivoting\n";
throw atmosphere_error();
}
for(int row = p+1; row < n; ++row) {
real_t factor = A(row,p) / A(p,p);
for(int col = p; col < n; ++col) {
A(row,col) = A(row,col) - A(p,col)*factor;
}
b.at(row) = b.at(row)- b.at(p)*factor;
}
}
x.at(n-1) = b.at(n-1) / A(n-1,n-1); // back substitution
for(int row = n-1; row >= 0; --row) {
real_t the_sum = 0;
for(int col = row+1; col < n; ++col) {
the_sum += A(row,col) * x.at(col);
}
x.at(row) = (b.at(row) - the_sum) / A(row,row);
}

193

194

The Puppeteer Pattern

110
111
112
113

return ptr_t(new atmosphere(air(x.at(0), x.at(1)),

cloud(x.at(2), cloud_.rho()),
ground(x.at(3), ground_.beta())));

114
115
116
117
118
119
120
121
122
123
124
125

ptr_t atmosphere::operator+=(ptr_t other) {

self_t added = cast<atmosphere>(other);
if(other == NULL) {
std::cerr << "atmosphere::operator+=: Invalid input type\n";
throw atmosphere_error();
}
air_
+= added->air_;
cloud_ += added->cloud_;
ground_ += added->ground_;
return ptr_t(this);
}

126
127
128
129
130
131
132
133
134
135
136
137

ptr_t atmosphere::operator-=(ptr_t other) {

self_t subbed = cast<atmosphere>(other);
if(other == NULL) {
std::cerr << "atmosphere::operator-=: Invalid input type\n";
throw atmosphere_error();
}
air_
-= subbed->air_;
cloud_ -= subbed->cloud_;
ground_ -= subbed->ground_;
return ptr_t(this);
}

138
139
140
141
142
143
144

ptr_t atmosphere::operator*=(real_t value) {

air_
*= value;
cloud_ *= value;
ground_ *= value;
return ptr_t(this);
}
Figure 8.8. Atmosphere class denitions in C++. (a) Class denition for atmosphere. (b)
Implementation of class atmosphere. (c) Denition for time-integration operation.

1
2

#ifndef _H_AIR_
#define _H_AIR_

Figure 8.9(a): air.h

3
4

#include "mat.h"

5
6
7
8

class air {
public:
air(real_t x, real_t sigma);

9
10
11
12
13
14

const crd_t& coordinate() const;

air d_dt(const crd_t&) const;
mat_t d_dAir(const crd_t&) const;
mat_t d_dy(const crd_t&) const;
air& operator+=(const air&);

8.2 The Solution

15
16

air& operator-=(const air&);

air& operator*=(real_t);

17
18

inline dim_t dimensions() const { return dim_t(dim_, dim_); }

19
20
21
22
23

private:
static const int dim_;
crd_t x_; // sigma is stored at x_[1]
};

24
25

#endif //!_H_AIR_

#include "air.h"

Figure 8.9(b): air.cpp

2
3

const int air::dim_ = 2;

4
5
6
7
8

air::air(real_t x, real_t sigma) {

x_.push_back(x);
x_.push_back(sigma);
}

9
10
11
12

const crd_t& air::coordinate() const {

return x_;
}

13
14
15
16

air air::d_dt(const crd_t &y) const {

return air((x_.at(1) * (y.at(0) - x_.at(0))), 0);
}

17
18
19
20
21
22
23

mat_t air::d_dAir(const crd_t& y) const {

mat_t result(dim_, dim_);
result(0, 0) = -x_.at(1);
result(0, 1) = y.at(0) - x_.at(0);
return result;
}

24
25
26
27
28
29

mat_t air::d_dy(const crd_t &y) const {

mat_t result(dim_, y.size());
result(0, 0) = x_.at(1);
return result;
}

30
31
32
33
34
35

air& air::operator+=(const air &other) {

x_.at(0) += other.x_.at(0);
x_.at(1)
+= other.x_.at(1);
return *this;
}

36
37
38
39
40
41
42

air& air::operator-=(const air &other) {

x_.at(0) -= other.x_.at(0);
x_.at(1)
-= other.x_.at(1);
return *this;
}

195

196

The Puppeteer Pattern

43
44
45
46
47

1
2

air& air::operator*=(real_t value) {

x_.at(0) *= value;
x_.at(1) *= value;
return *this;
}
Figure 8.9(c): cloud.h

#ifndef _H_CLOUD_
#define _H_CLOUD_

3
4

#include "mat.h"

5
6
7
8

class cloud {
public:
cloud(real_t y, real_t rho);

9
10
11
12
13
14
15
16
17
18

const crd_t& coordinate() const;

real_t rho() const;
cloud d_dt(const crd_t&, const crd_t&) const;
mat_t d_dCloud(const crd_t&, const crd_t&) const;
mat_t d_dx(const crd_t&, const crd_t&) const;
mat_t d_dz(const crd_t&, const crd_t&) const;
cloud& operator+=(const cloud&);
cloud& operator-=(const cloud&);
cloud& operator*=(real_t);

19
20

inline dim_t dimensions() const { return dim_t(dim_, dim_); }

21
22
23
24
25
26

private:
static const int dim_;
crd_t y_;
real_t rho_;
};

27
28

#endif //!_H_CLOUD_

#include "cloud.h"

Figure 8.9(d): cloud.cpp

2
3

const int cloud::dim_ = 1;

4
5
6
7

cloud::cloud(real_t y, real_t rho) :

y_(1, y), rho_(rho)
{}

8
9
10
11

const crd_t& cloud::coordinate() const {

return y_;
}

12
13
14
15

real_t cloud::rho() const {

return rho_;
}

16
17
18

cloud cloud::d_dt(const crd_t &x, const crd_t &z) const {

return cloud((x.at(0) * (rho_ - z.at(0)) - y_.at(0)), 0);

8.2 The Solution

19

20
21
22
23
24
25

mat_t cloud::d_dCloud(const crd_t&, const crd_t&) const {

mat_t result(dim_, dim_);
result(0, 0) = -1;
return result;
}

26
27
28
29
30
31

mat_t cloud::d_dx(const crd_t &x, const crd_t &z) const {

mat_t result(dim_, x.size());
result(0, 0) = rho_ - z.at(0);
return result;
}

32
33
34
35
36
37

mat_t cloud::d_dz(const crd_t &x, const crd_t &z) const {

mat_t result(dim_, z.size());
result(0, 0) = -x.at(0);
return result;
}

38
39
40
41
42
43
44
45
46
47
48

cloud& cloud::operator+=(const cloud &other) {

y_.at(0) += other.y_.at(0);
rho_ += other.rho_;
return *this;
}
cloud& cloud::operator-=(const cloud &other) {
y_.at(0) -= other.y_.at(0);
rho_ -= other.rho_;
return *this;
}

49
50
51
52
53
54

1
2

cloud& cloud::operator*=(real_t value) {

y_.at(0) *= value;
rho_ *= value;
return *this;
}

#ifndef _H_GROUND_
#define _H_GROUND_

Figure 8.9(e): ground.h

3
4

#include "mat.h"

5
6
7
8

class ground {
public:
ground(real_t y, real_t rho);

9
10
11
12
13
14
15
16
17

const crd_t& coordinate() const;

real_t beta() const;
ground d_dt(const crd_t&, const crd_t&) const;
mat_t d_dGround(const crd_t&, const crd_t&) const;
mat_t d_dx(const crd_t&, const crd_t&) const;
mat_t d_dy(const crd_t&, const crd_t&) const;
ground& operator+=(const ground&);
ground& operator-=(const ground&);

197

198

The Puppeteer Pattern

18

ground& operator*=(real_t);

19
20

inline dim_t dimensions() const { return dim_t(dim_, dim_); }

21
22
23
24
25
26

private:
static const int dim_;
crd_t z_;
real_t beta_;
};

27
28

#endif //!_H_GROUND_

#include "ground.h"

Figure 8.9(f): ground.cpp

2
3

const int ground::dim_ = 1;

4
5
6
7

ground::ground(real_t z, real_t beta) :

z_(1, z), beta_(beta)
{}

8
9
10
11

const crd_t& ground::coordinate() const {

return z_;
}

12
13
14
15

real_t ground::beta() const {

return beta_;
}

16
17
18
19

ground ground::d_dt(const crd_t& x, const crd_t& y) const {

return ground((x.at(0) * y.at(0) - beta_ * z_.at(0)), 0);
}

20
21
22
23
24
25

mat_t ground::d_dGround(const crd_t&, const crd_t&) const {

mat_t result(dim_, dim_);
result(0, 0) = -beta_;
return result;
}

26
27
28
29
30
31

mat_t ground::d_dx(const crd_t &x, const crd_t &y) const {

mat_t result(dim_, x.size());
result(0, 0) = y.at(0);
return result;
}

32
33
34
35
36
37

mat_t ground::d_dy(const crd_t &x, const crd_t &y) const {

mat_t result(dim_, y.size());
result(0, 0) = x.at(0);
return result;
}

38
39
40
41
42

ground& ground::operator+=(const ground &other) {

z_.at(0) += other.z_.at(0);
beta_ += other.beta_;
return *this;

8.3 The Consequences

43

44
45
46
47
48
49

ground& ground::operator-=(const ground& other) {

z_.at(0) -= other.z_.at(0);
beta_ -= other.beta_;
return *this;
}

50
51
52
53
54
55

ground& ground::operator*=(real_t value) {

z_.at(0) *= value;
beta_ *= value;
return *this;
}
Figure 8.9. Puppet class denitions in C++. (a)(b) Class denition of air. (c)(d) Class
denition of cloud. (e)(f) Class dention of ground.
Figure 8.10: main.cpp

1
2

#include "atmosphere.h"
#include "fmt.h"

3
4

typedef integrand::ptr_t ptr_t;

5
6
7
8

int main() {
const int num_steps=1000;
const real_t x=1., y=1., z=1., sigma=10., rho=28, beta=8./3., dt=0.02;

air sky(x, sigma);

cloud puff(y, rho);
ground earth(z, beta);
ptr_t boundary_layer = ptr_t(new atmosphere(sky, puff, earth));

10
11
12
13
14
15
16
17
18
19
20
21
22
23
24

real_t t = 0.;
std::cout << fmt(t,5,2) << " "
<< fmt(boundary_layer->state_vector()) << "\n";
for(int step = 1; step <= num_steps; ++step) {
integrate(boundary_layer, dt);
t += dt;
std::cout << fmt(t,5,2) << " "
<< fmt(boundary_layer->state_vector()) << "\n";
}

Figure 8.10. C++ implementation of puppeteer pattern applied to an atmosphere system

following the same design as that by Fortran 2003.

8.3 The Consequences

The chief consequence of the PUPPETEER pattern derives from its separation of concerns. Domain experts can construct ADTs that encapsulate widely disparate physics
without knowing the implementation details of the other ADTs. In the atmospheric
modeling example, a uid dynamicist might build an Air abstraction that solves the
Navier-Stokes equations for wind velocities and pressures, while a chemist might

199

200

The Puppeteer Pattern

build a Cloud abstraction that predicts acid rain by modeling chemical species adsorption at droplet surfaces. The PUPPETEER would rst request droplet locations
from the Cloud instance, then request species concentrations at the droplet location from the Air instance, and nally pass these concentrations to the Cloud in the
 and R/
 V.

process of constructing R
A second important consequence derives from the aforementioned containment
relationships. Implemented using Fortrans allocatable components, the PUPPETEER
can be viewed as holding only references to its puppets (similar to the assignment of a
smart pointer in C++, Fortrans move_alloc transfers the Air, Cloud, and Ground to
the corresponding puppets during the constructor call in 8.4). The actual construction
of puppets data and their lifetime, therefore, can be independent from the existence
of the PUPPETEER (see main in 8.6). This opens the possibility of varying the physical
models dynamically mid-simulation. In the atmospheric boundary layer model, the
cloud model would not have to be included until the atmospheric conditions became
ripe for cloud formation. Of course, when an absent object would otherwise supply
information required by another object, the PUPPETEER must substitute default values.
Whether this substitution happens inside the PUPPETEER or inside each ADT (via
optional arguments) is implementation-dependent.
The cost of separating concerns and varying the physics at runtime lies in the
conceptual work of discerning who does what, where they do it, and in which format.
Consider the following lines from the Fortran implementation of the trapezoidal time
integration algorithm. This algorithm is implemented in the integrate() procedure
of 8.6(a):
dRHS_dState
= integrand_estimate%dRHS_dV()
...
jacobian = identity - 0.5*dt*dRHS_dState
residual = integrand_estimate - &
( integrand + (integrand%t() + integrand_estimate%t())*(0.5*dt))
integrand_estimate = integrand_estimate - &
(jacobian .inverseTimes. residual)
where the ellipses indicate deleted lines.3 The rst line represents an abstract cal V.
 All calculations, including the assignment, happen inside the
culation of R/
PUPPETEER. This design decision stems from the fact that the details of the PUPPETEER
 so the information-hiding philosophy of
(and the objects it contains) determine R,
OOP precludes exposing these to integrate(). By contrast, the formula for calculating the Jacobian in the next line depends on the chosen time-integration algorithm,
so while dt*dRHS_dState represents an overloaded operation implemented inside
the PUPPETEER, that operation returns a Fortran array of intrinsic type (C++ primitive type), and the remainder of the line represents arithmetic on intrinsic entities.
Thus, the identity matrix and the Jacobian are simple oating-point arrays. Finally,
the subsequent two lines would be the same for any algorithm that employs implicit

3 The name inverseTimes is intended to be suggestive of the ultimate result. The sample code in 8.4

employs Gaussian elimination rather than computing and premultiplying the inverse Jacobian.

8.4 Related Patterns

advancement of nonlinear equations, so those lines represent overloaded arithmetic

carried out wholly inside the PUPPETEER.
Ultimately, we believe the forethought that goes into deciding what gets calculated where and by whom pays off in keeping the code exible. There is no hardwiring
of algorithm-specic details of the Jacobian calculation into individual physics abstractions. Nor is there any hardwiring of physics-specic details of the residual
calculation into the time integrator or the nonlinear solver. Each can be reused if the
implementation of the other changes in fundamental ways.
The puppeteer provides an elegant solution to a common problem in managing
complexity in multiscale, multiphysics applications. For efciency and numerical
stability, individual model components often run in nonidentical dimensions (e.g.,
when coupling two- and three-dimensional simulation models) or incompatible (e.g.,
reduced) units. All such considerations can be delegated to the puppeteer, greatly
easing the reuse of existing simulation software for the individual single-physics
models.
In any multiphysics package employing the ABSTRACT CALCULUS pattern, the
integrand abstract type would be highly stable because it depends on no other
ADTs. Hence, the afferent couplings vanish (Ce = 0). Also, the associated integrate procedure depends only on the dynamical system being integrated, which will
typically be a single puppeteer in a multiphysics simulation.

8.4 Related Patterns

Other parts of this book detail the similarities and close relationships between a PUPPETEER and other design pattern. Section 8.1 describes the ways in which a PUPPETEER
serves a role similar to that of a MEDIATOR. Section 6.4 details the close relationship
between a PUPPETEER and an ABSTRACT CALCULUS in the context of time integration.
In this context, a PUPPETEER might bear an equally close relationship with a STRATEGY
if the latter supplies numerical quadrature methods used by an ABSTRACT CALCULUS
for time integration. A PUPPETEER can serve as the concrete implementation of an
abstract FACADE dened by an ABSTRACT CALCULUS or a STRATEGY.

EXERCISES
1. Develop your own PUPPET pattern that abstracts the notion of an individual subsystem to be manipulated by a PUPPETEER. Shield the PUPPETEER from knowledge
of the concrete PUPPETS by having the PUPPETEER manipulate a collection of an
abstract puppet objects, each one having a private state vector along with public
methods for accepting and providing coupling terms. Write an atmosphere PUPPETEER with an three-element puppet array and pass it puppets corresponding to
each of the Lorenz equations (4.5). Advance your PUPPETEER by an explicit time
advancement algorithm such as 2nd-order Runge Kutta.
2. Consider the number of possible interpretations of the time integration line (46)
in Figure 6.2(b) with each interpretation corresponding to a unique dynamic type
that extends integrand type. Calculate the growth in information entropy as
each new single-physics abstraction is added to a multiphysics package.

201

Factory Patterns

Separate the physics from the data.

Jaideep Ray

9.1 The Problem

The patterns discussed in Chapters 67 adhere closely to the GoF philosophy of
designing to an interface rather than an implementation. That maxim inspired the use
of abstract classes in dening an ABSTRACT CALCULUS, a STRATEGY, and a SURROGATE.
The PUPPETEER denition in Chapter 8 represented the one setting in which client
code manipulated concrete implementations directly although an exercise at the
end of that chapter describes a way to liberate the PUPPETEER from the tyranny of
implementations also.
Independent of whether the classes comprising the pattern demonstrations are
abstract, the client codes (main programs in the cases considered) exploit knowledge
of the concrete type of each object constructed in Chapters 68. Although we were
able to write polymorphic procedures such as integrate in Figures 6.2(b) and 6.3(b)
(c), in each case, the actual arguments passed to these procedures were references
to concrete objects the client constructed. One can write even more exible code
by freeing clients of even this minimal knowledge of concrete types. This poses the
dilemma of where object construction happens. Put simply, how does one construct
an object without directly invoking its constructor? Or from another angle, how does
an object come to be if the client knows only the interface dened by its abstract
parent, but that parents abstract nature precludes denition of a constructor? By
denition, no instance of an abstract type can be instantiated.
The desire for any reference to the object to be abstract from its inception conicts with the need for that inception to result in a concrete instantiation. To place
this dilemma in a scientic context, consider the Burgers equation (4.9):
ut = uxx uux
which we might consider advancing from tn to tn+1 tn + t according to the explicit
Euler marching algorithm:
 n

un uxn
un+1 = un + t uxx
202

9.2 The Solution

where un and un+1 are the solution at tn and tn+1 , respectively. Given an interface,
field, to a family of 1D scalar eld implementations, we desire to write Fortran
client code statements on a field reference u of the form:
u = u + dt*(nu*u%xx() - u*u%x())
or equivalent C++ client statements of the form:
u += dt*(nu*u.xx() - u*u.x());
where the Fortran and C++ field classes are both abstract, and u can dynamically
(at runtime) take on any type that implements the field interface.
The quandary surrounding the where of object instantiation poses related questions about the when and the how of instantiation. Eliminating the clients control
over the instantiations spatial position in the code offers the potential to also eliminate the clients control over its temporal position and its path to completion.
Flexibility in the timing matters most when an objects construction occupies signicant resources such as memory or execution time. In such cases, it might be desirable
for a constructor to return a lightweight reference or an identifying tag but to delay
the actual resource allocation until the time of the objects rst use. Flexibility in the
steps taken to complete the instantiation matters most when the process involves
constructing multiple parts that must t coherently into a complex whole.
The solutions discussed in Section 9.2 focus primarily on the where question,
moving the object constructor invocation outside the client code and into a class
hierarchy parallel to the one in question. Section 9.4 briey discusses some of the
options for addressing the when and how questions.
Give Me Liberty and Give Me Birth!
The freedom associated with keeping clients references to an object abstract can
be realized throughout the objects lifetime if these references are abstract from
moment of the objects creation. This precludes letting a client invoke the constructors dened in the objects concrete implementation and creates a quandary
regarding how, where, and when the objects instantiation occurs.

9.2 The Solution

The GoF dene several creational design patterns that resolve the dilemmas presented in Section 9.1. This section details two: the ABSTRACT FACTORY and FACTORY
METHOD patterns. The GoF describe the intent of an ABSTRACT FACTORY as follows:
Provide an interface for creating families of related or dependent objects without
specifying their concrete classes.

The GoF demonstrated the construction of an ABSTRACT FACTORY for manufacturing

mazes for computer games. A maze comprised a set of rooms, walls, and doors. In an
ABSTRACT FACTORY, one denes a method for each product manufactured. Thus, the
GoF C++ example comprises a MazeFactory class containing MakeMaze, MakeWall,
MakeRoom, and MakeDoor virtual member functions with default implementations
that return pointers to new product instances. Because their MazeFactory was not

203

204

Factory Patterns

abstract, it served as both the ABSTRACT FACTORY and the concrete factory. The
GoF dene two concrete MazeFactory subclasses: EnhantedMaze and BombedMaze.
These overload the aforementioned virtual member functions to produce spellbound
doors and boobytrapped walls, respectively.
In most designs, an ABSTRACT FACTORY contains one or more FACTORY METHOD
implementations, the intent of which the GoF dene as follows:
Dene an interface for creating an object, but let subclasses decide which class to
instantiate. Factory method lets a class defer instantiation to subclasses.

In MazeFactory, the MakeMaze, MakeWall, MakeRoom, and MakeDoor productmanufacturing methods are FACTORY METHODS. Whereas the MazeFactory base class
denes default implementations for these procedures, subclasses can override that
behavior to provide more specic implementations.
The GoF present an alternative based on parameterizing the FACTORY METHOD.
In this case, one denes a single FACTORY METHOD for the entire product family.
The value of a parameter passed to the FACTORY METHOD determines which family
member it constructs and returns.
By analogy with the MazeFactory, an ABSTRACT FACTORY that could support the
Burgers equation client code of Section 9.1 might construct a product family in which
each member abstracts one aspect of a 1D scalar eld. These might include Boundary, Differentiator, and Field classes. The Boundary abstraction could store values
associated with boundary conditions of various types. The Differentiator could provide methods for computing derivatives based on various numerical approximation
schemes. The Field class could store values internal to the problem domain and aggregate a Boundary object and a Differentiator object. Alternatively, all of these
classes could be abstract with default implementations of their methods or with
no method implementations, leaving to subclasses the responsibility for providing
implementations.
By analogy with the EnchantedFactory and BombedFactory, one might dene
additional product families by subclassing each of the classes described in the previous paragraph. For example, a Periodic6thOrder subclass of Field might aggregate
a PeriodicBoundary object and a PeriodicPade6th object. The PeriodicBoundary
subclass of Boundary might or might not store data, considering that the boundary
values are redundant in the periodic case. The PeriodicPade6th subclass of Differentiator could provide algorithms for a sixth-order-accurate Pad nite difference
scheme as described in Appendix A.
Figure 9.1 depicts a minimal class model for supporting the Burgers equation
client code from Section 9.1. The ABSTRACT FACTORY FieldFactory in that gure
publishes a FACTORY METHOD create(). The concrete Periodic6thFactory implements
the FieldFactory interface and therefore provides a concrete create() method. That
method constructs a Periodic6thOrder object, but returns a Field reference, freeing
the client code to manipulate only this reference without any knowledge of the class
of the actual object returned. The greatest benet of this freedom accrues from
the ability to switch between Periodic6thOrder and another Field implementation
without changing the client code. An exercise at the end of this chapter explores this
possibility.

9.2 The Solution

FieldFactory
Field

+create()
returns

Periodic6thFactory

Periodic6thOrder
constructs

+create()

Figure 9.1. UML class diagram of an ABSTRACT FACTORY implementation.

Periodic6thOrder constructs a Field product family that has only one member, Periodic6thOrder, under the assumption that any additional dependencies on
other products are either eliminated by subsuming the necessary logic into Periodic6thOrder or delegated to Periodic6thOrder for satisfying internally. The code
examples in Sections 9.2.19.2.2 take the former approach. An example at the end
of this chapter asks reader to design and implement the second approach.
Figure 9.1 debuts the UML diagrammatic technique for adorning relationships.
Adornments indicate the nature of adjacent relationships. The arrows indicate the
implied order for reading the adornment: a Periodic6thFactory object constructs a
Periodic6thOrder object and returns a Field reference.

9.2.1 Fortran Implementation

Figure 9.2 presents a Fortran implementation of Figure 9.1. Figure 9.2(a) shows
a main program that time advances the Burgers equation via 2nd-order RungeKutta, adapting the time step continually to maintain stability. Each module in
Figure 9.2(b)(e) corresponds to one class in the UML diagram of Figure 9.1.

1
2
3
4
5
6
7
8
9
10
11
12
13

Figure 9.2(a): main.f03

module initializer
use kind_parameters ,only : rkind
implicit none
contains
real(rkind) pure function u_initial(x)
real(rkind) ,intent(in) :: x
u_initial = 10._rkind*sin(x)
end function
real(rkind) pure function zero(x)
real(rkind) ,intent(in) :: x
zero = 0.
end function
end module

14
15
16
17
18
19

program main
use field_module ,only : field,initial_field
use field_factory_module ,only : field_factory
use periodic_6th_factory_module ,only : periodic_6th_factory
use kind_parameters ,only : rkind, ikind

205

206

Factory Patterns
20
21
22
23
24
25
26

use initializer ,only : u_initial,zero

implicit none
class(field) ,pointer :: u,half_uu,u_half
class(field_factory), allocatable :: field_creator
real(rkind) :: dt,half=0.5,t=0.,t_final=0.6,nu=1.
procedure(initial_field) ,pointer :: initial
integer, parameter :: grid_resolution = 16

27
28
29
30
31
32
33

allocate (periodic_6th_factory :: field_creator)

initial => u_initial
u => field_creator%create(initial,grid_resolution)
initial => zero
half_uu => field_creator%create(initial,grid_resolution)
u_half => field_creator%create(initial,grid_resolution)

34
35
36
37
38
39
40
41
42
43
44
45

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22

do while (t<t_final) ! 2nd-order Runge-Kutta:

dt = u%runge_kutta_2nd_step(nu ,grid_resolution)
half_uu = u*u*half
u_half = u + (u%xx()*nu - half_uu%x())*dt*half ! first substep
half_uu = u_half*u_half*half
u = u + (u_half%xx()*nu - half_uu%x())*dt ! second substep
t = t + dt
end do
print *,u at t=,t
call u%output()
end program
Figure 9.2(b): field.f03
module field_module
use kind_parameters ,only : rkind, ikind
implicit none
private
public :: field
public :: initial_field
type ,abstract :: field
contains
procedure(field_op_field) ,deferred :: add
procedure(field_eq_field) ,deferred :: assign
procedure(field_op_field) ,deferred :: subtract
procedure(field_op_field) ,deferred :: multiply_field
procedure(field_op_real) ,deferred :: multiply_real
procedure(real_to_real) ,deferred :: runge_kutta_2nd_step
procedure(derivative) ,deferred :: x ! 1st derivative
procedure(derivative) ,deferred :: xx ! 2nd derivative
procedure(output_interface) ,deferred :: output
generic :: operator(+)
=> add
generic :: operator(-)
=> subtract
generic :: operator(*)
=> multiply_real,multiply_field
generic :: assignment(=) => assign
end type

23
24
25
26
27

abstract interface
real(rkind) pure function initial_field(x)
import :: rkind
real(rkind) ,intent(in) :: x

9.2 The Solution

28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19

end function
function field_op_field(lhs,rhs)
import :: field
class(field) ,intent(in) :: lhs,rhs
class(field) ,allocatable :: field_op_field
end function
function field_op_real(lhs,rhs)
import :: field,rkind
class(field) ,intent(in) :: lhs
real(rkind) ,intent(in) :: rhs
class(field) ,allocatable :: field_op_real
end function
real(rkind) function real_to_real(this,nu,grid_resolution)
import :: field,rkind,ikind
class(field) ,intent(in) :: this
real(rkind) ,intent(in) :: nu
integer(ikind),intent(in) :: grid_resolution
end function
function derivative(this)
import :: field
class(field) ,intent(in) :: this
class(field) ,allocatable :: derivative
end function
subroutine field_eq_field(lhs,rhs)
import :: field
class(field) ,intent(inout) :: lhs
class(field) ,intent(in) :: rhs
end subroutine
subroutine output_interface(this)
import :: field
class(field) ,intent(in) :: this
end subroutine
end interface
end module
Figure 9.2(c): periodic_6th_order.f03
module periodic_6th_order_module
use field_module ,only : field,initial_field
use kind_parameters ,only : rkind,ikind
use matrix_module
,only : matrix, new_matrix
implicit none
private
public :: periodic_6th_order
type ,extends(field) :: periodic_6th_order
private
real(rkind) ,dimension(:) ,allocatable :: f
contains
procedure :: add => total
procedure :: assign => copy
procedure :: subtract => difference
procedure :: multiply_field => product
procedure :: multiply_real => multiple
procedure :: runge_kutta_2nd_step => rk2_dt
procedure :: x => df_dx
! 1st derivative w.r.t. x
procedure :: xx => d2f_dx2 ! 2nd derivative w.r.t. x

207

208

Factory Patterns
20
21

procedure :: output
end type

22
23

real(rkind) ,dimension(:) ,allocatable :: x_node

24
25
26
27

interface periodic_6th_order
procedure constructor
end interface

28
29
30
31
32
33

real(rkind) ,parameter :: pi=acos(-1._rkind)

real(rkind) ,parameter :: first_coeff_6th(5)= (/1.0_rkind/3.0_rkind,&
0.0_rkind,14.0_rkind/9.0_rkind,1.0_rkind/9.0_rkind,0.0_rkind/)
real(rkind) ,parameter :: second_coeff_6th(5)=(/2.0_rkind/11.0_rkind,&
0.0_rkind,12.0_rkind/11.0_rkind,3.0_rkind/11.0_rkind,0.0_rkind/)

34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53

contains
function constructor(initial,grid_resolution)
type(periodic_6th_order) ,pointer :: constructor
procedure(initial_field) ,pointer :: initial
integer(ikind) ,intent(in) :: grid_resolution
integer :: i
allocate(constructor)
allocate(constructor%f(grid_resolution))
if (.not. allocated(x_node)) x_node = grid()
forall (i=1:size(x_node)) constructor%f(i)=initial(x_node(i))
contains
pure function grid()
integer(ikind) :: i
real(rkind) ,dimension(:) ,allocatable :: grid
allocate(grid(grid_resolution))
forall(i=1:grid_resolution) &
grid(i) = 2.*pi*real(i-1,rkind)/real(grid_resolution,rkind)
end function
end function

54
55
56
57
58
59
60
61
62
63
64
65

real(rkind) function rk2_dt(this,nu, grid_resolution)

class(periodic_6th_order) ,intent(in) :: this
real(rkind) ,intent(in) :: nu
integer(ikind) ,intent(in) :: grid_resolution
real(rkind)
:: dx, CFL, k_max
dx=2.0*pi/grid_resolution
k_max=grid_resolution/2.0_rkind
CFL=2.0/(24.0*(1-cos(k_max*dx))/11.0/(1.0+4.0/11.0*cos(k_max*dx))+ &
3.0*(1.0-cos(2.0*k_max*dx))/22.0/(1.0+4.0/11.0*cos(k_max*dx)))
rk2_dt = CFL*dx**2/nu
end function

66
67
68
69
70
71
72
73
74

function total(lhs,rhs)
class(periodic_6th_order) ,intent(in) :: lhs
class(field) ,intent(in) :: rhs
class(field) ,allocatable :: total
type(periodic_6th_order) ,allocatable :: local_total
select type(rhs)
class is (periodic_6th_order)
allocate(local_total)

9.2 The Solution

75
76
77
78
79
80

local_total%f = lhs%f + rhs%f

call move_alloc(local_total,total)
class default
stop periodic_6th_order%total: unsupported rhs class.
end select
end function

81
82
83
84
85
86
87
88
89
90
91
92
93
94
95

function difference(lhs,rhs)
class(periodic_6th_order) ,intent(in) :: lhs
class(field) ,intent(in) :: rhs
class(field) ,allocatable :: difference
type(periodic_6th_order) ,allocatable :: local_difference
select type(rhs)
class is (periodic_6th_order)
allocate(local_difference)
local_difference%f = lhs%f - rhs%f
call move_alloc(local_difference,difference)
class default
stop periodic_6th_order%difference: unsupported rhs class.
end select
end function

96
97
98
99
100
101
102
103
104
105
106
107
108
109
110

function product(lhs,rhs)
class(periodic_6th_order) ,intent(in) :: lhs
class(field) ,intent(in) :: rhs
class(field) ,allocatable :: product
type(periodic_6th_order) ,allocatable :: local_product
select type(rhs)
class is (periodic_6th_order)
allocate(local_product)
local_product%f = lhs%f * rhs%f
call move_alloc(local_product,product)
class default
stop periodic_6th_order%product: unsupported rhs class.
end select
end function

111
112
113
114
115
116
117
118
119
120

function multiple(lhs,rhs)
class(periodic_6th_order) ,intent(in) :: lhs
real(rkind) ,intent(in) :: rhs
class(field) ,allocatable :: multiple
type(periodic_6th_order) ,allocatable :: local_multiple
allocate(local_multiple)
local_multiple%f = lhs%f * rhs
call move_alloc(local_multiple,multiple)
end function

121
122
123
124
125
126
127
128
129

subroutine copy(lhs,rhs)
class(field) ,intent(in) :: rhs
class(periodic_6th_order) ,intent(inout) :: lhs
select type(rhs)
class is (periodic_6th_order)
lhs%f = rhs%f
class default
stop periodic_6th_order%copy: unsupported copy class.

209

210

Factory Patterns
130
131

end select
end subroutine

132
133
134
135
136
137
138
139
140
141
142
143

function df_dx(this)
class(periodic_6th_order) ,intent(in) :: this
class(field) ,allocatable :: df_dx
class(matrix),allocatable,save :: lu_matrix
integer(ikind) :: i,nx, x_east, x_west
integer(ikind) :: x_east_plus1,x_east_plus2
integer(ikind) :: x_west_minus1,x_west_minus2
real(rkind) ,dimension(:,:) ,allocatable :: A
real(rkind) ,dimension(:)
,allocatable :: b,coeff
real(rkind) :: dx
class(periodic_6th_order) ,allocatable :: df_dx_local

144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184

nx=size(x_node)
dx=2.*pi/real(nx,rkind)
coeff = first_coeff_6th
if (.NOT. allocated(lu_matrix)) allocate(lu_matrix)
if (.NOT. lu_matrix%is_built()) then
allocate(A(nx,nx))
!__________Initialize coeffecient matrix A _____
A=0.0_rkind
do i=1, nx
x_east = mod(i,nx)+1
x_west = nx-mod(nx+1-i,nx)
if (i==2) then
x_east_plus1=x_east+1; x_west_minus1=nx
x_east_plus2=x_east+2; x_west_minus2=nx-1
else if (i==3) then
x_east_plus1=x_east+1; x_west_minus1=1
x_east_plus2=x_east+2; x_west_minus2=nx
else if (i==nx-1) then
x_east_plus1=1; x_west_minus1=x_west-1
x_east_plus2=2; x_west_minus2=x_west-2
else if (i==nx-2) then
x_east_plus1=nx; x_west_minus1=x_west-1
x_east_plus2=1; x_west_minus2=x_west-2
else
x_east_plus1=x_east+1; x_west_minus1=x_west-1
x_east_plus2=x_east+2; x_west_minus2=x_west-2
end if
A(i,x_west_minus1) =coeff(2)
A(i,x_west)
=coeff(1)
A(i,i)
=1.0_rkind
A(i,x_east)
=coeff(1)
A(i,x_east_plus1) =coeff(2)
end do
lu_matrix=new_matrix(A)
deallocate(A)
end if
allocate(b(nx))
b=0.0
do i=1,nx
x_east = mod(i,nx)+1

9.2 The Solution

185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201

x_west = nx-mod(nx+1-i,nx)
if (i==2) then
x_east_plus1=x_east+1; x_west_minus1=nx
x_east_plus2=x_east+2; x_west_minus2=nx-1
else if (i==3) then
x_east_plus1=x_east+1; x_west_minus1=1
x_east_plus2=x_east+2; x_west_minus2=nx
else if (i==nx-1) then
x_east_plus1=1; x_west_minus1=x_west-1
x_east_plus2=2; x_west_minus2=x_west-2
else if (i==nx-2) then
x_east_plus1=nx; x_west_minus1=x_west-1
x_east_plus2=1; x_west_minus2=x_west-2
else
x_east_plus1=x_east+1; x_west_minus1=x_west-1
x_east_plus2=x_east+2; x_west_minus2=x_west-2
end if

202
203
204
205
206
207
208
209
210

b(i)=(0.25*coeff(4)*(this%f(x_east_plus1)-this%f(x_west_minus1))+&
0.5*coeff(3)*(this%f(x_east)-this%f(x_west))+ &
coeff(5)/6.0*(this%f(x_east_plus2)-this%f(x_west_minus2)))/dx
end do
allocate(df_dx_local)
df_dx_local%f=lu_matrix .inverseTimes. b
call move_alloc(df_dx_local, df_dx)
end function

211
212
213
214
215
216
217
218
219
220
221
222

function d2f_dx2(this)
class(periodic_6th_order) ,intent(in)
:: this
class(field) ,allocatable :: d2f_dx2
class(matrix),allocatable, save
:: lu_matrix
integer(ikind) :: i,nx,x_east,x_west
integer(ikind) :: x_east_plus1,x_east_plus2
integer(ikind) :: x_west_minus1,x_west_minus2
real(rkind) ,dimension(:,:) ,allocatable :: A
real(rkind) ,dimension(:)
,allocatable :: coeff,b
real(rkind)
:: dx
class(periodic_6th_order) ,allocatable :: d2f_dx2_local

223
224
225
226
227
228
229

nx=size(this%f)
dx=2.*pi/real(nx,rkind)
coeff = second_coeff_6th
if (.NOT. allocated(lu_matrix)) allocate(lu_matrix)
if (.NOT. lu_matrix%is_built()) then
allocate(A(nx,nx))

230
231
232
233
234
235
236
237
238
239

!__________Initialize coeffecient matrix A _____

A=0.0_rkind
do i=1, nx
x_east = mod(i,nx)+1
x_west = nx-mod(nx+1-i,nx)
if (i==2) then
x_east_plus1=x_east+1; x_west_minus1=nx
x_east_plus2=x_east+2; x_west_minus2=nx-1
else if (i==3) then

211

212

Factory Patterns
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290

x_east_plus1=x_east+1; x_west_minus1=1
x_east_plus2=x_east+2; x_west_minus2=nx
else if (i==nx-1) then
x_east_plus1=1; x_west_minus1=x_west-1
x_east_plus2=2; x_west_minus2=x_west-2
else if (i==nx-2) then
x_east_plus1=nx; x_west_minus1=x_west-1
x_east_plus2=1; x_west_minus2=x_west-2
else
x_east_plus1=x_east+1; x_west_minus1=x_west-1
x_east_plus2=x_east+2; x_west_minus2=x_west-2
end if
A(i,x_west_minus1) =coeff(2)
A(i,x_west)
=coeff(1)
A(i,i)
=1.0_rkind
A(i,x_east)
=coeff(1)
A(i,x_east_plus1) =coeff(2)
end do
lu_matrix=new_matrix(A)
deallocate(A)
end if
allocate(b(nx))
do i=1, nx
x_east = mod(i,nx)+1
x_west = nx-mod(nx+1-i,nx)
if (i==2) then
x_east_plus1=x_east+1; x_west_minus1=nx
x_east_plus2=x_east+2; x_west_minus2=nx-1
else if (i==3) then
x_east_plus1=x_east+1; x_west_minus1=1
x_east_plus2=x_east+2; x_west_minus2=nx
else if (i==nx-1) then
x_east_plus1=1; x_west_minus1=x_west-1
x_east_plus2=2; x_west_minus2=x_west-2
else if (i==nx-2) then
x_east_plus1=nx; x_west_minus1=x_west-1
x_east_plus2=1; x_west_minus2=x_west-2
else
x_east_plus1=x_east+1; x_west_minus1=x_west-1
x_east_plus2=x_east+2; x_west_minus2=x_west-2
end if
b(i)=(0.25*coeff(4)* &
(this%f(x_east_plus1)-2.0*this%f(i)+this%f(x_west_minus1))+ &
coeff(3)*(this%f(x_east)-2.0*this%f(i)+this%f(x_west))+ &
coeff(5)/9.0* &
(this%f(x_east_plus2)-this%f(i)+this%f(x_west_minus2)))/dx**2
end do
allocate(d2f_dx2_local)
d2f_dx2_local%f=lu_matrix .inverseTimes. b
call move_alloc(d2f_dx2_local, d2f_dx2)
end function

291
292
293
294

subroutine output(this)
class(periodic_6th_order) ,intent(in) :: this
integer(ikind) :: i

9.2 The Solution

295
296
297
298

do i=1,size(x_node)
print *, x_node(i), this%f(i)
end do
end subroutine

299
300

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21

1
2

end module
Figure 9.2(d): field_factory.f03
module field_factory_module
use field_module ,only : field,initial_field
implicit none
private
public :: field_factory
type, abstract :: field_factory
contains
procedure(create_interface), deferred :: create
end type
abstract interface
function create_interface(this,initial,grid_resolution)
use kind_parameters ,only : ikind
import :: field, field_factory ,initial_field
class(field_factory), intent(in) :: this
class(field) ,pointer :: create_interface
procedure(initial_field) ,pointer :: initial
integer(ikind) ,intent(in) :: grid_resolution
end function
end interface
end module
Figure 9.2(e): periodic_6th_factory.f03
module periodic_6th_factory_module
use field_factory_module! , only : field_factory
use field_module ! ,only : field,initial_field
implicit none
private
public :: periodic_6th_factory
type, extends(field_factory) :: periodic_6th_factory
contains
procedure :: create=>new_periodic_6th_order
end type
contains
function new_periodic_6th_order(this,initial,grid_resolution)
use periodic_6th_order_module ,only : periodic_6th_order
use kind_parameters ,only : ikind
class(periodic_6th_factory), intent(in) :: this
class(field) ,pointer :: new_periodic_6th_order
procedure(initial_field) ,pointer :: initial
integer(ikind) ,intent(in) :: grid_resolution
new_periodic_6th_order=> periodic_6th_order(initial,grid_resolution)
end function
end module
Figure 9.2(f): matrix.f03
module matrix_module
use kind_parameters ,only : rkind, ikind

213

214

Factory Patterns
3
4
5
6
7
8
9
10
11
12
13
14
15
16

implicit none
private
public :: matrix
public :: new_matrix
type :: matrix
integer(ikind), allocatable :: pivot(:)
real(rkind), allocatable :: lu(:,:)
contains
procedure :: back_substitute
procedure :: is_built
procedure :: matrix_eq_matrix
generic
:: operator(.inverseTimes.) => back_substitute
generic
:: assignment(=) => matrix_eq_matrix
end type

17
18
19
20

interface new_matrix
procedure constructor
end interface

21
22
23
24
25
26

contains
logical function is_built(this)
class(matrix),intent(in)
:: this
is_built = allocated (this%pivot)
end function

27
28
29
30
31

function constructor(A) result(new_matrix)

class(matrix) ,allocatable
:: new_matrix
real(rkind), intent(in) ,dimension(:,:) :: A
integer
:: n,info

32
33
34
35
36
37
38

n=size(A,1)
allocate(new_matrix)
allocate(new_matrix%pivot(n), new_matrix%lu(n,n))
new_matrix%lu=A
call dgetrf(n,n,new_matrix%lu,n,new_matrix%pivot,info)
end function

39
40
41
42
43
44
45

subroutine matrix_eq_matrix(lhs,rhs)
class(matrix) ,intent(in) :: rhs
class(matrix) ,intent(out) :: lhs
lhs%pivot = rhs%pivot
lhs%lu
= rhs%lu
end subroutine

46
47
48
49
50
51
52
53
54

function back_substitute(this,b) result(x)

class(matrix) ,intent(in)
real(rkind) ,intent(in) ,dimension(:)
real(rkind) ,allocatable ,dimension(:)
real(rkind) ,allocatable ,dimension(:,:)
real(rkind) ,allocatable ,dimension(:)
integer(ikind)
real(rkind)

55
56
57

n=size(this%lu,1)
allocate(lower(n,n), upper(n,n))

:: this
:: b
:: x, temp_x
:: lower, upper
:: local_b
:: n,i,j
:: temp

9.2 The Solution

58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100

lower=0.0_rkind
upper=0.0_rkind
do i=1,n
do j=i,n
upper(i,j)=this%lu(i,j)
end do
end do
do i=1,n
lower(i,i)=1.0_rkind
end do
do i=2,n
do j=1,i-1
lower(i,j)=this%lu(i,j)
end do
end do
allocate(local_b(n))
local_b=b
do i=1,n
if (this%pivot(i)/=i) then
temp=local_b(i)
local_b(i)=local_b(this%pivot(i))
local_b(this%pivot(i))=temp
end if
end do
allocate(temp_x(n))
temp_x(1)=local_b(1)
do i=2,n
temp_x(i)=local_b(i)-sum(lower(i,1:i-1)*temp_x(1:i-1))
end do
allocate(x(n))
x(n)=temp_x(n)/upper(n,n)
do i=n-1,1,-1
x(i)=(temp_x(i)-sum(upper(i,i+1:n)*x(i+1:n)))/upper(i,i)
end do
do i=n,1,-1
if (this%pivot(i)/=i) then
temp=x(i)
x(i)=x(this%pivot(i))
x(this%pivot(i))=temp
end if
end do
end function
end module
Figure 9.2. Fortran demonstration of an abstract factory: (a) main client, (b) abstract eld
class, (c) periodic 6th-order Pad eld, (d) abstract eld factory, (e) periodic 6th-order Pad
factory, (f) matrix module.

Fortran constructs appearing for the rst time in Figure 9.2 include typed allocation and procedure pointers. Line 28 in the main program of Figure 9.2(a)
demonstrates how to allocate the dynamic type of a field_factory instance to
be its child class periodic_6th_factory. That allocation yields a concrete object to
which results can be assigned and on which type-bound procedures can be invoked.

215

216

Factory Patterns

Line 25 in main demonstrates the declaration of a procedure pointer, initial

that conforms to the abstract interface, initial_field, published by the field class
at line 6 and subsequently prototyped at line 2528 of Figure 9.2(b). Lines 29 and 31
in main show pointer assignments of initial to functions intended for initializing u
and for initially defaulting other elds to zero, respectively. Lines 30 and 3233 pass
initial to the eld factory.
The matrix module shown in Figure 9.2(f) is used to solve the 6th-order rst
derivatives and second derivatives from Equations A.57 and A.58 using LU decomposition method. LU decomposition method is an efcient procedure for solving
a system of linear algebraic equations Ax = b, the details of LU decomposition
method is introduced in A.2.2. Line 37 in matrix module is calling the function
dgetrf from Linear Algebra PACKage (LAPACK, http://www.netlib.org/lapack),
which performs the LU factorization of a general matrix.

9.2.2 C++ Implementation

Figure 9.3 presents a C++ implementation of Figure 9.1. Figure 9.3(a) shows a main
program that time advances the Burgers equation via 2nd-order Runge-Kutta, adapting the time step continually to maintain stability. The class structure implemented
in Figure 9.3(b)(h) mirrors that in the Fortran implementation of Figure 9.2.

1
2
3
4
5

#include
#include
#include
#include
#include

Figure 9.3(a): main.cpp

"field.h"
"field_factory.h"
"periodic_6th_factory.h"
"fmt.h"
"initializer.h"

6
7
8
9
10

int main ()
{
typedef field::ptr_t field_ptr_t;
typedef field_factory::ptr_t field_factory_ptr;

11
12
13
14
15

const
const
const
const

real_t t_final = 0.6;

real_t half = 0.5;
real_t nu = 1.0;
int grid_resolution = 128;

16
17

real_t t = 0.0;

18
19
20

field_factory_ptr field_creator =
field_factory_ptr (new periodic_6th_factory());

21
22

field_ptr_t u

23

= field_creator->create(u_initial,
grid_resolution);

24
25
26

field_ptr_t u_half = field_creator->create(zero,

grid_resolution);

27
28

real_t dt;

9.2 The Solution

29

while (t < t_final) {

// use 2nd-order Runge-Kutta
dt = u->runge_kutta_2nd_step(nu, grid_resolution);

30
31
32
33

// first substep
u_half = u + (u->xx()*nu - (u*u*half)->x())*dt*half;

34
35
36

// second substep
u += (u_half->xx()*nu - (u_half*u_half*half)->x())*dt;

37
38
39
40

41

t += dt;

42

std::cout << " u at t = " << fmt(t) << std::endl;

u->output();

43
44
45
46
47

1
2

return 0;

#ifndef _FIELD_H_
#define _FIELD_H_ 1

Figure 9.3(b): field.h

3
4
5

#include "RefBase.h"
#include "globals.h"

6
7
8
9

class field : public virtual RefBase {

public:
typedef Ref<field> ptr_t;

10

virtual field() {};

11
12

virtual
virtual
virtual
virtual

13
14
15
16

ptr_t
ptr_t
ptr_t
ptr_t

operator+=(ptr_t)
operator-(ptr_t)
operator*(real_t)
operator*(ptr_t)

= 0;
const = 0;
const = 0;
const = 0;

17

virtual ptr_t x()

const = 0;
virtual ptr_t xx()
const = 0;
virtual real_t runge_kutta_2nd_step(real_t,
int)
const = 0;

18
19
20
21
22
23
24
25
26
27

};

virtual ptr_t clone()

virtual void output()
protected:
field() {};

const = 0;
const = 0;

28
29
30
31
32
33
34

//
// define operators to be used for field ptr_t
//
inline field::ptr_t operator+= (field::ptr_t a, field::ptr_t b) {
*a += b;

217

218

Factory Patterns
35
36

return a;

37
38
39

inline field::ptr_t operator+ (field::ptr_t a, field::ptr_t b) {

field::ptr_t tmp = a->clone();

40
41
42
43

*tmp += b;
return tmp;

44
45
46

inline field::ptr_t operator- (field::ptr_t a, field::ptr_t b) {

field::ptr_t tmp = a->clone();

47
48
49
50

tmp = *tmp - b;
return tmp;

51
52
53
54

inline field::ptr_t operator*(field::ptr_t a, field::ptr_t b) {

field::ptr_t tmp = a->clone();

55
56
57
58

tmp = *tmp * b;
return tmp;

59
60
61

inline field::ptr_t operator*(field::ptr_t a, real_t b) {

field::ptr_t tmp = a->clone();

62
63
64
65

tmp = *tmp * b;
return tmp;

66
67

1
2

#endif
Figure 9.3(c): field_factory.h
#ifndef _FIELD_FACTORY_H_
#define _FIELD_FACTORY_H_ 1

3
4
5

#include "RefBase.h"
#include "field.h"

6
7
8
9

class field_factory : public virtual RefBase {

public:
typedef Ref<field_factory> ptr_t;

10
11
12
13
14
15
16

};

virtual field_factory() {};

virtual field::ptr_t create(real_t (*initial) (real_t x),
int grid_resolution) = 0;
protected:
field_factory() {};

17
18

#endif

9.2 The Solution

1
2

Figure 9.3(e): periodic_6th_factory.h

#ifndef PERIODIC_6TH_FACTORY_H_
#define PERIODIC_6TH_FACTORY_H_

3
4
5

#include "field.h"
#include "field_factory.h"

6
7
8
9
10

class periodic_6th_factory : public field_factory {

public:
periodic_6th_factory ();
virtual periodic_6th_factory();

11
12
13
14
15

1
2

};
#endif

virtual field::ptr_t create(real_t (*initial) (real_t x),

int grid_resolution);

Figure 9.3(f): periodic_6th_factory.cpp

#include "periodic_6th_factory.h"
#include "periodic_6th_order.h"

3
4
5

periodic_6th_factory::periodic_6th_factory ()
{ }

6
7
8

periodic_6th_factory::periodic_6th_factory ()
{ }

9
10
11
12
13
14
15
16

1
2

field::ptr_t periodic_6th_factory::create(real_t (*initial) (real_t x),

int grid_resolution)
{
return field::ptr_t
(new periodic_6th_order(initial, grid_resolution));
}
Figure 9.3(g): periodic_6th_order.h
#ifndef _PERIODIC_6TH_ORDER_H_
#define _PERIODIC_6TH_ORDER_H_ 1

3
4
5
6

#include <vector>
#include "field.h"
#include "globals.h"

7
8
9
10

class periodic_6th_order : public field {

public:
typedef field::ptr_t ptr_t;

11
12
13

periodic_6th_order(real_t (*initial) (real_t x),

int grid_resolution);

14
15
16

periodic_6th_order(const crd_t& other);

virtual periodic_6th_order();

17
18
19

virtual ptr_t operator+=(ptr_t);

virtual ptr_t operator-(ptr_t) const;

219

220

Factory Patterns
virtual ptr_t operator*(real_t) const;
virtual ptr_t operator*(ptr_t) const;

20
21
22

virtual ptr_t x()

virtual ptr_t xx()
virtual real_t runge_kutta_2nd_step(real_t nu,
int grid_resolution)

23
24
25
26

const;
const;
const;

27

virtual ptr_t clone()

virtual void output()

28
29

const;
const;

30

static void set_grid(int num_grid_pts);

static const crd_t & get_grid();
static int get_grid_size();

31
32
33
34
35
36
37
38

};

private:
crd_t f_;
static crd_t x_node_;

39
40

1
2
3
4
5
6

#endif

#include
#include
#include
#include
#include
#include

Figure 9.3(h): periodic_6th_order.cpp

<cmath>
<exception>
"periodic_6th_order.h"
"fmt.h"
"mat.h"
"gaussian_elimination.h"

7
8
9
10

struct periodic_6th_order_error : public std::exception {

virtual periodic_6th_order_error() throw () {};
};

11
12
13

const real_t pi = 3.14159265f;

14
15
16

const real_t first_coeff_6th[5] =

{1.0/3.0, 0.0, 14.0/9.0, 1.0/9.0, 0.0};

17
18
19

const real_t second_coeff_6th[5] =

{2.0/11.0, 0.0, 12.0/11.0, 3.0/11.0, 0.0};

20
21
22

typedef field::ptr_t ptr_t;

23
24

crd_t periodic_6th_order::x_node_ = crd_t();

25
26
27
28
29
30
31
32

void periodic_6th_order::set_grid(int grid_resolution)

{
if (x_node_.empty()) {
for (int i=0; i< grid_resolution;++i) {
x_node_.push_back (2.0*pi*(i)/real_t(grid_resolution));
}

9.2 The Solution

33
34

35
36
37

int periodic_6th_order::get_grid_size() const {

return (x_node_.size());

38
39

40
41
42
43

const crd_t & periodic_6th_order::get_grid () const {

return x_node_;
}

44
45
46
47
48

periodic_6th_order::periodic_6th_order(real_t (*initial) (real_t x),

int grid_resolution) {
set_grid(grid_resolution);

49

f_ = get_grid();

50
51
52
53
54
55

for (int i = 0; i < f_.size(); ++i) {

f_.at(i) = initial(f_.at(i));
}

56
57
58
59

periodic_6th_order::periodic_6th_order(const crd_t & other)

: f_(other)
{}

60
61
62
63

periodic_6th_order::periodic_6th_order() {
}

64
65
66

ptr_t periodic_6th_order::operator+=(ptr_t rhs) {

67

Ref<periodic_6th_order> other = cast<periodic_6th_order> (rhs);

68
69

if ((other == NULL) || (f_.size() != other->f_.size())) {

std::cerr << "periodic_6th_order::operator+= " <<
"rhs is invalid" << std::endl;

70
71
72
73
74

75

throw periodic_6th_order_error();

76

for (int i = 0; i < f_.size(); ++i) {

f_.at(i) += other->f_.at(i);
}

77
78
79
80
81
82

return ptr_t(this);

83
84
85
86
87

ptr_t periodic_6th_order::operator-(ptr_t rhs) const {

Ref<periodic_6th_order> other = cast<periodic_6th_order> (rhs);

221

222

Factory Patterns
if ((other == NULL) || (f_.size() != other->f_.size())) {
std::cerr << "periodic_6th_order::operator- " <<
"rhs is invalid" << std::endl;

88
89
90
91
92

93

throw periodic_6th_order_error();

94

Ref<periodic_6th_order> result =
Ref<periodic_6th_order>(new periodic_6th_order(*this));

95
96
97

for (int i = 0; i < f_.size(); ++i) {

result->f_.at(i) -= other->f_.at(i);
}

98
99
100
101
102
103

return result;

104
105
106
107
108

ptr_t periodic_6th_order::operator*(real_t rhs) const {

Ref<periodic_6th_order> result =
Ref<periodic_6th_order>(new periodic_6th_order(*this));

109

for (int i = 0; i < f_.size(); ++i) {

result->f_.at(i) *= rhs;
}

110
111
112
113
114
115

return result;

116
117
118
119

ptr_t periodic_6th_order::operator*(ptr_t rhs) const {

Ref<periodic_6th_order> other = cast<periodic_6th_order> (rhs);

120

if ((other == NULL) || (f_.size() != other->f_.size())) {

std::cerr << "periodic_6th_order::operator* " <<
"rhs is invalid" << std::endl;

121
122
123
124
125

126

throw periodic_6th_order_error();

127

Ref<periodic_6th_order> result =
Ref<periodic_6th_order>(new periodic_6th_order(*this));

128
129
130

for (int i = 0; i < f_.size(); ++i) {

result->f_.at(i) *= other->f_.at(i);
}

131
132
133
134
135
136

return result;

137
138
139
140
141
142

ptr_t periodic_6th_order::clone() const {

return ptr_t (new periodic_6th_order(*this));
}

9.2 The Solution

143
144
145
146
147
148
149

void periodic_6th_order::output() const {

for (int i = 0; i < f_.size(); ++i) {
std::cout << fmt(x_node_.at(i)) << " " << fmt(f_.at(i))
<< std::endl;
}
}

150
151
152
153

ptr_t periodic_6th_order:: x() const {

int nx = get_grid_size();
real_t dx = 2.0 * pi /real_t(nx);

154
155
156
157

// __Initialize coeffecient matrix A and right handside b__

mat_t A;
A.clear_resize(nx, nx, 0.0);

158
159

crd_t b = crd_t(nx, 0.0);

160
161

crd_t coeff;

162
163
164

for (int i = 0; i < 5; ++i)

coeff.push_back(first_coeff_6th[i]);

165
166
167

int x_east, x_west, x_east_plus1,

x_east_plus2,x_west_minus1,x_west_minus2;

168
169
170
171

for (int i = 0; i < nx; ++i) {

x_east = (i+1)%nx;
x_west = nx -1 - (nx-i)%nx;

172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197

if (i == 1) {
x_east_plus1=x_east+1;
x_east_plus2=x_east+2;
x_west_minus1=nx-1;
x_west_minus2=nx-2;
}
else if (i == 2) {
x_east_plus1=x_east+1;
x_east_plus2=x_east+2;
x_west_minus1=0;
x_west_minus2=nx-1;
}
else if (i == nx-2) {
x_east_plus1=0;
x_east_plus2=1;
x_west_minus1=x_west-1;
x_west_minus2=x_west-2;
}
else if (i == nx-3) {
x_east_plus1=nx-1;
x_east_plus2=0;
x_west_minus1=x_west-1;
x_west_minus2=x_west-2;
}
else {

223

224

Factory Patterns
198
199
200
201

202

x_east_plus1=x_east+1;
x_east_plus2=x_east+2;
x_west_minus1=x_west-1;
x_west_minus2=x_west-2;

203

A(i,x_west_minus1)
A(i,x_west)
A(i,i)
A(i,x_east)
A(i,x_east_plus1)

204
205
206
207
208

=coeff.at(1);
=coeff.at(0);
=1.0;
=coeff.at(0);
=coeff.at(1);

209
210
211
212
213
214

215

b.at(i) = (0.25*coeff.at(3)*
(f_.at(x_east_plus1)-f_.at(x_west_minus1))+
0.5*coeff.at(2)*(f_.at(x_east) - f_.at(x_west)) +
coeff.at(4)/6.0 * (f_.at(x_east_plus2)f_.at(x_west_minus2)))/dx;

216
217
218

return ptr_t(new periodic_6th_order(gaussian_elimination(A,b)));

219
220
221
222
223

ptr_t periodic_6th_order:: xx() const {

int nx = get_grid_size();
real_t dx = 2.0 * pi /real_t(nx);

224
225
226
227

// __Initialize coeffecient matrix A and right handside b__

mat_t A;
A.clear_resize(nx, nx, 0.0);

228
229

crd_t b = crd_t(nx, 0.0);

230
231

crd_t coeff;

232
233
234

for (int i = 0; i < 5; ++i)

coeff.push_back(second_coeff_6th[i]);

235
236
237

int x_east, x_west, x_east_plus1,

x_east_plus2,x_west_minus1,x_west_minus2;

238
239
240
241

for (int i = 0; i < nx; ++i) {

x_east = (i+1)%nx;
x_west = nx -1 - (nx-i)%nx;

242
243
244
245
246
247
248
249
250
251
252

if (i == 1) {
x_east_plus1=x_east+1;
x_east_plus2=x_east+2;
x_west_minus1=nx-1;
x_west_minus2=nx-2;
}
else if (i == 2) {
x_east_plus1=x_east+1;
x_east_plus2=x_east+2;
x_west_minus1=0;

9.2 The Solution

x_west_minus2=nx-1;
}
else if (i == nx-2) {
x_east_plus1=0;
x_east_plus2=1;
x_west_minus1=x_west-1;
x_west_minus2=x_west-2;
}
else if (i == nx-3) {
x_east_plus1=nx-1;
x_east_plus2=0;
x_west_minus1=x_west-1;
x_west_minus2=x_west-2;
}
else {
x_east_plus1=x_east+1;
x_east_plus2=x_east+2;
x_west_minus1=x_west-1;
x_west_minus2=x_west-2;
}

253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273

A(i,x_west_minus1)
A(i,x_west)
A(i,i)
A(i,x_east)
A(i,x_east_plus1)

274
275
276
277
278

=coeff.at(1);
=coeff.at(0);
=1.0;
=coeff.at(0);
=coeff.at(1);

279
280
281
282
283
284

285

b.at(i) = (0.25*coeff.at(3)*
(f_.at(x_east_plus1)-2.0*f_.at(i)+f_.at(x_west_minus1))+
coeff.at(2)*(f_.at(x_east)-2.0*f_.at(i)+f_.at(x_west)) +
coeff.at(4)/9.0 * (f_.at(x_east_plus2)f_.at(i)+f_.at(x_west_minus2)))/(dx*dx);

286
287
288

return ptr_t(new periodic_6th_order(gaussian_elimination(A,b)));

289
290
291
292
293

real_t periodic_6th_order:: runge_kutta_2nd_step(real_t nu,

int grid_resolution) const
{
real_t dx, CFL, k_max;

294

dx=2.0*pi/grid_resolution;

295
296

k_max=grid_resolution*0.5;

297
298

CFL=2.0/(24.0*(1-cos(k_max*dx))/11.0/(1.0+4.0/11.0*cos(k_max*dx))+
3.0*(1.0-cos(2.0*k_max*dx))/22.0/(1.0+4.0/11.0*cos(k_max*dx)));
return (CFL*dx*dx/nu);

299
300
301
302
303

1
2

#ifndef INITIALIZER_H_
#define INITIALIZER_H_

Figure 9.3(i): initializer.h

225

226

Factory Patterns
3
4

#include "globals.h"

5
6
7

real_t u_initial (real_t x);

real_t zero
(real_t x);

8
9

#endif

Figure 9.3(j): initializer.cpp

#include "initializer.h"
#include <cmath>

2
3
4
5
6
7

real_t u_initial (real_t x)

{
return (10.0 * sin (x));
}

8
9
10
11
12
13

real_t zero (real_t x)

{
return 0.0;
}
Figure 9.3. Abstract factory demonstration in C++: (a) main client, (b) eld interface, (c) eld
factory header, (d) abstract eld factory, (e) concrete periodic 6th-order Pad eld factory, (f)
abstract eld factory, (g) periodic 6th-order Pad factory, (i) eld initializing routine interfaces,
(j) eld initializing routines.

9.3 The Consequences

The GoF list four consequences of using an ABSTRACT FACTORY: its isolation of concrete classes, its easing of exchanging product families, its promotion of product
consistency, and its potential hindering of extending support to new classes. The
rst two of these reiterate the primary purpose of an ABSTRACT FACTORY. The third
prevents mistakenly mixing products from different product families. Centralizing
product creation and hiding implementations prevents, for example, a sixth-order
Differentiator instance from being used with an eighth-order Pad eld instance.
The fourth consequence crops up generally in all uses of abstract classes as interfaces: Any changes to the interface cascade down through every implementation.
Adding a product to the family necessitates adding corresponding factory methods
to every implementation.
The FACTORY METHOD consequences overlap with those of the ABSTRACT FACTORY in which it plays a role. A FACTORY METHOD accomplishes the aforementioned
concrete product isolation. Additionally, the GoF note that a FACTORY METHOD provides hooks for subclasses and connects parallel class hierarchies. In the rst of
these roles, the FACTORY METHOD dened in a base class can provide a default implementation. Client code that links to the base class retains this fully functional
hook for use when extended implementations are unavailable or unnecessary. In
the second role, the FACTORY METHOD enables one class to delegate responsibilities
to another without exposing implementation details. The parallel class then takes
on some aspects of the client when it needs to construct its own instances of its

9.4 Related Patterns

sister class. It then benets from the same exibility a FACTORY METHOD affords
clients.
Another consequence of this chapters two patterns lies in the resulting proliferation of classes. Were the Burgers equation solver architect willing to expose
Periodic6thOrder to the client, she could eliminate the other three classes. Obviously,
the exibility afforded by not exposing it must be balanced against the complexity
associated with hiding it.
The latter paragraph strikes a theme that recurs throughout this text. Flexibility
often comes at the cost of increased complexity. Engendering exibility in an application requires increasing the sophistication of the infrastructure that supports it.
This proved true in Part I of the text, wherein we argued that the sophistication of
OOP matters most as a software application grows larger. A similar statement holds
for the design patterns in Part II. They prove more worthwhile with increasing scale
in OOP projects. A successful designer allows for growth while consciously avoiding
slicing butter with a machete.
Up to this point, we have emphasized the role of factory patterns in determining
where object construction occurs: inside factories. In the context of private data,
where also determines who. To the extent factories hide concrete classes from clients,
they assume sole responsibility for object construction.
Other creational patterns also manage the how of object construction. The GoF
BUILDER pattern, for example, adds a Director abstraction between the client a
Builder interface. Clients direct the Director to construct a product. The concrete
Builder steps through the construction of the parts required to piece together a
complex object. This isolates the client from even knowing the construction steps,
in contrast to the MazeFactory example in which the client individually invokes
methods to manufacture the mazes constituent parts.
Lastly, the GoF SINGLETON pattern address one aspect of when construction
happens: upon the rst request and only then. A SINGLETON instance ensures that
only one instance of a class gets instantiated. This provides an OOD alternative to
global data, ensuring all clients access the same instance.

9.4 Related Patterns

ABSTRACT FACTORY and FACTORY METHOD represent the smallest of this books patterns. Whereas OBJECT potentially impacts every ADT in a package, ABSTRACT
FACTORY and FACTORY METHOD impact only those for which it proves valuable to
vary the class of the object instantiated at a given point in the code. This could apply
to all classes in a given project or none. Furthermore, these two patterns always play
a supporting role: They exist only to support the instantiation of another class. In
these senses, these patterns sit at the end of, and complete, a long process in multiphysics model construction. That process starts with the universally useful basic
services in OBJECT, progresses through the expression of the mathematical problem
in an ABSTRACT CALCULUS, the enabling of adaptive, semidiscrete time integration via
a STRATEGY, the aggregation of individual submodels into a PUPPETEER, and nally the
exible object construction in an ABSTRACT FACTORY containing one or more FACTORY
METHODS.

227

228

Factory Patterns

In a different sense, the ABSTRACT FACTORY could have a large footprint if a

substantial number of ADTs need the exibility it affords at object construction time.
Unlike some of the other patterns for which the corresponding implementation sits
in one node or one node hierarchy of a class diagram, the ABSTRACT FACTORY and
FACTORY METHOD patterns might be replicated at multiple locations across a package
wherever their services nd useful application.
The role an ABSTRACT FACTORY plays also relates closely to that of a STRATEGY
pattern. One might use a STRATEGY to vary the nite difference schemes discussed
in this chapter instead of using the ABSTRACT FACTORY approach. One distinction
lies in the timing and duration, with an ABSTRACT FACTORY acting only at the time
of instantiation to vary the scheme by choosing the corresponding objects identity;
whereas a STRATEGY acts throughout the duration of the schemes use and thereby
allows for additional variation beyond the point of instantiation.
The viewpoint and expected knowledge of the user harbors a more important
distinction. The STRATEGY client code in Figure 7.2(a), for example, exploits explicit
knowledge of which time-integration scheme is being used. The ABSTRACT FACTORY
code in the current chapter, by contrast, remains agnostic with respect to the concrete
type of field it uses.
The GoF listed several other relationships between this chapters two patterns
and patterns not discussed in this text. The SINGLETON pattern, for example, ensures
that only one instance of a class gets instantiated. Because applications typically
require only a single factory, an ABSTRACT FACTORY can be implemented protably
as a SINGLETON. A TEMPLATE METHOD denes a procedure abstractly by building up
statements (e.g., expressions in an ABSTRACT CALCULUS) from procedures specied in
an interface. Since a TEMPLATE METHOD works only with interfaces, its only option for
object instantiation is via FACTORY METHODS. See (Gamma et al. 1995) for additional
pattern relationships.

EXERCISES
1. Augment the UML class diagram of Figure 9.1 by including interfaces and corresponding implementations of Boundary and Differentiator abstractions. Modify
the Fortran code in Figure 9.2 or the C++ code in Figure 9.3 so that the Periodic6thOrder constructor instantiates and stores Boundary and Differentiator
abstractions.
2. Refactor the code from Exercise 2 so that the Periodic6thOrder stores a pointer
to a DifferentationStrategy using the STRATEGY pattern of Chapter 7. Supply an
additional DifferentiatorStrategy8th eighth-order-accurate Pad nite difference
scheme as dened in Appendix A.

PA R T III

G UMBO S OOP

10 Formal Constraints

Learn the rules so you know how to break them properly.

Tenzin Gyatso, The 14th Dalai Lama

10.1 Why Be Formal?

Whereas Parts I and II focused on canonical examples, Part III marches toward
complete applications, resplendent with runtime considerations. The current chapter
addresses code and compiler correctness. Chapter 11 discusses language interoperability. Chapter 12 addresses scalability and weaves elements of the entire book into
a vision for multiphysics framework design.
Formal methods form an important branch of software engineering that has
apparently been applied to the design of only a small percentage of scientic simulation programs (Bientinesi and van de Geijn 2005; van Engelen and Cats 1997). Two
pillars of formalization are specication and verication that is, specifying mathematically what a program must do and verifying the correctness of an algorithm with
respect to the specication. The numerical aspects of scientic programming are already formal. The mathematical equations one wishes to solve in a given scientic
simulation provide a formal specication, whereas a proof of numerical convergence
provides a formal verication. Hence, formal methods developers often cite a motivation of seeking correctness standards for non-scientic codes as rigorous as those
for scientic codes (Oliveria 1997). This ignores, however, the nonnumerical aspects of scientic programs that could benet from greater rigor. One such aspect
is memory management. The current chapter species formal constraints on memory allocations in the Fortran implementation of the Burgers equation solver from
Chapter 9.
There have been long-standing calls for increased use of formal methods in scientic programming to improve reliability. Stevenson (1997) proclaimed a crying need
for highly reliable system simulation methodology, far beyond what we have now. He
cited formal methods among the strategies that might improve the situation. Part of
his basis was Hattons measurement of an average of 10 statically detectable serious
faults per 1,000 lines among millions of lines of commercially released scientic C and
Fortran 77 code (Hatton 1997). Hatton dened a fault as a misuse of the language
231

232

Formal Constraints

that will very likely fail in some context. Examples include interface inconsistencies
and using uninitialized variables. Hatton suggested that formal methods reduce defect density by a factor of three (cited in J3 Fortran Standards Technical Committee
1998). This chapter shows how code inspired by formal constraints can also detect
compiler faults.
Pace (2004) expressed considerable pessimism about the prospects for adopting formal methods in scientic simulation because of the requisite mathematical
training, which often includes set theory and a predicate calculus. A candidate for
adoption must balance such rigor against ease of use. The Object Constraint Language (OCL) strikes such a balance, facilitating expressing of formal statements
about software models without using mathematical symbols unfamiliar to nonspecialists. OCLs designers intended for it to be understood by people who are not
mathematicians or computer scientists (Warmer and Kleppe 2003).
To attract scientic programmers, any software development strategy must address performance. Fortunately, program specication and verication are part of
the software design rather than the implementation. Thus, they need not impact
run-time performance. However, run-time checking of assertions, a third pillar of formal methods, is part of the implementation (Clarke and Rosenblum 2006; Dahlgren
2007). Assertions are tool-based or language-based mechanisms for gracefully terminating execution whenever programmer-inserted Boolean expressions fail. When
the Boolean expressions are sufciently simple, they occupy a negligible fraction of
the run time compared to the long loops over oating point calculations typical of
scientic software.
A nal factor inuencing adoption of formal methods is the lack of a common approach for describing the structure of traditional scientic codes beyond ow charts.
OCLs incorporation into the UML standard (Warmer and Kleppe 2003) suggests
that newcomers must simultaneously leap two hurdles: learning OCL and learning UML. Fortunately, increasing interest in object-oriented scientic programming
has led to more frequent scientic program structural descriptions resembling UML
class diagrams (Akin 2003; Barton and Nackman 1994), and several scientic programming projects use automated documentation tools that produce class diagrams
(Heroux et al. 2005).
This chapter details how exposure to formal methods inspired systematic runtime assertion checking in a production research code, and how these assertions
enabled the detection of an inscrutable compiler bug. The chapter also explains how
applying OCL constraints to a UML software model forced us to think abstractly
about an otherwise language-specic construct, pointers, and how this abstraction
process inspired a workaround to a dilemma at the intersection of language design
and existing compiler technology.
The remaining sections apply constraints on memory use in the Burgers
equation solver of Chapter 9. Section 10.2 denes the problem that the constraints address. Section 10.3 describes an aspect of Fortran that poses a relevant
dilemma. Section 10.4 species a useful set of constraints on a UML class diagram.
Section 10.5 describes a useful design pattern that facilitates applying the constraints.
Section 10.7 details performance results from a case study culled from a research
application.

10.2 Problem Statement

Im Not as Think as You Drunk I Am!

Runtime checking of assertions inserted during software implementation can help
detect source- and compiled-code faults. Design constraints can motivate the
assertions. OCL provides a formal way to specify constraints.

10.2 Problem Statement

OCL constraints are meaningful only in the context of an object-oriented class model
(Warmer and Kleppe 2003). Figure 10.1 presents a UML class model for a Burgers
equation solver. As we move toward more complex examples in this part of the
book, many implementation details will be left to the reader as exercises at the end
of each chapter. We provide here a broad description of such an implementation
with emphasis on the perspective of client code.
In the sample Burgers solver design of Figure 10.1, the Burgers class sits at
the core and provides the primary interface with which most clients would interact.
Burgers implements the ABSTRACT CALCULUS dened by the Integrand interface. In
doing so, it likely makes itself an eligible argument for a time integration procedure
such as that the integrate subroutine packaged in Figure 6.1. Clients could declare

Surrogate

Hermetic

Integrand

IntegrationStrategy

uses

Burgers

RK2
FieldFactory

Field
ret

urn

HermeticField

HermeticFieldFactory
constructs

Figure 10.1. UML class diagram of a hermetic Burgers solver.

233

234

Formal Constraints

and construct a Burgers object and pass it repeatedly to integrate for advancement
one time step per invocation. Advancing from an initial time t to a nal time t_final
over time increments dt might take the form:
type(burgers) :: cosmos
cosmos = burgers()
do while (t<t_final)
call integrate(cosmos,dt)
t = t + dt
end do
where we have assumed the existence of default component initialization in the
overloaded structure constructor burgers.
In Figure 10.1, the Burgers class aggregates a reference to a Field component u.
Since Field is abstract, this reference must be a pointer or allocatable component
in Fortran. The structure constructor would use this component to dynamically allocate a Field instance. The closed diamond on the Burgers end of the Burgers/Field
relationship indicates that this instances lifetime coincides with that of the Burgers
object. It gets constructed when the Burgers object is constructed and nalized when
the Burgers object gets nalized.
Because this chapter focuses on situations where one needs to monitor compilers compliance with the memory management stipulations in the Fortran 2003
standard, the Field class in Figure 10.1 extends the Hermetic class described in
Chapter 5. This gives HermeticField utilities to assist in avoiding memory leaks
and dangling pointers. It also forces the HermeticField developer to dene a
force_finalization procedure intended to facilitate manual invocation of the nal
subroutine.
The Burgers class of Figure 10.1 makes use of an ABSTRACT FACTORY and
its concrete implementations FACTORY METHOD. The Burgers constructor uses a
HeremticFieldFactory to construct a HermeticField. It returns a pointer to an
abstract Field.
The Integrand interface uses STRATEGY and SURROGATE pattern implementations
to advance the Burgers object. Advancement happens via the RK2 algorithm in
the design shown. As demonstrated in Figure 9.1, each RK2 substep involves a
invoking the time differentiation method t() to evaluate the RHS of the Burgers
equation. This evaluation might take the form of the code in Figure 10.2, where
burgers extends the integrand of Figure 6.2(b) and aggregates a pointer to a Field
from Figure 9.2(b) and where we have written Burgers equation at line 22 in the
nonconservative form of equation 4.9 for simplicity.
The associate construct debuts in this text in burgers_module at line 23. It allows the creation of local aliases. In the case shown, v and dv_dt become a shorthand
for this%u and d_dt%u, respectively.
The procedures a compiler invokes to evaluate the RHS of line 24 include the
type-bound operators operator(*) and operator(-) and the type-bound functions
xx() and x(). Figure 10.3 shows the corresponding call tree.
The object dv_dt in Figure 10.3 ultimately becomes the result of the t()
type-bound procedure call in a time integration procedure such as integrate in

10.2 Problem Statement

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28

module burgers_module
use integrand_module ,only : integrand
implicit none
private
public :: burgers
type ,extends(integrand) :: burgers
class(field) ,allocatable :: u
contains
procedure :: t
procedure :: add => total
procedure :: assign
procedure :: multiply => product
generic :: operator(*) => product
generic :: assignment(=) => assign
end type
contains
function t(this) result(d_dt)
class(burgers) :: this
type(burgers) :: d_dt
class(field) ,pointer :: v, dv_dt
if (.not. allocated(this%u)) stop invalid argument inside t()
allocate(d_dt%u,source=this%u)
associate( v => this%u , dv_dt =>d_dt%u)
dv_dt = v%xx()*nu - v*v%x()
end associate
end function
! (additional type-bound procedures omitted)
end module burgers_module
Figure 10.2. Partial implementation of a Burgers solver class.

Figure 7.2(h). Line 27 of that gure reads:

this_half = this + this%t()*(0.5*dt) ! predictor step
which results in the additional cascade of calls detailed in Figure 10.4.
Each function call in Figure 10.3 produces a field result that is no longer
needed once the procedure or operator of lower precedence (lower on the call
tree) completes. Each function call in Figure 10.4 likewise produces an equally transient integrand containing a field. Thus, the combined Figures 10.310.4 represent
eight intermediate instantiations. These correspond to the results of u%xx(), u%x(),
operator(-), and two invocations of operator(*) in Figure 10.3 along with the
results of this%t(), operator(*), and operator(+) in Figure 10.4.
Judicious handling of the memory allocations and deallocations associated with
intermediate results proves critical when they occupy sufcient memory to crash a
run in the case of a memory leak. In 3D simulations of uid turbulence, for example, direct calculation of even modest laboratory conditions requires storing discrete
elds with roughly 5123 grid point values (de Bruyn Kops and Riley 1998). Given
typical 4-byte precision per value, each 5123 -point field would occupy 0.5 gigabytes. State-of-the-art simulations on the largest supercomputers involve distributed
memory allocations one to two orders of magnitude larger. Furthermore, coupling

235

236

Formal Constraints

dv_dt = v%xx()*nu - v*v%x()

(a)
nu

v%xx()

operator(*)

v%x()
operator(*)

operator(-)

dv_dt

assignment(=)

(b)
Figure 10.3. Burgers equation: (a) abstract calculus and (b) the resultant call tree.

this_half = this + this%t()*(0.5*dt) ! predictor step

(a)
this%t()

0.5*dt

operator(*)

this_half

this

operator(+)

assignment(=)

(b)
Figure 10.4. RK2 predictor step: (a) abstract calculus and (b) the resultant call tree.

in additional physics such as magnetic or quantum effects multiplies the memory

footprint several-fold more. If compilers do not free the intermediate results, the
burden falls on the programmer to do so. When the nal assignment at the bottom
of each call tree executes, the names associated with intermediate allocations are
out of scope, and it would be difcult or impossible for the programmer to free the
associated memory at that point.
Rouson, Morris, and Xu (2005) and Markus (2003) independently developed
essentially equivalent rules for freeing the intermediate memory allocated for
results produced in the process of evaluating ABSTRACT CALCULUS expressions.
Rouson, Xu, and Morris (2006) formalized these rules in a set of OCL constraints.
Here we update those constraints, applying the memory deallocation strategy of

10.3 Side Effects in Abstract Calculus

Stewart (2003), which handles a broader set of cases and allows us to leverage the
OBJECT pattern implementation of Chapter 5.
For the remainder of Chapter 10, we refer to all functions called in ABSTRACT
CALCULUS expressions as operators. In practice, some of these would be generic
type-bound operators and dened assignments, for example, operator(*) and
assignment(=) in Figure 10.3(b). Others would be functions that mimic integral
or differential mathematical operators such as xx(). With this nomenclature, we
dene intermediate results as temporary according to the following:
Denition 10.1. Any object is dened as temporary if it can be safely nalized at the
termination of the rst operator to which it is passed as an argument.
Corollary 10.1. All objects that are not temporary are persistent.
In the OBJECT implementation of Figure 5.2, any object that extends hermetic
starts life with a temporary component pointer (in its hermetic parent) initialized
to null(). This defaults all objects to being persistent. With this denition and its
corollary in hand, one can summarize the Stewart memory management scheme in
ve rules:
Rule 10.1. Mark all operator results as temporary.
Rule 10.2. Mark all left-hand arguments to dened assignments as persistent.
Rule 10.3. Guard all temporary operator arguments upon receipt.
Rule 10.4. Finalize all temporary operator arguments that were received unguarded.
Rule 10.5. Finalize persistent objects inside the procedure that instantiated them.
These rules constrain code behavior. The challenge for this chapter lies in expressing these constraints in OCL and incorporating the resulting formalisms into
the UML class model.
It Beats Flipping Burgers
The continual turnover of Burgers object allocations illustrates the necessity for
disciplined memory management by the compiler or the developer. Sorting out
when to free an objects memory can be difcult or impossible in the absence of
rules that distinguish temporary from persistent objects.

10.3 Side Effects in Abstract Calculus

OCL constraints can be satised by code in any language at development time.
Nonetheless, several Fortran-specic issues motivate the content of this chapter.
This section outlines these issues.
In cases where the field component inside a burgers object is allocatable,
the Fortran 2003 standard obligates the compiler to free the associated memory after the object goes out of scope. In cases where the field component is a pointer,
the most straightforward way is to free the memory inside a final subroutine. The
developer can then rely on the compiler to invoke the final subroutine after the
object goes out of scope. As explained in Section 5.2, the allocatable-component

237

238

Formal Constraints

approach has proved problematic with recent versions of several compilers. That
section also explains that the pointer-component approach remains infeasible for
most compilers. Even when either approach works, developers might desire to control the memory utilization more closely in order to recycle allocated space or to free
it sooner than the compiler might.
Fortrans requirement that type-bound operators and the procedures they call be
free of side effects greatly complicates programmers ability to free memory allocated
for operator arguments. The intent(in) attribute enforces the requirement but
also poses a dilemma. In nested invocations of operators, an argument passed to
one operator might be a result calculated by another operator of higher precedence.
That higher-precedence operator is likely to have allocated memory for its result
just as the lower-precedence operator might allocate memory for its result. The
easiest and most efcient place for the programmer to release memory that was
dynamically allocated inside the result of one operator is inside the operator to which
this result is passed. However, the operator receiving the result cannot modify it. A
similar dilemma relates to dened assignments such as the assignment(=) binding
in Figure 10.2, wherein there frequently arises a need to free memory associated
with allocatable components of the RHS argument if it is the result of an expression
evaluation.
Even compilers that do the proper deallocations might not do so economically.
For example, the rst author has analyzed intermediate code received from one vendor and found that this vendors compiler carries along all the memory allocated at
intermediate steps in the call tree, performing deallocations only after the nal assignment at the top of the tree (Rouson et al. 2006). In complicated expressions written
on large objects, that vendors approach could make ABSTRACT CALCULUS infeasible.
Hence, in addition to correcting for compiler nonconformance with the Fortran standards memory management stipulations, the rules and resulting constraints in this
chapter economize memory usage in ways the standard does not require.
Warning: Calculus May Cause Memory Loss, Bloating, and Insomnia
Fortrans prohibition against side effects in type-bound operators complicates the
process of explicitly deallocating objects before they go out of scope and can lead
to unrestrained growth in memory usage. Even compilers that correctly automate
deallocation or nalization might not do so economically, resulting in developers
losing sleep strategizing to free memory sooner than would the compiler.

10.4 Formal Specication

10.4.1 Modeling Arrays with OCL
Although the field class component array in our Burgers solver is 1D, the lions
share of scientic research problems are more naturally expressed with multidimensional arrays. Modeling these arrays in OCL therefore bears some discussion.
Although OCL has a sequence type that could naturally be used to model 1D
arrays, the language does not contain an intrinsic multidimensional array type.

10.4 Formal Specication

Because OCL is a subset of UML, however, types dened by a UML model are considered OCL model types. We can therefore incorporate the Array type in Figure 2.3
into models for use in OCL. We do so in the remainder of this chapter for the 1D
case and leave the multidimensional case as a modeling exercise for the reader.

10.4.2 Hermeticity
We can now specify the constraints that preclude a category of memory leaks in
composite operator calls. We refer to leak-free execution as hermetic memory
management, or simply hermeticity. The memory of interest is associated with the
allocatable array components inside data structures passed into, through, and out
of call trees of the form of Figures 10.310.4. We assume no other large memory
allocations inside operators. This assumption is discussed further vis--vis economy
in Section 10.4.3.
Figure 10.5 provides a partial class diagram, including several ABSTRACT CALCULUS operators: the unary operator x(), binary operator add, and a dened assignment
assign. Our primary task is to constrain the behavior of such operators using the
memory management rules listed in Section 10.2. We model Denition 10.1 and
Corollary 10.1 simply by including in the HermeticField class a temporary attribute,
modeled as an instance of the Boolean model type described in Figure 2.3. The value
of this attribute classies the object in the set of objects dened by the Denition or
in the complementary set dened by the Corollary. The use of the Boolean temporary attribute demonstrates several differences between a software model and a
software implementation. For example, an implementation of HermeticField would
likely extend Hermetic as in Figure 10.1, thereby inheriting the temporary and
depth attributes as well as the force_nalization() method. That level of detail
is unnecessary for purposes of writing the intended constraints. Along this vein, to
consider the hermetic derived type component temporary in Figure 5.2 to be an implementation of the HermeticField temporary attribute, one has to map the integer
value 0 for the component to the value True for the Boolean type in Figure 2.3, and
map all nonzero values to False. Finally, even the last artice misses the detail that
the temporary component in the hermetic implementation Figure 5.2 is a pointer.
UML does not contain a pointer type or attribute, but the modeling of pointers in
UML plays a central role in Section 10.4.3, so we defer further discussion to that
section.

+
+
+
+

HermeticField
values : Array<Real>
temporary : Boolean
depth : Integer
x() : HermeticField
add(HermeticField) : HermeticField
assign(HermeticField)
force_nalization()

Figure 10.5. HermeticField partial class diagram.

239

240

Formal Constraints

In formal methods, constraints take three forms: preconditions that must be true
before a procedure commences; postconditions that must be true after it terminates;
and invariants that must be true throughout its execution. OCL preconditions bear
the prex pre:, whereas post: precedes postconditions and inv: precedes
invariants. The contexts for OCL pre- and postconditions are always operations
or methods. When writing OCL constraints as standalone statements, a software
designer writes the constraint context above the constraint. Hence, Rule 10.1 applied
to the Hermetic method SetTemp() in Figure 10.6 could be written:
context : Hermetic :: SetTemp()
post: temporary = true
Alternatively, the designer indicates the context by linking the constraint to the
corresponding operations visually in a UML diagram. The visual link might involve
writing a constraint just after the method it constrains in a class diagram or by drawing
connectors between the constraint and the method as in Figure 10.6. When using
connectors, one commonly encloses the constraint in a UML annotation symbol as
Figure 10.6 depicts.
The argument and return value declarations in Figure 10.6 reect OCL conventions along with a desire for succinct presentation. For both of these reasons,
Figure 10.6 omits the implicit passing of the object on which the method is invoked.
In C++ and in many of the Fortran examples in this book, that object is referenced
by the name this. In OCL, that object is named self and can be omitted in the
interest of brevity. The postconditions applied to the Hermetic in Figure 10.6 omit
self because it is the only object in play. Several postconditions applied to HermeticField, however, include it to explicitly distinguish its attributes from those of the
other the objects received or returned by the methods.
We formally specify Rule 10.1 through postconditions on the arithmetic and
differential operators x() and add(), respectively, and Rule 10.2 through a postcondition on the dened assignment assign(). Specically, two postconditions in
Figure 10.6 specify that the results of x() and add() must be temporary. Another
postcondition species that the LHS argument (self) passed implicitly to assign
must be persistent.
Rule 10.3 governs the treatment of the subset of temporary objects that must
be guarded against premature nalization. According to the memory management
scheme of Stewart (2003), the GuardTemp() procedure guards a temporary object by
incrementing its depth count. The postcondition applied to the like-named method
in Figure 10.6 stipulates this.
Formalization of the guarded state dened by rule 10.3 requires making a
statement about the state of a class of objects throughout a segment of the program
execution. One naturally expresses such concepts via invariants. In OCL invariants, the context is always a class, an interface, or a type. For modeling purposes,
Figure 10.6 therefore denes a GuardedField class that exists only to provide a welldened subset of HermeticField specically for purposes of specifying the invariant
that subset satises: depth>1.
As the GoF demonstrated, class diagrams also provide an effective medium
through which to describe conforming implementations. The note linked to the

10.4 Formal Specication

Hermetic
temporary : Boolean

{post: temporary = True}

depth : Integer
{post: temporary=True implies depth = depth + 1}

+ SetTemp()
+ GuardTemp()

{post: temporary=True and depth>1) implies depth = depth 1}

+ CleanTemp()
...
if (depth==1)call force_finalization()

+ force_finalization()

{post: d_dx.temporary = True}

{post: self.depth=1 implies not self.values.allocated()}

HermeticField
values : Array<Real>
+ x() d_dx : HermeticField
+ add(rhs : HermeticField) total : HermeticField
+ assign(rhs : HermeticField)

{post: self.temporary = False}

{post: rhs.depth=1 implies not rhs.values.allocated()}

GuardedField

{post: total.temporary = True}

{post: rhs.depth=1 implies not rhs.values.allocated()}
{post: self.depth=1 implies not self.values.allocated()}

{inv: depth > 1}

...
type (GuardeField),pointer :: guarded_self, guarded_rhs
call precondition(self,rhs)
call self.GuardTemp(); call rhs.GuardTem()
guarded_self => GuardeField(self)
guarded_rhs => GuardedField(rhs)
call do_add(guarded_self,guarded_rhs)
call self.CleanTemp(); call rhs.CleanTemp()
Call postcondition(self,rhs)
...

Figure 10.6. UML class diagram of a hermetic Burgers solver.

add() operation in Figure 10.6 displays a skeletal implementation. In a complete

implementation, the depicted precondition() procedure would likely check a set
of assertations inspired by formal preconditions. Then invoking GuardTemp() on
self and rhs would guard these objects by increasing their depth counts if they
are temporary. Next the GuardedField constructor calls might verify that the depth
count exceeds unity before returning pointers to the HermeticField objects, giving
them the dynamic type GuardedField.

241

242

Formal Constraints

Next the do_add() procedure would then sum its GuardedField arguments, the
type denition for which would be sufciently minimal and intimate in its connection to HermeticField that it would make sense for it to have access to the
private data inside its parent HermeticField. In Fortran, this would take place
most naturally by dening both types in one module. The CleanTemp() procedure would unguard temporary objects by decrementing their depth count and
would call force_finalization if the depth count drops to unity, as demonstrated
in the sample code linked to that method in Figure 10.6. Finally, the procedure
postcondition() would check assertions inspired formal postconditions.
Rule 10.4 governs the nalization of temporary objects. It applies to all operators
and dened assignments. As applied to x(), the corresponding formal constraint
takes the form of the second postcondition linked to that method in Figure 10.6.
Predicate calculus stipulates that this postcondition evaluates to true if the expression
after the implies operator is true whenever the expression before the implies is true.
The constraint also evaluates to true whenever the expression before the implies is
false.
Rule 10.5 governs the nalization of persistent objects. Only dened assignments
create persistent objects. Since all HermeticField operators have access to private
HermeticField components, they can make direct assignments to those components.
HermeticField operators therefore do not call the dened assignment procedure, so
most persistent HermeticField objects are created in client code outside the HermeticField class. For this reason, no postcondition corresponding to Rule 10.5 appears
in Figure 10.6.
As valid OCL expressions, the constraints in Figure 10.6 are backed
by a formal grammar dened in Extended Backus-Naur Form (EBNF)
(Rouson and Xiong 2004). An EBNF grammar species the semantic structure of
allowable statements in a formal language. The statements meanings can therefore
be communicated unambiguously as part of the UML design document. An additional advantage of formal constraints is their ability to express statements that could
not be written in an executable language. One example is the relationship between
the Boolean expressions in the implied statement in Figure 10.6. Another is the fact
that these constraints must be satised on the set of all instances of the class. Programs that run on nite-state machines cannot express conditions on sets that must
in theory be unbounded. A third example is any statement about what must be true
at all times, such as the invariant constraint on the GuardedField class in Figure 10.6.
In general, an additional benet of formal methods accrues from the potential to
use set theory and predicate logic to prove additional desirable properties from the
specied constraints. However, such proofs are most naturally developed by translating the constraints into a mathematical notation that is opaque to those who are
not formal methods specialists. Since OCL targets non-specialists, proving additional
properties mathematically would detract from the main purpose of simply clarifying
design intent and software behavior.
10.4.3 Economy
Rening Rules 10.110.5 can lead to more economical memory usage. Although
the memory reductions are modest in some instances, we hope to demonstrate

10.4 Formal Specication

an important tangential benet from the way OCL forces designers to think more
abstractly about otherwise language-specic constructs. For this purpose, consider
again the function add(), which in the model of Figure 10.6 takes the implicit argument self and the explicit argument rhs and returns total. We can ensure
economical memory usage by specifying that temporary memory be recycled. To facilitate this, the scalars self, rhs, and total must have the pointer or allocatable
attribute. In many respects, the allocatable approach is more robust: It obligates
the compiler to free the memory when it goes out of scope and the move_alloc()
intrinsic facilates moving the memory allocation from one object to another without costly deallocation and reallocation (see Section 6.2.1). However, because
allocatable scalars are a Fortran 2003 construct without universal compiler support,
the pointer approach is desirable from the standpoint of portability. We consider
that approach for the remainder of this chapter.
To pass pointers in Fortran, one writes the add() function signature as:
function add(self,rhs) result(total)
class(HermeticField), pointer, intent(in) :: self
type(HermeticField), pointer, intent(in) :: rhs
type(HermeticField), pointer :: total
Since OCL does not have a pointer type, however, we model pointers as a UML association between two classes. Specically, we model self, rhs, and total as instances
of a HermeticFieldPointer class that exists solely for its association with the HermeticField class. We further assume HermeticFieldPointer implements the services of
Fortran pointers, including an associated() method that returns a Boolean value
specifying whether the object on which it is invoked is associated with the object
passed explicitly as an argument.
Figure 10.7 diagrams the HermeticFieldPointer/HermeticField relationship. A
diagrammatic technique appearing for the rst time in that gure is the adornment
of either end of the relationship with multiplicities (see Section 2.5) describing the
allowable number of instances of the adjacent classes. The adornments in Figure 10.7
indicate that one or more HermeticFieldPointer instances can be associated with any
one HermeticField instance.
In Figure 10.7, we apply the label target to the HermeticField end of the association. From a HermeticFieldPointer object, this association can be navigated
through the OCL component selection operator ., so an economical postcondition might take the form shown in the annotation linked to HermeticFieldPointer,
wherein the implied conditions stipulate that total must be associated with one of

HermeticFieldPointer
1*
+ Boolean : associated(HermeticField)
+ add(rhs : HermeticFieldFieldPointer)

HermeticField

target

{post: self.target.temporary implies associated (total, self.target)}

{post: ((not self.target.temporary) and rhs.target.temporary)
implies associated (total, rhs.target)

Figure 10.7. HermeticField and HermeticFieldPointer class association.

243

244

Formal Constraints

the temporary targets. In writing the constraint, we arbitrarily chose to overwrite

self in the case in which both self and rhs are recyclable.
We advocate reusing the entire temporary object while giving its array component, if any, the allocatable attribute, rather than the pointer attribute. In
addition to the aforementioned convenience of allocatable arrays, they offer performance advantages over pointers. An allocatable array references contiguous
memory that can be accessed with unit stride, whereas a pointer can target an array subsection that might not be contiguous (Metcalf, Reid, and Cohen 2004). Any
resulting lack of spatial locality could retard cache performance.

10.5 The Shell Pattern

From a programmability standpoint, the greatest benet of developing the economizing constraints in Figure 10.7 stems from the conceptual leap required to model
pointers in OCL. Fortrans prohibition against side effects in dened operators
necessitates the ability to specify pointer arguments intent another technology
for which compiler support lagged the publication of the Fortran 2003 standard by
several years. Developing the HermeticFieldPointer class model in UML naturally
inspires the notion of constructing an actual class analogous HermeticFieldPointer.
Doing so facilitates emulating pointer intent without compiler support for it.
The construction of a class analogous to HermeticFieldPointer served
as the critical linchpin that enabled the rst ADT calculus syntax in Fortran (Rouson, Xu, and Morris 2006). Without it, it would have been impossible to
satisfy the hermeticity constraints in Figure 10.6 with recent compiler versions. In the
absence of correct, automatic compiler deallocation of memory, the programmers
hands are tied for the reasons described in Section 10.3. For this reason, earlier attempts at ADT calculus in Fortran used an awkward syntax, evaluating derived type
expressions via explicit procedure invocation with an adjacent comment displaying
the desired operator syntax:
! total = lhs + rhs
call assign(total,add(lhs,rhs))
Although this approach worked well for simple expressions, more complicated examples render code nearly indecipherable on rst inspection. Consider adopting a
similar approach to writing the 3D heat equation in a form patterned after Table 3.1:
! dT_dt = alpha *(T%xx() + T%yy() + T%zz())
call assign(dT_dt,times(alpha,(addT%xx(),add(T%yy(),T%zz()))))
where times implements the multiplication operator (*) in the commented code. The
readability of such code gets progressively worse with increasingly complicated sets of
equations such as those encountered in multiphysics applications. (See Chapter 12.)
The construction of a HermeticFieldPointer class such as:
type HermeticFieldPointer
type(HermeticField) ,pointer :: target_field=>null()
end type

10.6 An Assertion Utility for Fortran

along with a corresponding constructor and various dened operators such as:
function add(self,rhs) result(total)
class(HermeticFieldPointer) ,intent(in) :: self
type(HermeticFieldPointer) ,intent(in) :: rhs
type(HermeticFieldPointer)
:: total
end function
facilitates satisfying the Fortrans intent(in) requirement and restricts us from
changing with what target object the target_field is associated while allowing us
to deallocate any allocatable entities inside the target with compilers that fail to do so.
Since HermeticFieldPointer exists only to wrap a HermeticField instance, one
might refer to the former as a shell and the latter as a kernel. Together they comprise
a SHELL pattern. Chapter 11 describes similar instances in which a lightweight class
serves as a wrapper or identication tag for another class. An exercise at the end
of the current chapter asks the reader to write a SHELL pattern description using the
three-part rule of Section 4.1.

10.6 An Assertion Utility for Fortran

Formal constraints make statements that must always be true for an application to
conform to the given specication. Relative to this level of certitude, run-time assertion checking must necessarily be considered incomplete: Each run tests only one
set of inputs. Nonetheless, the regularity of many scientic simulations mitigates the
impact of this limitation. In the Burgers solver, for example, the memory allocations,
reallocations, and deallocations are the same at every time step independent of the
initial data.
Ideally, constraints could be translated into program statements automatically.
Tools exist for generating Java code from OCL. Although some of the earliest work
on automated assertion insertion targeted Fortran 77 (Stucki and Foshee 1971), we
know of no such efforts for Fortran 2003. Fortran lacks even a language mechanism
intended specically for writing assertions as exists in C, C++, and Java. In the
absence of a language construct, the options remain to write ad hoc conditional
(if-then) statements or to link to an assertion toolkit.
Press et al. (1996) published assertion tools similar to those in Figure 10.8. The
two subroutines in that gure terminate execution after printing a message, if provided one, to error_unit, which on Unix-like systems corresponds to printing to
the standard error stream. The subroutine assert() traverses an logical array,
each element of which would likely be a Boolean expression. Execution terminates
gracefully (without a platform-dependent, system-generated error message) if any
element of the array evaluates to .false.. The subroutine assert_identical()
traverses an integer array, checking whether each element matches the rst one.
If not, the execution halts gracefully after reporting the mismatches. As suggested
by Press and colleagues, a common use of the latter utility involves passing a list of
array sizes. The rst executable line of assert() invokes assert_identical() for
this purpose.
To be viable in production settings, an assertion facility must have negligible
impact on run time. Languages with built-in assertion capabilities typically allow the

245

246

Formal Constraints

developers to disable assertions at compile-time. In such cases, the developers can

turn assertion checking on and off, using assertions when debugging but eliminating
them during production runs. In Fortran, one can emulate this capability by providing
a global, compile-time, logical constant assertions value and conditioning all
assertion checking on this value in the manner:
logical ,parameter :: assertions=.false.
if (assertions) call assert(sanity_checks,insanity_warnings)
where sanity_checks and insanity_warnings would contain arrays of logical expressions and corresponding error messages, respectively. With this approach, most
compilers would eliminate the assertion calls during a dead-code-removal optimization phase because it is clear that the call can never be made. (Section 1.7 described
a similar scenario for a global debugging parameter.) This behavior would likely
depend on the optimization level used; however, most production scientic codes
use the highest optimization levels available.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35

module assertion_utility
use iso_fortran_env ,only : error_unit
implicit none
private
public :: error_message,assert,assert_identical
type error_message
character(:) ,allocatable :: string
end type
contains
subroutine assert(assertion,text)
logical ,dimension(:) ,intent(in) :: assertion
type(error_message) ,dimension(:) ,intent(in) :: text
integer :: i
logical :: any_failures
call assert_identical( [size(assertion),size(text)] )
any_failures=.false.
do i=1,size(assertion)
if (.not. assertion(i)) then
any_failures=.true.
write(error_unit,*) Assertion failed with message:
if (allocated(text(i)%string)) then
write(error_unit,*) text(i)%string
else
write(error_unit,*) (no message provided).
end if
end if
end do
if (any_failures) stop Execution halted on failed assertion(s)!
end subroutine
subroutine assert_identical(integers)
integer ,dimension(:) ,intent(in) :: integers
integer :: i
logical :: any_mismatches
any_mismatches = .false.
do i=2,size(integers)

10.7 Case Study: A Fluid Turbulence Solver

36
37
38
39
40
41
42
43
44

if (integers(i) /= integers(1)) then

any_mismatches = .true.
write(error_unit,*) &
Value ,i, does not match expected value ,integers(1)
end if
end do
if (any_mismatches) stop Execution halted on failed assertion!
end subroutine
end module
Figure 10.8. Fortran assertions utility.

10.7 Case Study: A Fluid Turbulence Solver

This section describes results associated with a production research code that employed run-time assertions based on a set of rules and constraints that were simpler
than Rules 10.110.5 and Figure 10.6. The code solves the Navier-Stokes equations
for uid dynamics for a canonical yet computationally demanding turbulent ow.
Rouson et al. (2008a) gave a comprehensive description of the software architecture
and the numerical methods employed. Rouson et al. (2008b) and Morris et al. (2008)
described new science illuminated via multiphysics applications of the software. Here
we focus on the performance impact of checking run-time assertions inspired by OCL
constraints.
The rules employed in this case inspired the inclusion of a Boolean temporary
ag in the derived type denition but no depth count. Given the role the depth
count plays in guarding temporary objects from premature deletion, the lack of a
depth count necessitates dropping Rule 10.3 and modifying 10.4. The new set of
rules follow:
Rule 10.6. Mark all operator results as temporary.
Rule 10.7. Mark all left-hand arguments to dened assignments as persistent.
Rule 10.8. Deallocate all memory associated with temporary operator arguments.
Rule 10.9. Deallocate all memory associated with persistent objects inside the
procedure that instantiated them.
where the meanings of temporary and persistent remain based on Denition 10.1
and Corollary 10.1. These are equivalent to the rules put forth by Markus (2003) and
Rouson, Morris, and Xu (2005). Rules 10.610.9 use deallocate instead of nalize to address situations in which it is necessary or desirable for the programmer to
explicitly deallocate memory rather than to rely on a final binding.
In the absence of a depth count, Rules 10.610.9 can cause problems in
nested invocations. For illustrative purposes, consider the following case adapted
from Stewart (2003):
1
2
3
4

type(field) :: a, b, c
a = field()
b = field()
call do_something(a + b, c)

247

248

Formal Constraints

where we assume the field type contains one or more allocatable components and
a logical component that identies it as temporary or persistent. We further assume
the field() constructor allocates memory for allocatable components inside a and
b and gives those components default values. In this scenario, the addition operator
invoked inside the rst argument slot at line 4 would mark its result temporary based
on Rule 10.6.
Next consider the following denition of the above do_something procedure:
1
2
3
4
5
6

subroutine do_something(x, y)
type(field) ,intent(in) :: x
type(field) ,intent(inout) :: y
y =x
y = x + y
end subroutine

where the temporary object now referenced by x inside do_something() would have
its allocatable components deallocated by the dened assignment at line 4 in the
latter code in accordance with Rule 10.6. That deallocation would in turn preclude
proper execution of the addition operator in the subsequent line. Nonetheless, the
current case study involved no such scenarios and the simplied Rules 10.610.9
sufced.
Figure 10.9 presents precondition() and postcondition() subroutines
adapted from Rouson et al. (2006) using the assertion utility of Figure 10.8. In
the subject turbulence code, each field instance contains two allocatable array
components: One stores a physical-space representation and another stores the corresponding Fourier coefcients. Since one can always compute Fourier coefcients
from the physical-space samples that generated those coefcients, and vice versa,
having both allocated is either redundant or inconsistent. It is therefore never valid
for both to be allocated except momentarily in the routine that transforms one to
the other. This condition is checked in the Figure 10.9.
The aforementioned turbulence simulations employed analogous routines. The
code called a precondition() procedure at the beginning of each field method. It
called a postcondition() procedure at the end of each method.
Besides monitoring the source code behavior, these assertions proved eminently
useful in checking for compiler bugs. Using these routines, Rouson et al. (2006) discovered that one vendors compiler was not always allocating arrays as requested.
Despite repeated attempts to reproduce this problem in a simple code to submit in
a bug report, the simplest demonstration of the error was a 4,500-line package in
which the error only occurred after several time steps inside a deeply nested call
tree. The related saga cost several weeks of effort plus a wait of more than a year
for a compiler x. The disciplined approach to checking assertions at the beginning
and end of each procedure, as inspired by the OCL pre- and postconditions, paid its
greatest dividend in motivating a switch to a different compiler.
Figure 10.10 presents a procedure that veries economical memory recycling
based on a SHELL class named field_pointer. Rouson et al. (2006) used analogous
code to check whether memory allocated for temporary arguments had been recycled
by associating that memory with the function result.

10.7 Case Study: A Fluid Turbulence Solver

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18

subroutine precondition(this,constructor)
use assertion_utility ,only : assert
implicit none
type(field) ,intent(in) :: this
logical ,intent(in) :: constructor
logical :: both_allocated ,at_least_one_allocated
both_allocated= allocated(this%fourier) .and. allocated(this%physical)
call assert( [.not. both_allocated] &
,[error_message(redundant or inconsistent argument)])
at_least_one_allocated = &
allocated(this%fourier) .or. allocated(this%physical)
if (constructor) then
call assert( [.not. at_least_one_allocated] &
,[error_message(constructor argument pre-allocated)])
else; call assert( [at_least_one_allocated] &
,[error_message(argument data missing)])
end if
end subroutine

19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42

subroutine postcondition(this,public_operator,deletable)
use assertion_utility ,only : assert
implicit none
type(field) ,intent(in) :: this
logical ,intent(in) :: public_operator,deletable
logical :: both_allocated,at_least_one_allocated
both_allocated= allocated(this%fourier) .and. allocated(this%physical)
call assert( [.not. both_allocated] &
,[error_message(redundant or inconsistent result)])
at_least_one_allocated = &
allocated(this%fourier) .or. allocated(this%physical)
if (public_operator .and. deletable) then
if (this%temporary) then
call assert([.not. at_least_one_allocated] &
,[error_message(temporary argument persisting]))
else; call assert([at_least_one_allocated] &
,[error_message(persistent argument deleted)])
end if
else
call assert([at_least_one_allocated] &
,[error_message(invalid result]))
end if
end subroutine
Figure 10.9. Memory allocation assertions in a Fourier-spectral turbulence code.

Since our pre- and postcondition subroutines contain simple Boolean expressions, they require insignicant amounts of execution time. Table 10.1 shows that the
most costly procedure for the case studied by Rouson et al. (2006) the transform(),
which contained 3D FFT calls. That procedure accounted for 34 percent of the
processor time. By contrast, precondition() and postcondition() occupied immeasurably low percentages of execution time even though the number of calls to
these routines exceeds the calls to transform() by roughly an order of magnitude.

249

250

Formal Constraints
1
2
3
4
5
6
7
8
9
10
11
12
13

subroutine economical_postcondition(lhs,rhs,result)
use assertions ,only : assert
implicit none
type(field_pointer) ,intent(in) :: lhs,rhs,result
if (lhs%target%temporary) then
call assert( [associated(result%target,lhs%target)] &
,[error_message(lhs not recycled]))
else if (right%target%temorary) then
call assert( &
[associated(result%target,rhs%target)] &
,[error_message(rhs not recycled]))
end if
end subroutine
Figure 10.10. Memory economy assertions in a Fourier-spectral turbulence code.

Table 10.1. Procedural run-time distribution

Function Name
transform
assign
xx
dealias
add
times
...
precondition
postcondition

Number of Calls

% of Total Runtime

108
167
156
36
75
75
...
1477
885

34.42
19.67
16.18
7.70
7.43
5.03
...
0
0

Thus, the capability to turn assertions off appears to be nonessential in the chosen application. This observation is likely to hold for the majority of demanding,
leading-edge scientic applications.

EXERCISES
1. Write a set of preconditions prescribing desirable memory management behavior
for the methods in Figure 10.6.
2. Using the three-part rule of Section 4.1, write a more complete description of the
SHELL pattern.
3. Modify the assertion utility of Figure 10.8 so that the error_message derived
type has private data along with a error_message generic procedure that overloads the intrinsic structure constructor and supports the constructor call syntax
in Figure 10.9.

11 Mixed-Language Programming

But if thought corrupts language, language can also corrupt thought.

George Orwell

11.1 Automated Interoperability Tools

In human cultures, membership in a community often implies sharing a common
language. This holds as much for computer programming languages as for natural human tongues. Certain languages dominate certain research communities.
As a result, the modern drive towards multidisciplinary studies inevitably leads to
mixed-language development. As noted in the preface, survey evidence suggests
that approximately 85% of high-performance computing users write in some avor
of C/C++/C#, whereas 60% write in Fortran. These data imply that at least 45% write
in both Fortran and C/C++/C#, so the union of these two language families likely
contains the most important pairings in mixed-language scientic programming.
In many settings, the language version matters as much as the language identity.
For example, interfacing other languages with object-oriented Fortran 2003 poses
a much broader set of challenges than does interfacing with procedural Fortran 77
or even object-based Fortran 95. It also matters whether one language invokes code
in a second language, the second invokes code in the rst, or both. As suggested by
Rasmussen et al. (2006), one can account for invocation directionality by considering
ordered pairs of languages, where the order determines which language is the caller
and which is the callee.
Consider a set {A, B, C, . . .} of N dialects, where each member is a language
or language version distinct from each other member. In the worst case, when each
member comes from a different language or when language versions do not maintain
perfect backward compatibility, a project that involves code in each dialect invoking
code in each other dialect generates N(N 1) = O(N 2 ) ordered dialect pairings.
Most such pairings require glue code (also called bridging code) that ensures type,
variable, and procedure compatibility. Supporting language interoperability in such
a setting may be cost-prohibitive.
Broadly, two categories of solutions reduce the cost of supporting language
pairings. One class of tools reduces the programming overhead associated with
each bilingual pairing by generating the relevant glue code automatically. A more
251

252

Mixed-Language Programming

ambitious class of tools eliminates the quadratic growth in ordered pairings by inserting an intermediary through which an object declared in one language communicates
with objects in all other languages without either side knowing the identity of the
other side. The intermediary plays a role analogous to the MEDIATOR mentioned in
Sections 4.2.3 and 8.1. Such tools generally work with a common interface representation: an Interface Denition Language (IDL). The remainder of this section
discusses tools that follow both approaches.
The Simplied Wrapper and Interface Generator (SWIG) automates the interface construction process. SWIG interfaces C and C++ code with codes written in
several other languages, including Python, Perl, Ruby, PHP, Java, and C#1 . To access
C functions, for example, from another language, one writes a SWIG interface le.
This le species the function prototypes to be exported to other languages. From an
interface le, SWIG generates a C wrapper le that, when linked with the rest of the
project code, can be used to build a dynamically loadable extension that is callable
from the desired language. Often, the developer need not write the interface le line
by line and can instead incorporate a C/C++ header le in a SWIG interface le via
a single #include directive.
SWIG does not provide direct support for interfacing to Fortran. Since SWIG
interfaces to Python, one workaround involves using the PyFort tool2 that enables
the creation of Python language extensions from Fortran routines as suggested by
Pierrehumbert3 . PyFort supports Fortran 77 with plans for supporting Fortran 90,
but Pierrehumbert found it easier to manually construct C wrappers for Fortran
routines.
The Chasm language interoperability toolkit targets scientic programmers.
Chasm enables bidirectional invocations between statically typed languages that is
languages in which type information can be discovered in the compiled code prior to
runtime including Fortran, C, C++, and Java. (Chasm also enables unidirectional
invocations from dynamically typed languages such as Ruby and Python to statically
typed languages but not the reverse.) Because many scientic applications are written in Fortran and many libraries in C++, the most oft-cited discussion of Chasm
focuses on Fortran/C++ pairings (Rasmussen et al. 2006). In this setting, Chasm facilitates invoking methods on C++ objects from Fortran 95 and vice versa. Chasm
aims for the interoperability holy grail articulated by Barrett (1998):
Informally, a polylingual system is a collection of software components, written in diverse languages, that communicate with one another transparently. More specically, in
a polylingual system it is not possible to tell, by examining the source code of a polylingual software component, that it is accessing or being accessed by components of other
languages. All method calls, for example, appear to be within a components language,
even if some are interlanguage calls. In addition, no IDL or other intermediate, foreign
type model is visible to developers, who may create types in their native languages and
need not translate them into a foreign type system.

Chasm therefore eschews the use of a specially tailored IDL. Chasm instead
automatically generates an Extensible Markup Language (XML) representation of
1 http://www.swig.org
2 http://sourceforge.net/projects/pyfortran
3 http://geosci.uchicago.edu/ rtp1/itr/Python.html

11.1 Automated Interoperability Tools

procedure interfaces. Chasm does so by parsing a compilers intermediate language

representation of a codes abstract syntax tree, which is a simplied description of
the codes syntactical structure.
A complete multilanguage build process with Chasm takes three steps:
1. Static analysis,
2. Bridging-code generation, and
3. Compilation and linking.
Step 1 produces the XML description of the procedure interfaces found in the compilers intermediate-language representation. With the XML output from Step 1 as
input, Step 2 outputs stub and skeleton interfaces as illustrated in Figure 11.1. Step 3
takes whatever form the user chooses. For the smallest packages, manually invoking
the compiler and linker might sufce. For moderately larger packages on Windows
and POSIX-compliant (Unix) platforms, building might involve writing a Makele
that species le dependencies and compilation and linking syntax. For the largest
collections of packages, Makele construction itself might be automated with tools
such as Gnu4 Autotools or CMake5 (Martin and Hoffman 2008).
In the schema of Figure 11.1, a calling procedure in language A invokes a
language-A stub. Stubs exemplify the GoF ADAPTER pattern. The GoF describes
the following goal of an ADAPTER: Convert the interface of a class into another
interface clients expect. A language-A stub provides a natural interface in A, hiding compiler-specic conventions the language-B compiler uses to generate symbol
names in object les and libraries. Section 11.2 details some of these conventions.
Continuing with Figure 11.1, the language-A stub invokes the language-B
skeleton, which addresses additional compiler-specic handling of argument and
return-value data representations, specically the layout of Fortran-derived types,
assumed-shape arrays, and pointers thereto. For some argument and return types,
no skeleton need be constructed. Furthermore, in some cases, the compiler can
inline the stub, incorporating the associated executable code directly into that associated with the calling routine. With an inlined stub and no skeleton, the overhead
associated with A-B interoperability is minimal.
When language syntax and semantics lead to ambiguities in transferring data to
another language, Chasm requires the user to tweak some of the generated XML

Language A
Caller

Language A
Stub

Language B
Skeleton

Language B
Callee

Generated automatically by Chasm

Figure 11.1. Chasm stub and skeleton interfaces for language A procedure invoking language
B procedure.
4 http://www.gnu.org
5 http://www.cmake.org

253

254

Mixed-Language Programming

manually. An important case involves pointers in C and C++. An argument listed as a

pointer to a C/C++ value, for example, an int* or double*, might resolve to a single
value passed by reference in Fortran or to a 1D array. Knowledge about the development conventions of a particular project can assist in resolving such ambiguities,
as Section 11.3 discusses.
According to Rasmussen et al. (2006), Chasm does not support distributed objects. This limits its utility for the high-performance computing applications that
dominate much of scientic programming. Furthermore, its most recently published
version precedes the December 2008 release of the rst fully compliant Fortran 2003
compiler by at least two years. Chasm is therefore unlikely to support the OOP
constructs of Fortran 2003.
The Babel tool described by Allan et al. (2006) takes the second of the aforementioned two paths to reducing interoperability complexity: eliminating the quadratic
explosion of dialect-pairings by inserting a MEDIATOR. A Babel user writes an interface le in a specication language: the Scientic Interface Denition Language
(SIDL) described by Dahlgren et al. (2009). Babel uses a single SIDL interface to
generate language-specic bindings for C, C++, Fortran 77, Fortran 90, Python, and
Java pairings.
As its name implies, SIDL is tailored to scientic programming. For example, it
supports a rich set of mathematical constructs, including, for example, dynamically,
allocated, multidimensional, complex arrays of non-unit stride. A desire to support
scientic computing also sets the agenda for future Babel development. The Babel
team cites support for massively parallel, distributed objects as a priority for future
development.
Babel mediates communications between languages via an intermediate object
representation (IOR) written in ANSI C. Babel enables OOP in each supported
language, including languages that do not provide native support for OOP. In each
dialect, Babel enables inheritance, polymorphism, encapsulation, and object identity. For example, Babel implements object identity via struct pointers in C, via 64-bit
integers in Fortran 77, and via derived types with 64-bit integer components in Fortran 90/95. With Babel, one could implement each level of an inheritance hierarchy
in a different language.
Owing to the recent advent of compiler support for Fortran 2003, none of the
above tools supported object-oriented Fortran 2003 constructs as of the time of this
writing. Also, none supported distributed objects. For authors seeking to implement
the design patterns from Part II in mixed-language settings or with distributed objects, manual binding construction remains the only option. Section 11.2 discusses
a strategy for interfacing Fortran 95 and C++. Section 11.3 updates that strategy
using two open-source software projects to describe the C interoperability feature
set of Fortran 2003 and a particular SHELL pattern instantiation that arises in constructing shadow objects. The case study involves the semiautomatic construction of
object-oriented Fortran interfaces to a set of parallel, distributed C++ objects.

11.2 Manual Interoperability: C++/Fortran 95

The remainder of this chapter outlines a strategy for wrapping object-oriented C++
with object-oriented Fortran 2003 using the C interoperability constructs in Fortran

11.2 Manual Interoperability: C++/Fortran 95

Real object

Flat
interface

Language A

Shadow object

Language B

Figure 11.2. A at interface exported to a shadow object in language B by a real object in

language A.

2003. The strategy builds on and updates the approach Gray et al. (1999) outlined for
interfacing object-based Fortran 95 with object-oriented C++.Gray and colleagues
aimed to make user-dened types in Fortran available in C++ and vice versa. In their
approach, a real object in one language exports its behavior to the other language
via a at interface consisting of intrinsic types and 1D arrays thereof. The original
object exports its identity and state to the second language via a shadow object
that presents a logical interface. Figure 11.2 illustrates this architecture. Through
use of the logical interface, the shadow object may appear as a native object to a user
in language B.
Gray and colleagues distinguish an objects logical interface from its physical
interface. In their schema, constructing a physical interface requires three steps:
1. Unmangling procedure names,
2. Flattening interfaces with interoperable built-in data types,
3. Ensuring code initialization.
We explain these steps next.
Unmangling reverses the name-mangling practices compilers use to map global
identiers in source code to unique identiers in object code. The root cause for
name mangling stems from the C++ and Fortran languages allowing certain global
identiers to have the same name. For example, C++ allows two procedures to have
the same name while being distinguishable by only their signatures, and Fortran
allows different modules to have the same module procedure names. Compilers
generate these global names and pass them to linkers as entries in a symbol table.
Linkers generally reject duplicate symbols in producing the object or executable
code. Thus, a mangling scheme enables successful linking by eliminating symbol
duplication.
C++ compilers mangle all procedure names by default. In Fortran, the most
common global identiers are procedure names (external procedures or module
procedures)6 . However, Fortran restricts the name conicts on global identiers
used in a program so that only module procedures require name mangling.
Name-mangling rules are not standardized, which results in different namemangling schemes used in different compilers. Therefore, object codes produced
by different compilers are not linkable even on the same platform. C++ provides
6 Internal procedure names in Fortran are not global identiers. Also generic names in Fortran are

not global identiers.

255

256

Mixed-Language Programming
Table 11.1. Corresponding C++ and Fortran 95 types: Which
primitive C++ types maps to the given intrinsic Fortran type
depending on the platform
Type

C++

Fortran 95

Integer
Real
Character

long/int
double/oat
char

integer
real
character

the extern "C" construct to unmangle the procedure name, assuming there is no
function overloading in the program, thus enabling interfacing to C, which disallows
function overloading. In Fortran, one can use external procedures to avoid name
mangling. Since Fortran is case-insensitive, it further requires C++ to use lower
case in procedure names. Gray and colleagues took this approach to unmangle the
procedure names: Their physical interfaces consisted of a set of C++ extern "C"
procedures in lower case and Fortran 95 external procedures. Since the Fortran
2003 standard introduced features to enhance interoperability with C, we can update the Gray and colleagues unmangling scheme by requiring C bindings on Fortran
procedures in the physical interface. Section 11.3.1 describes the mechanics of this
process.
Flattening (interfaces), the second step in the Gray et al. strategy, is required for
physical interfacing through nonmember functions in C++ or external procedures in
Fortran 95. Flattening requires passing only intrinsic data types that are common
to C++ and Fortran 95. Gray et al. suggest that the rst types listed in Table 11.1
match Institute of Electrical and Electronics Engineers (IEEE) standards on most
platforms, even though the C++ and Fortran 95 standards do not guarantee matching
representations. The second types listed in the C++ column, however, might match
the corresponding type in the Fortran column on other platforms. Section 11.3.1
eliminates this ambiguity by leveraging the Fortran 2003 standards interoperable
data types.
In addition, Gray et al. limit array arguments to be contiguous, 1D arrays in
the attened physical interface because of the semantic disparity in treating arrays
between the two languages. In Fortran 95, a multidimensional array has its rigid shape
and its elements can only be accessed via array subscripts. In C++, a multidimensional
array is treated the same as an 1D array, and the elements can be accessed via a
pointer using pointer arithmetic. Furthermore, pointer arrays and assumed-shape
arrays in Fortran can represent noncontiguous portions of a data array. This concept
is not naturally supported in C++.7 Thus it is reasonable to the restrict the array
arguments to the contiguous, 1D arrays.
Since Fortran 95 passes all arguments by reference, Gray et al. further require
all C++ arguments to be passed as pointers or references. With this approach, all
parameters passed by value in the C++ logical interface must be passed by reference
in the physical interface. One benet of this calling convention is that it can be
extended to support code that conforms to earlier Fortran standards, for example
7 Currently Fortran language committee, J3, is working on a TR to provide utilities for C programmers

to access Fortrans assumed-shape and pointer arrays, and so on.

11.2 Manual Interoperability: C++/Fortran 95

Fortran 77. Nevertheless, one can eliminate the need for this convention in Fortran
2003 by leveraging that standards provision for passing by value. Section 11.3.1
explains further.
In applications, interface attening may expose safety and portability issues. One
issue Gray et al. have to address is the this pointer passed implicitly to all nonstatic
member functions. By the rules of interface attening, the this pointer has to be
passed using a reference to a long by C++ and an integer by Fortran 95 in the
physical interface. This raises concerns as to its safety and portability. To address
these issues, Gray et al. chose to use an opaque pointer in encapsulating the details
of casting between the this pointer and a long or integer formal parameter in
the calls. In addition, Gray et al. strongly recommended automating the conversion
process from the logical interface seen by end-users to its corresponding physical
interface, which is normally hidden from the same users.
Initialization of data, the last step in the Gray et al. strategy, relates to the underlying operating systems (OS) and has been a long-standing portability issue for
programs written in C, C++, and Fortran. The issue almost exclusively concerns the
extern const and static objects in C/C++, and objects with the save attribute
in Fortran. In C++ extern const and static objects can be explicitly initialized.
Similarly, a save object in Fortran can be explicitly initialized either by its declaration statement or by a separate data statement. These initialized data objects are
stored in a .data segment and are all initialized before main program starts executing. However neither C++ nor Fortran stipulate the explicit initializations for these
objects, that is, extern const, static, or save objects can be declared without being initialized in their native languages. These uninitialized objects are inserted into
a .bss segment, whose initialization is dependent on the OS. These .bss objects
may cause porting issues. For example, a programmer who solely develops on an
ELF-based OS may comfortably leave all the static variables uninitialized because
the operating system mandates data in .bss be lled up with zeros at startup time.
However, on other systems, such as XCOFF-based AIX, these objects are left with
uninitialized states.
In their strategy, Gray et al. built code using shared libraries on ELF-based systems. They reported success with this technique on Linux, Sun, and SGI platforms
but not on IBMs AIX machines.
In the Gray et al. approach, the logical interface uses the the physical interfacing techniques to enable the shadow-object relationship of Figure 11.2. The logical
interface consists of C++ class denitions or Fortran 95-derived type denitions in
a module that a programmer would see naturally in her native language. We illustrate this approach by using the astronaut example from Section 2.2. Lets assume
the astronaut class in Figure 2.4(b) has to be implemented in Fortran 95. Thus the
logical interfaces in Fortran 95 for astronaut class would be the following:
module astronaut_class
implicit none
public :: astronaut
type astronaut
private
character(maxLen), pointer :: greeting

257

258

Mixed-Language Programming

end type
interface astronaut
module procedure constructor
end interface
interface destruct
module procedure free_greeting
end interface
contains
function constructor(new_greeting) result(new_astronaut)
type(astronaut) :: new_astronaut
character(len=*), intent(in) :: new_greeting
! ...
end function
subroutine greet(this)
type(astronaut), intent(in) :: this
! ...
end subroutine
subroutine free_greeting (this)
type(astronaut), intent(inout) :: this
! ...
end subroutine
end module
To enable accessing the astronaut class in C++, one might construct the following
logical interface in C++:
class astronaut
{
public:
astronaut (const string &);
void greet() const;
astronaut();
private:
string greeting;
};
The physical interface consists of a set of procedures with at interfaces seen by
compilers but not necessarily by application programmers. The Fortran 95 physical
interface corresponding to the C++ logical interface comprises the following external
procedures:
subroutine astronaut_construct(this, new_greeting)
integer :: this
! this pointer

11.3 Case Study: ForTrilinos and CTrilinos

integer, intent(in) :: new_greeting ! points to message

! code omitted, use opaque pointers to cast this pointer ...
end function
subroutine astronaut_greet(this)
integer, intent(in) :: this
! this pointer
! code omitted, use opaque pointers to cast this pointer ...
end function
subroutine astronaut_free_greeting (this)
type(astronaut), intent(inout) :: this ! this pointer
! code omitted, use opaque pointers to cast this pointer ...
end subroutine
The following lists the C++ physical interface for the astronaut class:
// in flat interfaces parameter name "self" is used because
// "this" is a reserved keyword in C++
extern "C"
{
void astronaut_construct(long & self, long & new_greeting);
void astronaut_greet (long & self);
void astronaut_free_greeting (long & self);
}
Gray et al. has demonstrated this approach can be successfully applied to export
the behaviors of a real object from C++ to Fortran 95, and vice versa. As with any
other techniques, there are also pitfalls in the strategy. As admitted by Gray et al.,
this set of physical interfaces using long or integer to represent C pointers is not
type-safe. The outcome could be disastrous if an invalid value is passed in the call.
Gray et al. mitigated this safety risk by using an opaque pointer technique, which
insulates users from the need to do direct type casting. Today, the data safety issue
can be sufciently addressed by C-interoperability features introduced in Fortran
2003, as Section 11.3.1 discusses.
A separate issue omitted by Gray et al. relates to their method requiring special
glue-code or wrapper code by both C++ and Fortran 95 to wrap the logical interface and generate calls to the physical interfaces. The amount of work required might
limit the utility of applying their approach to the large, preexisting C++ and Fortran
95 programs that are already well tested for use within their language of origin.

11.3 Case Study: ForTrilinos and CTrilinos

This section uses examples from the ForTrilinos and CTrilinos open-source software
projects to present an updated approach to interfacing object-oriented Fortran and
C++. ForTrilinos provides Fortran interfaces to the C++ packages published by the
Trilinos project8 . CTrilinos exports C interfaces to Trilinos packages and serves as
an intermediary between ForTrilinos and other Trilinos packages.
8 http://trilinos.sandia.gov

259

260

Mixed-Language Programming

Heroux et al. (2005) described the goals, philosophy, and practices of the
Trilinos project:
The Trilinos Project is an effort to facilitate the design, development, integration, and
ongoing support of mathematical software libraries within an object-oriented framework
for the solution of large-scale, complex multiphysics engineering and scientic problems.
Trilinos addresses two fundamental issues of developing software for these problems:
(i) providing a streamlined process and set of tools for development of new algorithmic implementations and (ii) promoting interoperability of independently developed
software.
Trilinos uses a two-level software structure designed around collections of packages.
A Trilinos package is an integral unit usually developed by a small team of experts in
a particular algorithms area such as algebraic preconditioners, nonlinear solvers, etc.
Packages exist underneath the Trilinos top level, which provides a common look-and-feel,
including conguration, documentation, licensing, and bug-tracking.

With more than 50 packages in its most recent release, Trilinos might best be described as a meta-project aimed at encouraging consistently professional software
engineering practices in scalable, numerical algorithm development projects. Object
orientation plays a central role in the Trilinos design philosophy. Hence, when requests came in for Fortran interfaces to Trilinos, which is mostly written in C++, the
Trilinos developers assigned a high value to writing object-oriented interfaces while
using idioms that feel natural to Fortran programmers.
Figure 11.3 shows the global architecture of a ForTrilinos-enabled application.
The top layer comprises application code a user writes by instantiating Fortranderived type instances and invoking type-bound procedures on those instances.
These instances are lightweight in that they hold only identifying information about
the underlying Trilinos C++ objects.

Application
Object-Oriented Interface
ForTrilinos
Interface Bodies
Procedural
bindings
C Header Files
CTrilinos
External Linkage Wrappers
Trilinos

Figure 11.3. Sample ForTrilinos-enabled application architecture.

11.3 Case Study: ForTrilinos and CTrilinos

The next layer down, the ForTrilinos layer, supports a users application code
by dening public, extensible derived types arranged in hierarchies that mirror
the hierarchies in the underlying Trilinos packages. As such, the Fortran layer
reintroduces properties like polymorphism, inheritance, function overloading, and
operator overloading, each of which gets sacriced internally in the CTrilinos layer
due to the C languages lack of support for OOP. That internal sacrice ensures
portability by exploiting the C interoperability constructs in the Fortran 2003 standard to guarantee type compatibility, procedural name consistency, and compatible
argument passing conventions between Fortran and C.
Each ForTrilinos method wraps one or more CTrilinos functions and invokes
those functions via procedural bindings consisting of Fortran interface bodies and
corresponding C header les. CTrilinos produces the ForTrilinos procedural bindings automatically via scripts that parse the underlying C++ and simultaneously
generate the CTrilinos header les and the ForTrilinos interface bodies. Although
the CTrilinos scripts output C function prototypes for the header les, the actual implementation of each function is written in C++. Embedding the C++ in a
extern "C"{} constructs provides external linkage that is, the ability to link with
any C procedures that do not use function overloading.
In the bottom layer of Figure 11.3 lie the C++ class implementations dened by
various Trilinos packages. This layered design approach relieves the developers of
Trilinos numerical packages from any burden of designing for interoperability. In
this sense, the ForTrilinos/CTrilinos effort is noninvasive.
Sections 11.3.111.3.4 adapt material from Morris et al. (2010), presenting code
snippets and class diagrams from ForTrilinos and CTrilinos to demonstrate the C
interoperability features of Fortran 2003 in 11.3.1, method invocation in 11.3.2, object
construction and destruction in 11.3.3, and mirroring C++ inheritance hierarchies in
Fortran in 11.3.4.
11.3.1 C Interoperability in Fortran 2003
The C interoperability constructs of Fortran 2003 ensure the interoperability of types,
variables, and procedures. These constructs provide for referencing global data and
procedures for access by a companion processor. That companion processor need
not be a C compiler and might even be a Fortran compiler. In the case of a procedure,
for example, the Fortran standard merely constrains the procedure to be describable
by a C function prototype.
A ForTrilinos/CTrilinos procedural binding example provides a broad sampling
of the Fortran 2003 C interoperability features. For that purpose, we consider here
one of the Trilinos Petra (Greek for foundation) packages: Epetra (essential
petra). Epetra provides the objects and functionality needed for the development of
linear and nonlinear solvers in a parallel or serial computational environment. All
Trilinos packages build atop one of the Petra packages, most commonly Epetra, so
enabling Fortran developers to build Epetra objects is an essential rst step toward
enabling their use of other Trilinos packages.
The Epetra MultiVector class facilitates constructing and using dense multivectors, vectors, and matrices on parallel, distributed-memory platforms. The

261

262

Mixed-Language Programming

original C++ prototype for an Epetra_MultiVector method that lls the multi-vector
with random elements has the following prototype:
class Epetra_MultiVector {
int Random();
}
which CTrilinos scripts parse and automatically export the corresponding C wrapper
prototype:
int Epetra_MultiVector_Random (CT_Epetra_MultiVector_ID_t selfID);
as well as the corresponding ForTrilinos interface block:
1
2
3
4
5
6
7

interface
integer(c_int) function Epetra_MultiVector_Random ( selfID ) &
bind(C,name=Epetra_MultiVector_Random)
import :: c_int ,FT_Epetra_MultiVector_ID_t
type(FT_Epetra_MultiVector_ID_t), intent(in), value :: selfID
end function
end interface

where the Epetra_Multivector_Random interface body demonstrates several useful

C interoperability constructs. An examination of this example from left to right and
top to bottom drives much of the remainder of this section.
The interface block beginning and ending at lines 1 and 7, respectively, contains
one interface body. Most Trilinos interface blocks contain many more. Each Trilinos
package contains multiple classes, each of which exposes multiple methods. CTrilinos
typically generates one interface block per package. Each block contains the interface
bodies for all methods on all classes in the package. The interface body count includes
at least one per method and often more if an overloaded method name presents
multiple possible combinations of argument types, ranks, and kind parameters. The
ForTrilinos Epetra procedural bindings, for example, contain over 800 interface
bodies.
The Epetra_MultiVector_Random interface body starts at line 2 in the example.
That line species an integer return type with a c_int kind parameter that ensures
a bit representation matching the int type provided by the companion C compiler.
Fortran 2003 compilers provide an intrinsic iso_c_binding module that contains the
c_int parameter along with more than 30 other interoperable kind parameters. See
Metcalf et al. (2004) for a complete listing. Table 11.2 updates Table 11.1 with types
provided by iso_c_binding. When a companion compiler fails to supply a C type
corresponding to one of the interoperable kind parameters, the parameters take on
negative values.
ForTrilinos supplies the valid_kind_parameters() function in Figure 11.4 to
test several kind parameters and report the meanings of any negative values found.
The function returns .true. if the tested parameters are all positive and .false.
otherwise. ForTrilinos also provides the test code in Figure 11.5 to report on the

11.3 Case Study: ForTrilinos and CTrilinos

Table 11.2. Fortran 2003 kind parameters and the
corresponding interoperable C types
Fortran
Type

Fortran 2003
Kind Parameter

C type

integer
integer
real
real
character

c_int
c_long
c_double
c_oat
c_char

int
long
double
oat
char

valid_kind_parameters() result. Trilinos uses the open-source CMake9 build systems ctest facility to run tests such as that in Figure 11.5 in an automated manner,
search the results, and report successes and failures.
The statement at line 2 of the Epetra_MultiVector_Random interface body continues on line 3, where the bind attribute species that the referenced procedure
is describable by an interoperable C prototype. The name= parameter enforces Cs
case sensitivity during link-time procedure name resolution. Because it is possible
for the linked procedure to be implemented in Fortran, this also provides a mechanism for making Fortran case sensitive insofar as procedure names are concerned.
The generalize_all() procedure in Figure 11.4 provides an example of a Fortran
implementation of a bind(C) interface body dened elsewhere in ForTrilinos. It also
demonstrates a C pointer argument received as a c_ptr derived type and converted
to a Fortran pointer by the c_f_pointer() subroutine. The iso_c_binding module
provides c_ptr and c_f_pointer().
The Epetra_MultiVector_Random interface body also ensures the default C
behavior of passing of its FT_Epetra_MultiVector_ID_t argument.10 Elsewhere in
ForTrilinos appears that types denition:
type, bind(C) :: FT_Epetra_MultiVector_ID_t
private
integer(ForTrilinos_Table_ID_t) :: table
integer(c_int) :: index
integer(FT_boolean_t) :: is_const
end type
where the bind(C) attribute ensures interoperability with the CTrilinos struct:
typedef struct {
CTrilinos_Table_ID_t table; /* Table of object references */
int index;
/* Array index of the object */
boolean is_const;
/* Whether object was declared const */
} CT_Epetra_MultiVector_ID_t;
9 http://www.cmake.org
10 Although not specically mandated by any Fortran standard, the default behavior of most For-

tran compilers resembles passing arguments by reference, which provides an objects address to
the receiving procedure, thereby allowing the receiving procedure to modify the object directly.
An example of an alternative approach that simulates the same behavior is the copy-in/copy-out
technique that allows for modifying the passed entity without passing its address.

263

264

Mixed-Language Programming

1
2
3
4
5
6
7

Figure 11.4
module fortrilinos_utils
implicit none
private
public :: length
public :: generalize_all
public :: valid_kind_parameters
contains

8
9
10
11
12
13
14
15

!
!
!
!
!
!
!

Fortran-style strings take two forms: one with no C counterpart and

a second that is an array of character values. A C string is interoperable with the second type of Fortran string with the primary
distinction being that Fortran strings carry their own string-length
information, whereas the length of C strings is indicated by a
terminating null character. The "count" procedure calculates C
string length by searching for the null terminator.

16
17
18
19
20
21
22
23
24
25

integer(c_int) function length(c_string) bind(C,name="forLinkingOnly")

use ,intrinsic :: iso_c_binding ,only: c_char ,c_int ,c_null_char
character(kind=c_char) :: c_string(*)
length = 0
do
if(c_string(length+1) == c_null_char) return
length = length + 1
end do
end function

26
27
28
29
30
31

!
!
!
!
!

This is a Fortran implementation of the functionality in the

Epetra_*_Abstract procedures in each CTrilinos/src/CEpetra* file.
It effectively casts any given Epetra derived type to a general type
that can represent any of the Epetra derived types in
ForTrilinos_enums.F90.

32

type(ForTrilinos_Universal_ID_t) function generalize_all(object_id) &

bind(C,name="for_linking_only")
use ForTrilinos_enums ,only : ForTrilinos_Object_ID_t
use ,intrinsic :: iso_c_binding ,only: c_ptr,c_f_pointer
type(c_ptr) ,value :: object_id
type(ForTrilinos_Universal_ID_t), pointer :: local_ptr

33
34
35
36
37
38
39

call c_f_pointer (object_id, local_ptr)

generalize_all = local_ptr
end function

40
41
42
43
44
45
46
47
48
49
50
51
52
53
54

!
!
!
!
!
!
!
!
!
!
!

This procedure checks the values of parameters required to interoperate with CTrilinos. The Fortran 2003 standard requires that these
parameters be defined in the intrinsic module iso_c_binding with
values that communicate meanings specified in the standard. This
procedures quoted interpretations of these values are largely excerpted from the standard. This procedure returns true if all of
the interoperating Fortran kind parameters required by ForTrilinos
have a corresponding C type defined by the companion C processor.
Otherwise, it returns false. (For purposes of ForTrilinos, the
Fortran standards use of the word processor is interpreted as
denoting the combination of a compiler, an operating system, and

11.3 Case Study: ForTrilinos and CTrilinos

55

! a hardware architecture.)

56
57
58
59
60
61
62

logical function valid_kind_parameters(verbose)

use ,intrinsic :: iso_c_binding ,only : &
c_int,c_char,c_double,c_ptr,c_long,c_bool,c_null_char
use ,intrinsic :: iso_fortran_env ,only : error_unit ,output_unit
logical ,optional :: verbose
logical
:: verbose_output

63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82

character(len=*) ,parameter :: no_fortran_kind= &

The companion C processor defines the corresponding C type, &
but there is no interoperating Fortran processor kind.
character(len=*) ,parameter :: no_c_type= &
The C processor does not define the corresponding C type.
character(len=*) ,parameter :: interoperable = &
This interoperating Fortran kind has a corresponding C type.
character(len=*) ,parameter :: imprecise= &
The C processor type does not have a precision equal to the &
precision of any of the Fortran processor real kinds.
character(len=*) ,parameter :: limited= &
The C processor type does not have a range equal to the range &
of any of the Fortran processor real kinds.
character(len=*) ,parameter :: limited_and_imprecise= &
The C processor type has neither the precision nor range of any &
of the Fortran processor real kinds.
character(len=*) ,parameter :: not_interoperable_nonspecific = &
There is no interoperating Fortran processor kind for &
unspecified reasons.

83
84

valid_kind_parameters = .true. ! default return value

85
86
87
88
89
90

if (present(verbose)) then
verbose_output=verbose
else
verbose_output=.false.
end if

91
92
93
94
95
96
97
98
99
100
101
102

select case(c_long)
case(-1)
write(error_unit ,fmt=(2a)) c_long error: ,no_fortran_kind
valid_kind_parameters = .false.
case(-2)
write(error_unit ,fmt=(2a)) c_long error: ,no_c_type
valid_kind_parameters = .false.
case default
if (verbose_output) write(output_unit,fmt=(2a)) &
c_long: ,interoperable
end select

103
104
105
106
107
108
109

select case(c_double)
case(-1)
write(error_unit ,fmt=(2a)) c_double error: ,imprecise
valid_kind_parameters = .false.
case(-2)
write(error_unit ,fmt=(2a)) c_double error: ,limited

265

266

Mixed-Language Programming
110
111
112
113
114
115
116
117
118
119
120
121
122

valid_kind_parameters = .false.
case(-3)
write(error_unit ,fmt=(2a)) &
c_double error: ,limited_and_imprecise
valid_kind_parameters = .false.
case(-4)
write(error_unit ,fmt=(2a)) &
c_double error: ,not_interoperable_nonspecific
valid_kind_parameters = .false.
case default
if (verbose_output) write(output_unit,fmt=(2a)) &
c_double: ,interoperable
end select

123
124
125
126
127
128
129
130
131
132

select case(c_bool)
case(-1)
write(error_unit ,fmt=(a)) c_bool error: invalid value for &
a logical kind parameter on the processor.
valid_kind_parameters = .false.
case default
if (verbose_output) write(output_unit ,fmt=(a)) c_bool: &
valid value for a logical kind parameter on the processor.
end select

133
134
135
136
137
138
139
140
141
142
143
144

select case(c_char)
case(-1)
write(error_unit ,fmt=(a)) c_char error: invalid value for &
a character kind parameter on the processor.
valid_kind_parameters = .false.
case default
if (verbose_output) write(output_unit ,fmt=(a)) c_char: &
valid value for a character kind parameter on the processor.
end select
end function
end module fortrilinos_utils
Figure 11.4. ForTrilinos utilities.

1
2
3
4
5
6
7
8
9
10
11
12

program main
use fortrilinos_utils ,only : valid_kind_parameters
use iso_fortran_env
,only : error_unit ,output_unit
implicit none
if (valid_kind_parameters(verbose=.true.)) then
write(output_unit,*)
write(output_unit,fmt=(a)) "End Result: TEST PASSED"
else
write(error_unit,*)
write(error_unit,fmt=(a)) "End Result: TEST FAILED"
end if
end program
Figure 11.5. ForTrilinos interoperability test.

11.3 Case Study: ForTrilinos and CTrilinos

where table identies which of several tables holds a reference to the underlying
Trilinos C++ object and index species which entry in the table refers to that object.
CTrilinos constructs one table for each class in the packages it wraps, and it stores
all instances of that class in the corresponding table.
In the CT_Epetra_MultiVector_ID_t struct, the table member is dened as an
enumerated type of the form:
typedef enum {
CT_Invalid_ID,
/*does not reference a valid table entry*/
/* lines omitted */
CT_Epetra_MultiVector_ID/*references Epetra_MultiVector entry*/
/* lines omitted */
} CTrilinos_Table_ID_t;
which aliases the int type and denes a list of int values ordered in unit increments
from CT_Invalid_ID=0. Likewise, ForTrilinos exploits the Fortran 2003 facility for
dening an interoperable list of enumerated integers according to:
enum ,bind(C)
enumerator ::
&
FT_Invalid_ID,
&
! lines omitted
FT_Epetra_MultiVector_ID, &
! lines omitted
end enum
which denes the list but does not alias the associated type. Since Fortran provides no mechanism aliasing intrinsic types, ForTrilinos employs the nearest Fortran
counterpart: an alias for the relevant kind parameter:
integer(kind(c_int)) ,parameter :: ForTrilinos_Table_ID_t=c_int
which takes on the same kind and value as the c_int parameter itself.
The nal FT_Epetra_MultiVector_ID_t component is_const, a Boolean
value, allows for marking structs that cannot be modied because they were declared as const. Since Cs support for a Boolean type only entered the language in
its most recent standard (C99), CTrilinos denes its own integer Boolean type for use
with earlier C compilers. ForTrilinos denes a corresponding FT_boolean_t integer
kind parameter and employs it in dening its multi-vector type.
CTrilinos reserves exclusive rights to modify its struct members even though
they have an interoperable representation in ForTrilinos. ForTrilinos honors these
rights by giving the FT_Epetra_MultiVector_ID_t components private scope within
a module that contains only type denitions and parameter initializations.
Finally, line 4 of the Epetra_MultiVector_Random interface body imports entities dened in the host module scoping unit. The Fortran standard does not make
names from a host scoping unit available within an interface body. Hence, although
each ForTrilinos procedural binding module contains use statements of the form:
use ,iso_c_binding ,only : c_int
use ,ForTrilinos_enums ,only : FT_Epetra_MultiVector_ID_t

267

268

Mixed-Language Programming
Table 11.3. Dummy array declaration interoperability
(N/A=none allowed): the interoperability restrictions apply to
the dummy (received) array, not to the actual (passed) array

Array category

Fortran declaration

Interoperable
C declaration

Explicit-shape
Assumed-size
Allocatable
Pointer
Assumed-shape

real(c_oat) :: a(10,15)
real(c_oat) :: b(10,*)
real(c_oat),allocatable:: c(:,:)
real(c_oat),pointer:: d(:,:)
real(c_oat) :: e(:,:)

oat a[15][10];
oat b[ ][10];
N/A
N/A
N/A

the entities these statements make accessible to the surrounding module are not automatically accessible within the interface bodies in that module. The import statement
brings the named entities into the interface body scoping unit. This completes the
discussion of the Epetra_MultiVector_Random binding.
An important interoperability scenario not considered up to this point involves
the interoperability of array arguments. Whereas Gray et al. (1999) mapped all array arguments to their attened, 1D equivalents, Fortran 2003 facilitates passing
multidimensional arrays. Standard-compliant compilers must copy passed data into
and out of bind(C) procedures in a manner that ensures contiguous memory on
the receiving side even when, for example, the actual argument passed is an array
pointer that references a noncontiguous subsection of a larger array. The restrictions
in doing so relate primarily to how the interface body for receiving procedure must
declare its array arguments. Only arrays declared on the receiving end as explicitshape or assumed-size are interoperable. Table 11.3 delineates the interoperable and
noninteroperable declarations.
An important special case concern character array arguments. Metcalf et al. (2004)
state:
In the case of default character type, agreement of character length is not required . . .
In Fortran 2003, the actual argument corresponding to a default character (or with kind
[c_char]) may be any scalar, not just an array element or substring thereof; this case is
treated as if the actual argument were an array of length 1 . . .

where the bracketed text corrects a typographical error in the original text. Passing
a character actual argument represents a special case of what Fortran 2003 terms
sequence association, which is an exception to the array shape rules that otherwise
apply. The standard allows sequence association only when the dummy argument
(declared in the receiving procedure interface) references contiguous memory that
is an explicit-shape or assumed-size arrays. In such cases, the actual (passed) argument need not have a rank and a shape matching those of the dummy (received)
argument. The actual argument rank and shape only need to provide at least as many
elements as required by the dummy argument.
As an extreme scenario, an actual argument can even be an array element, in
which case the dummy takes the rest of the array as input. Although originally

11.3 Case Study: ForTrilinos and CTrilinos

intended to provide backward compatibility with Fortran 90, which introduced

assumed-shape arrays, this feature now nds new legs in a C-interoperability world.
Figure 11.4 denes a length() procedure that makes use of the aforementioned rank and shape exibility. That procedure returns the character count of
null-terminated C strings. The calling procedure need not pass an actual argument
with a rank and a shape matching those of the dummy argument c_string.

11.3.2 Method Invocation

Trilinos exploits several C++ capabilities not supported by C. Among these are object
method invocation; inheritance; polymorphism; overloaded function names and operators; default argument values; class and function templates; string and bool types;
exception handling; and safe type-casting. Consequently, the C interface to Trilinos
proves rather awkward from a C++ viewpoint. The developers pay this price for
standards conformance but hide the inelegant features in the intermediate layers of
Figure 11.3. We describe next how CTrilinos copes with the missing C++ capabilities
and ForTrilinos restores similar capabilities using natural Fortran forms.
Within C++, one can dereference a pointer to an object to invoke a method on
the object directly. A pointer mv to an Epetra_MultiVector, for example, can be
used to invoke the Random() method as in the following pseudocode:
Epetra_MultiVector *mv = new Epetra_MultiVector(...);
mv->Random();
// arguments omitted above
wherein the ellipses stand in for arguments not shown in this snippet. In implementing
CEpetra, the CTrilinos wrapper for Epetra, however, a dilemma arises from Cs
ability to receive, pass, and return pointers to C++ classes but Cs inability to directly
invoke methods on the object the pointer references. One could resolve this dilemma
by wrapping the above code in C++ helper functions and embedding such functions in
an externC{} construct. That approach, however, would require manipulating
raw pointers throughout CTrilinos, leaving the code vulnerable to a variety of pitfalls
ranging from dangling pointers to memory leaks.
To circumvent the pitfalls associated with the direct manipulation of pointers,
CTrilinos grants users no direct access to the object pointer. Instead, users reference
objects by a handle that the developers refer to as a struct identication tag, or a struct
ID,. The aforementioned CT_Epetra_MultiVector_ID_t exemplies a struct ID.
Struct IDs contain identifying information that users should never directly modify.
Given a struct ID mv_id, the multivector randomize method invocation of
section 11.3.1 becomes:
Epetra_MultiVector_Random(mv_id);
Likewise, although the Fortran 2003 standard denes a c_ptr derived type specically for passing C pointers between Fortran and C, ForTrilinos accesses Trilinos via
CTrilinos, so Fortran users never directly manipulate the underlying Trilinos object
pointers.
ForTrilinos wraps each struct ID in a SHELL, names the shell type after the
underlying Trilinos class, and publishes type-bound procedures named after the

269

270

Mixed-Language Programming

type. Figure 11.6 illustrates the Fortran Epetra multivector construction code and
randomization method call is thus:
type(epetra_multivector) :: mv
mv = epetra_multivector(...)
call mv%Random()
which is the most natural Fortran equivalent to the original C++, and where ellipses
again mark arguments omitted from this pseudocode.
CTrilinos tabulates each struct a CTrilinos function returns. Recording the struct
in a table facilitates passing it into the appropriate destruction function, which
deletes the underlying object and its struct ID. Compile-time type checking is enforced by providing each wrapped class with its own custom (but internally identical)
struct ID.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24

Figure 11.6
module Epetra_MultiVector_module ! ---------- Excerpt -------------use ForTrilinos_enums ,only: FT_Epetra_MultiVector_ID_t,&
FT_Epetra_Map_ID_t,ForTrilinos_Universal_ID_t
use ForTrilinos_table_man
use ForTrilinos_universal,only:universal
use ForTrilinos_error
use FEpetra_BlockMap ,only: Epetra_BlockMap
use iso_c_binding
,only: c_int,c_double,c_char
use forepetra
implicit none
private
! Hide everything by default
public :: Epetra_MultiVector ! Expose type/constructors/methods
type ,extends(universal)
:: Epetra_MultiVector !"shell"
private
type(FT_Epetra_MultiVector_ID_t) :: MultiVector_id
contains
procedure :: cpp_delete => ctrilinos_delete_EpetraMultiVector
procedure
:: get_EpetraMultiVector_ID
procedure ,nopass :: alias_EpetraMultiVector_ID
procedure
:: generalize
procedure
:: Random
procedure
:: Norm2
procedure
:: NumVectors
end type

25
26
27
28

interface Epetra_MultiVector ! constructors

module procedure from_scratch,duplicate,from_struct
end interface

29
30

contains

31
32
33
34
35
36
37

type(Epetra_MultiVector) function from_struct(id)

type(FT_Epetra_MultiVector_ID_t) ,intent(in) :: id
from_struct%MultiVector_id = id
call from_struct%register_self
end function

11.3 Case Study: ForTrilinos and CTrilinos

38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54

type(Epetra_MultiVector) function from_scratch( &

BlockMap,Num_Vectors,zero &
)
use ForTrilinos_enums ,only: FT_boolean_t,FT_TRUE,FT_FALSE
use iso_c_binding
,only: c_int
class(Epetra_BlockMap) ,intent(in) :: BlockMap
integer(c_int)
,intent(in) :: Num_Vectors
logical
,intent(in) :: zero
integer(FT_boolean_t)
:: zero_in
type(FT_Epetra_MultiVector_ID_t)
:: from_scratch_id
if (zero) zero_in=FT_TRUE
if (.not.zero) zero_in=FT_FALSE
from_scratch_id = Epetra_MultiVector_Create( &
BlockMap%get_EpetraBlockMap_ID(),Num_Vectors,zero_in &
)
from_scratch = from_struct(from_scratch_id)
end function

55
56
57
58
59
60
61

type(Epetra_MultiVector) function duplicate(this)

type(Epetra_MultiVector) ,intent(in) :: this
type(FT_Epetra_MultiVector_ID_t) :: duplicate_id
duplicate_id = Epetra_MultiVector_Duplicate(this%MultiVector_id)
duplicate = from_struct(duplicate_id)
end function

62
63
64
65
66
67
68

type(FT_Epetra_MultiVector_ID_t) function get_EpetraMultiVector_ID( &

this &
)
class(Epetra_MultiVector) ,intent(in) :: this
get_EpetraMultiVector_ID=this%MultiVector_id
end function

69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89

type(FT_Epetra_MultiVector_ID_t) function alias_EpetraMultiVector_ID( &

generic_id &
)
use ForTrilinos_table_man,only: CT_Alias
use iso_c_binding
,only: c_loc,c_int
use ForTrilinos_enums
,only: ForTrilinos_Universal_ID_t, &
FT_Epetra_MultiVector_ID
type(ForTrilinos_Universal_ID_t) ,intent(in) :: generic_id
type(ForTrilinos_Universal_ID_t) ,allocatable ,target :: alias_id
integer(c_int) :: status
type(error) :: ierr
allocate(alias_id,source=CT_Alias( &
generic_id,FT_Epetra_MultiVector_ID),stat=status &
)
ierr=error(status,FEpetra_MultiVector:alias_EpetraMultiVector_ID)
call ierr%check_success()
alias_EpetraMultiVector_ID=degeneralize_EpetraMultiVector( &
c_loc(alias_id) &
)
end function

90
91
92

type(ForTrilinos_Universal_ID_t) function generalize(this)

use ForTrilinos_utils ,only: generalize_all

271

272

Mixed-Language Programming
93
94
95
96

use iso_c_binding
,only : c_loc
class(Epetra_MultiVector) ,intent(in) ,target :: this
generalize = generalize_all(c_loc(this%MultiVector_id))
end function

97
98
99
100
101
102
103
104
105
106
107

type(FT_Epetra_MultiVector_ID_t) function &

degeneralize_EpetraMultiVector(generic_id) bind(C)
use ForTrilinos_enums ,only : ForTrilinos_Universal_ID_t, &
FT_Epetra_MultiVector_ID_t
use ,intrinsic :: iso_c_binding ,only: c_ptr,c_f_pointer
type(c_ptr)
,value
:: generic_id
type(FT_Epetra_MultiVector_ID_t) ,pointer :: local_ptr=>null()
call c_f_pointer (generic_id, local_ptr)
degeneralize_EpetraMultiVector = local_ptr
end function

108
109
110
111
112
113
114
115

subroutine Random(this,err)
class(Epetra_MultiVector) ,intent(inout) :: this
type(error)
,optional ,intent(out)
:: err
integer(c_int)
:: error_out
error_out = Epetra_MultiVector_Random (this%MultiVector_id)
if (present(err)) err=error(error_out)
end subroutine

116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131

function Norm2(this,err) result(Norm2_val)

class(Epetra_MultiVector)
,intent(in) :: this
type(error)
,optional
,intent(out) :: err
real(c_double) ,dimension(:),allocatable :: Norm2_val
integer(c_int)
:: error_out
integer(c_int)
:: status
type(error)
:: ierr
if (.not.allocated(Norm2_val)) then
allocate(Norm2_val(this%NumVectors()),stat=status)
ierr=error(status,FEpetra_MultiVector:Norm2)
call ierr%check_success()
endif
error_out = Epetra_MultiVector_Norm2(this%MultiVector_id,Norm2_val)
if (present(err)) err=error(error_out)
end function

132
133
134
135
136

integer(c_int) function NumVectors(this)

class(Epetra_MultiVector) ,intent(in) :: this
NumVectors=Epetra_MultiVector_NumVectors(this%MultiVector_id)
end function

137
138
139
140
141

subroutine ctrilinos_delete_EpetraMultiVector(this)
class(Epetra_MultiVector),intent(inout) :: this
call Epetra_MultiVector_Destroy( this%MultiVector_id )
end subroutine

142
143

end module

144

Figure 11.6. ForTrilinos MultiVector class excerpt.

11.3 Case Study: ForTrilinos and CTrilinos

11.3.3 Tabulation, Construction, and Destruction

CTrilinos wraps constructors for each supported class. Every wrapped class is associated with a pair of tables of templated class CTrilinos_Table. Each table entry
is a RCP to an object created by the Trilinos C++ code. The reference-counting
ability does not extend beyond CTrilinos and cannot be utilized by applications.
Application code written in C must use a type-specic struct ID, such as the aforementioned CT_Epetra_MultiVector_ID_t, to track an object via its table entry.
Each type-specic struct ID has an identical form to the following generic struct ID:
typedef struct {
CTrilinos_Table_ID_t table;
int index;
boolean is_const;
} CTrilinos_Universal_ID_t;
where the struct members have the same meanings as in the aforementioned
Epetra_MultiVector_ID_t type and where ForTrilinos denes a corresponding
interoperable ForTrilinos_Universal_ID_t. Application code written in Fortran would use the extensible, native derived type rather than the interoperable
struct ID.
The CTrilinos table design prevents catastrophic runtime errors by catching
attempts to access invalid table entries. Although CTrilinos uses RCPs to decide
when to free the underlying object, a reference within a table will prevent the reference count from hitting zero; therefore, a CTrilinos user must explicitly remove
objects from the tables when they are no longer needed so that the underlying object
can be destroyed. CTrilinos assumes ownership of the lifetimes of objects created
with wrapped constructors. CTrilinos destroys such an object once no table entries
reference the object.
For every class CTrilinos wraps, it publishes a destruction procedure that can be
invoked in the manner:
Epetra_MultiVector_Destroy(&mv_id);
Such a procedure removes the object reference from the specied table entry and then
destroys the C++ object if there are no remaining references to it in the table. Since
the struct ID is the only way CTrilinos users can access the C++ object, prevention
of memory leaks necessitates that explicit invocation of the destruction procedure
before the object goes out of scope.
As a direct consequence of supporting object access only through type-specic
tables, error-free use of CTrilinos requires that application code never duplicate
a struct ID or pass the ID by value to a function that requests the destruction of
the underlying object. Were a user to duplicate an ID and later pass it to a destroy
function, the one remaining table entry associated with the original object would refer
to a nonexistent object. Future use of that table entry would result in unpredictable
behavior and potential access violations. Certain procedures, however, may require
creating a new ID struct for an object. In such cases, both struct IDs should be deleted
once the underlying object has been destroyed.

273

274

Mixed-Language Programming

ForTrilinos automates the CTrilinos destruction call. This automation requires

navigating several subtleties arising from the interplay between multiple languages.
These subtleties crop up even in the simple construction pseudocode on page 270, for
example. In the statement mv=Epetra_MultiVector(...), the constructor returns
a new object that becomes the RHS argument of a dened assignment. When the
assignment terminates, its RHS goes out of scope with no remaining references
to it, at which time a standard-compliant compiler must invoke the corresponding
objects nal subroutine. ForTrilinos nal subroutines invoke CTrilinos destruction
functions, which in turn delete the underlying Trilinos C++ object and any related
table entries. This cascade of calls catastrophically impacts future use of the LHS
argument into which the assignment would have just copied the RHS information,
causing an existential crisis: The construction process destroys the constructed object!
One way to circumvent premature destruction relies upon counting references to
ForTrilinos objects so the destruction of Trilinos C++ objects occurs only after all the
corresponding ForTrilinos objects are out of scope. Figure 11.7 shows the UML class
diagram for a ForTrilinos OBJECT hierarchy that provides such a service. Figure 11.8
shows the implementation. All ForTrilinos types extend the shown Universal class,
which aggregates a Ref_Counter named ref_count. The ref_count component
serves a bookkeeping role for all objects that share the same CTrilinos struct ID.
The Hermetic class serves as a SURROGATE for Universal when aggregated into the
Ref_Counter.
Similar to reference counting in C++, reference counting in ForTrilinos relies on
a language capability to override the behaviors of object construction, assignment,
and destruction provided by compilers. First, an object requiring a unique struct ID
has to be constructed using ref_counts overloaded constructor in order to participate in reference counting. Also, a type-bound assignment overrides the default
assignment, so a reference-counted LHS object does not automatically get nalized
on object assignment. Finally, the nalizer of this type calls release() to decrease
the reference count by one.

Hermetic
+ cpp_delete ()

Ref_Counter

Universal
counter : Ref_Counter
+ register_self ()
+ force_finalize ()

count : Pointer<Integer>
obj : Pointer<Hermetic>
<<reference count>>
+ grab ()
{if this%count=0,
+ release ()
call obj%cpp_delete()}
1 <<assignment (=)>>
+ assign ()
{call lhs%release() ...
<< constructor >>
call lhs%grab()}
+ ref_counter ()
<< destructor >>

+ finalize_ref_counter ()

{call this%release()}

Figure 11.7. ForTrilinos OBJECT hierarchy, including the Hermetic SURROGATE.

11.3 Case Study: ForTrilinos and CTrilinos

The grab() and release() methods respectively increment and decrement the
reference count. When the count reaches zero, release() calls the CTrilinos destruction procedure for the underlying C++ object. To achieve this, ref_counter
keeps a reference to the object using the obj component, which is of Hermetic type.
Hermetic provides a hook for ref_counter to use to invoke the CTrilinos destruction
functions. As shown in Figure 11.8(b), the abstract Hermetic class only denes the
interface for invoking the CTrilinos destruction procedures. Hermetic delegates to
subclasses the implementation of that interface, including the CTrilinos destruction
function invocation.
Figure 11.8(c) provides the Universal implementation. All ForTrilinos objects
must invoke their inherited register_self() method to register themselves to participate in reference counting before they can be used that is, after instantiation
of a Trilinos C++ object and its ForTrilinos shadow and assignment of the shadows
struct ID, the shadow object must call register_self(). Afterward, the shadow can
be used freely in assignments and other operations during which all new references
to this object are counted automatically by ref_counter. Universal also provides
a force_finalize() method for programmers to manually invoke the release()
method of ref_counter. This method is needed only for ForTrilinos objects created in the main program or with the save attribute, for which the language does not
mandate automatic object nalizations and the programmer must therefore manually
release the objects memory.
Fortran provides no facility for overriding its default copying behavior.11 The
aforementioned reference-counting scheme therefore requires no object copying.
Thus, ForTrilinos objects cannot be used in places, such as array constructors, where
object copying is required.
The need for reference counting in Fortran is unique to the mixed-language
programming paradox: An objects creation and deletion in Fortran is constrained
by the underlying C++ objects creation and deletion. In the vast majority of cases,
developers writing solely in Fortran avoid object creation, deletion, and nalization concerns altogether by using allocatable components instead of pointers and
shadows. This practice obligates standard-compliant compilers to automatically deallocate the memory associated with such components when the encompassing object
goes out of scope.

1
2
3
4
5
6
7
8
9
10

Figure 11.8(a):ref_count
module ref_counter_module
use hermetic_module, only : hermetic
private
public :: ref_counter
type ref_counter
private
integer, pointer :: count => null()
class(hermetic), pointer :: obj => null()
contains
procedure, non_overridable :: grab
11 Fortran 2003 provides no facility analogous to C++ copy constructors.

275

276

Mixed-Language Programming
11
12
13
14
15

procedure, non_overridable :: release

procedure :: assign
final :: finalize_ref_counter
generic :: assignment(=) => assign
end type

16
17
18
19

interface ref_counter
module procedure constructor
end interface

20
21

contains

22
23
24
25
26
27
28
29
30

subroutine grab(this)
class(ref_counter), intent(inout) :: this
if (associated(this%count)) then
this%count = this%count + 1
else
stop Error in grab: count not associated
end if
end subroutine

31
32
33
34
35

subroutine release(this)
class (ref_counter), intent(inout) :: this
if (associated(this%count)) then
this%count = this%count - 1

36
37
38
39
40
41
42
43
44

else

if (this%count == 0) then
call this%obj%cpp_delete
deallocate (this%count, this%obj)
end if

stop Error in release: count not associated

end if
end subroutine

45
46
47
48

subroutine assign (lhs, rhs)

class (ref_counter), intent(inout) :: lhs
class (ref_counter), intent(in) :: rhs

49
50
51
52
53

lhs%count => rhs%count

lhs%obj => rhs%obj
call lhs%grab
end subroutine

54
55
56
57
58

recursive subroutine finalize_ref_counter (this)

type(ref_counter), intent(inout) :: this
if (associated(this%count)) call this%release
end subroutine

59
60
61
62
63
64
65

function constructor (object)

class(hermetic), intent(in) :: object
type(ref_counter), allocatable :: constructor
allocate (constructor)
allocate (constructor%count, source=0)
allocate (constructor%obj, source=object)

11.3 Case Study: ForTrilinos and CTrilinos

66
67
68

1
2
3
4
5
6
7

call constructor%grab
end function
end module

Figure 11.8(b):hermetic
module hermetic_module
private
public :: hermetic
type, abstract :: hermetic
contains
procedure(free_memory), deferred :: cpp_delete
end type

8
9
10
11
12
13
14
15

1
2
3
4

abstract interface
subroutine free_memory (this)
import
class(hermetic), intent(inout) :: this
end subroutine
end interface
end module

Figure 11.8(c):universal
module universal_module
use hermetic_module ,only : hermetic
use ref_counter_module, only : ref_counter
implicit none

5
6
7
8

type ,abstract ,extends(hermetic) :: universal

private
type(ref_counter) :: counter

9
10
11
12
13

contains
procedure, non_overridable :: force_finalize
procedure, non_overridable :: register_self
end type

14
15

contains

16
17
18

subroutine force_finalize (this)

class(universal), intent(inout) :: this

19
20
21

call this%counter%release
end subroutine

22
23
24

subroutine register_self (this)

class(universal), intent(inout) :: this

25
26
27
28

this%counter = ref_counter(this)
end subroutine
end module
Figure 11.8. Implementation of reference-counted ForTrilinos OBJECT hierarchy.

277

278

Mixed-Language Programming

11.3.4 Polymorphism and Hierarchy Mirroring

Whereas C supports neither polymorphism nor inheritance, Trilinos employs these
two technologies ubiquitously. For example, each Trilinos class has one overloaded
constructor with an average of six different implementations. In CTrilinos, all
wrapped overloaded functions must bear a unique name.
Life without inheritance is nontrivial. Without C support for the inheritance
concept, a separate wrapper must be created for each method a child class inherits
from its parent classes. For example, a wrapper for the method:
int Epetra_MultiVector::Norm2(double * Result) const;
will invoke the Norm2() method on an Epetra_MultiVector instance only. Even
though Trilinos publishes an Epetra_Vector class that inherits this method, a
separate wrapper must be written for Epetra_Vector objects.
As another example, a separate wrapper must be created for each combination of argument classes that could result when a child instance is passed
where the function prototype species a parent instance. Consider the class hierarchy in Figure 11.9 showing that Epetra_RowMatrix has four child classes,
including Epetra_BasicRowMatrix, Epetra_CrsMatrix, Epetra_MsrMatrix, and
Epetra_VbrMatrix. Three of these classes also have children. In C++, one can invoke a hypothetical method involving two different Epetra_RowMatrix children as
follows:
Epetra_CrsMatrix *A = new Epetra_CrsMatrix(...);
Epetra_JadMatrix *B = new Epetra_JadMatrix(...);
A->TwoRowMatrixOp(B);
A particular method of the Epetra_RowMatrix class can take as an argument an instance of any of class that implements the Epetra_RowMatrix interface. Therefore, in
the specic case of an Epetra_CrsMatrix and Epetra_JadMatrix, the corresponding

Epetra_RowMatrix

Epetra_CrsMatrix

Epetra_MsrMatrix

Epetra_VbrMatrix

Epetra_BasicRowMatrix

Epetra_FECrsMatrix
Epetra_JadMatrix

Epetra_OskiMatrix

Epetra_VbrRowMatrix

Figure 11.9. Epetra_Row Matrix class hierarchy.

Epetra_FEVbrMatrix

11.3 Case Study: ForTrilinos and CTrilinos

C code would be:

CT_Epetra_CrsMatrix_ID_t A;
CT_Epetra_JadMatrix_ID_t B;
A = Epetra_CrsMatrix_Create(...);
B = Epetra_JadMatrix_Create(...);
Epetra_RowMatrix_TwoRowMatrixOp_Crs_Jad(A,B);
This simple example requires creating 36 wrappers with unique names in order to
allow for all Epetra_RowMatrix methods with all possible combinations of types that
derive from Epetra_RowMatrix. CTrilinos manages inheritance by dening a set of
union types corresponding to each of the struct ID types. Within each union, there is a
member corresponding to itself, its generic, and each class above it in the inheritance
hierarchy. Following this convention, CTrilinos denes the exible union ID for the
Epetra_CrsMatrix class as follows:
typedef union{
CTrilinos_Universal_ID_t universal;
CT_Epetra_CrsMatrix_ID_t Epetra_CrsMatrix;
CT_Epetra_BLAS_ID_t Epetra_BLAS;
CT_Epetra_CompObject_ID_t Epetra_CompObject;
CT_Epetra_DistObject_ID_t Epetra_DistObject;
CT_Epetra_Object_ID_t Epetra_Object;
CT_Epetra_Operator_ID_t Epetra_Operator;
CT_Epetra_RowMatrix_ID_t EPetra_RowMatrix;
CT_Epetra_SrcDistObject_ID_t Epetra_SrcDistObject;
} CT_Epetra_CrsMatrix_ID_Flex_t;
This union type allows CTrilinos to invoke the previously discussed method as:
CT_Epetra_CrsMatrix_ID_Flex_t A;
CT_Epetra_JadMatrix_ID_Flex_t B;
A.Epetra_CrsMatrix = Epetra_CrsMatrix_Create(...);
B.Epetra_JadMatrix = Epetra_JadMatrix_Create(...);
Epetra_RowMatrix_TwoRowMatrixOp(
A.Epetra_RowMatrix,B.Epetra_RowMatrix
);
where only one version of the wrapper is needed. This allows safe struct ID type
conversion without using explicit type casting in C.
ForTrilinos reintroduces the inheritance and polymorphism that most Trilinos
packages employ but CTrilinos cannot. The ForTrilinos Epetra_Vector wrapper,
for example, extends the Epetra_MultiVector derived type and represents the
special case of a multivector comprised of one vector. Figure 11.10 excerpts the
Epetra_Vector implementation.
The Epetra_Vector excerpt shows several features common to all ForTrilinos derived types, including the procedures generalize(), degeneralize_*(),
get_*_ID(), and alias_*_ID(), where the asterisks substitute for derived type

279

280

Mixed-Language Programming

names. The object-construction process provides a useful guide to the use of each of
these procedures. Consider the following pseudocode:
1
2
3
4
5
6

program main
use Epetra_Vector_module, only : Epetra_Vector
implicit none
type(Epetra_Vector) :: A
A = Epetra_Vector(...) ! arguments omitted
end main

where the structure constructor invoked on the RHS of line 5 returns a new
Epetra_Vector which has a new Fortran struct ID tag for the underlying C++
object that the ForTrilinos constructor directs CTrilinos to create at lines 5557
in Figure 11.10. The new struct ID is also passed to the structure constructor of
Epetra_Vector, from_struct, at line 58 to create a FT_Epetra_MultiVector_ID_t
alias struct ID and construct the new Epetra_Vector objects Epetra_MultiVector
parent component at lines 3537 in Figure 11.10. Finally, the intrinsic assignment
assigns the new Epetra_Vector object to LHS and invokes the dened assignment
in the ref_counter parents component. This completes the construction process.
Creation of the alias struct ID proceeds in three steps. First, starting with
the innermost invocation on lines 3537 of Figure 11.10, generalize() converts
the FT_Epetra_Vector_ID_t struct to the generic ForTrilinos_Universal_ID_t.
Second, alias_EpetraMultiVector_ID() converts the returned generic ID into
the parent FT_Epetra_MultiVector_ID_t. Third, the Epetra_MultiVector() constructor uses the latter ID to construct an Epetra_MultiVector that gets assigned
to the parent component of the new Epetra_Vector at line 3537.
Lines 7089 of Figure 11.6 present alias_EpetraMultiVector_ID(). At
lines 8183, that procedure passes its generic_id argument to the CTrilinos
CT_Alias() function along with an enumerated value identifying the table of
the desired return ID. Lines 8688 copies that return ID into a newly allocated
generic alias_id, the C address of which is then passed to the module procedure degeneralize_EpetraMultiVector(), which converts alias_id into a
class-specic struct ID.
Being able to reference the same object by either its own struct ID or that of
the parent guarantees the ability to invoke the Epetra_MultiVector type-bound
procedure Norm2() on an Epetra_Vector or on an Epetra_MultiVector. Thus, the
ForTrilinos equivalent of the presented CTrilinos example follows:
type(Epetra_CrsMatrix) :: A
type(Epetra_JadMatrix) :: B
A=Epetra_CrsMatrix(...) ! arguments omitted
B=Epetra_JadMatrix(...) ! arguments omitted
call A%TwoRowMatrixOp(B)
Supporting this syntax requires that Epetra_RowMatrix_TwoRowMatrixOp() be
wrapped by TwoRowMatrixOp(). ForTrilinos uses the type-bound procedure
get_EpetraRowMatrix_ID() to pass the required arguments to the wrapped function. Consequently, users can pass any Epetra_RowMatrix subclass instances to an
Epetra_RowMatrix method.

11.3 Case Study: ForTrilinos and CTrilinos

In the implementation just described, we have again encountered issues that

developers writing solely in Fortran 2003 (or even in C++ in this case) need never
confront. The need to explicitly traverse branches of an inheritance hierarchy by
converting an instance of a child type to a generic type for subsequent conversion to
its parent type arises from using a non-OOP language as an intermediary between
two OOP languages. In pure Fortran or C++, one could invoke a method dened
in a parent class on an instance of a child class without ever even discovering the
instances dynamic type.

11.3.5 Discussion
Experience with implementing the strategy outlined in this article suggests that it
provides a noninvasive approach to wrapping the packages in Trilinos. Developers
of individual packages need not concern themselves with designing for interoperability. Several packages have been successfully wrapped without modication to
the underlying package.
The strategy has also proven to be less labor intensive and more exible than
prior practices of manually constructing wrappers for individual packages for specic
end-users. The labor reduction stems from the possibility of automating portions
of the process, namely the generation of procedural bindings consisting of the C
header les and the Fortran interface bodies. The increased exibility is manifold.
First, the adherence to language-standard features, designed specically to promote interoperability, all but guarantees compatible bit representations, linking,
and argument-passing conventions. Second, the ability to construct Petra objects,
which are used by all Trilinos packages, ensures that wrapping Petra packages immediately decreases the subsequent effort required to wrap new packages. Third,
the ability for Fortran programmers to directly control the construction of Epetra objects, in particular, eliminates the need for the developers of individualized
wrappers to hardwire assumptions about parameter values into the construction
process.
Drawbacks of the above approach include the amount of new code that must
be generated and the inability to pass objects directly between Fortran and C++.
The desired portability comes at the cost of manipulating only lightweight, Fortran
shadows of the underlying C++ objects. Direct manipulation of the objects happens
only in C++, and the functionality that supports that manipulation gets exported in
C header les, which necessitates sacricing OOP in the intermediate layer. Fortunately, this sacrice is purely internal and is not exposed to the end users. Even
though there has been some discussion in the Fortran community of adding interoperability with C++ to a future Fortran standard, no formal effort is underway and
the standardization of any such effort would likely be many years away.

1
2
3
4

Figure 11.10
module Epetra_Vector_module
use ForTrilinos_enums
,only: &
FT_Epetra_MultiVector_ID_t,FT_Epetra_Vector_ID_t,&
FT_Epetra_BlockMap_ID_t,ForTrilinos_Universal_ID_t,FT_boolean_t

281

282

Mixed-Language Programming
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24

use ForTrilinos_table_man
use ForTrilinos_universal
use ForTrilinos_error
use FEpetra_MultiVector ,only: Epetra_MultiVector
use FEpetra_BlockMap
,only: Epetra_BlockMap
use iso_c_binding
,only: c_int
use forepetra
implicit none
private
! Hide everything by default
public :: Epetra_Vector ! Expose type/constructors/methods
type ,extends(Epetra_MultiVector)
:: Epetra_Vector !"shell"
private
type(FT_Epetra_Vector_ID_t) :: vector_id
contains
procedure
:: cpp_delete => &
ctrilinos_delete_EpetraVector
procedure
:: get_EpetraVector_ID
procedure ,nopass :: alias_EpetraVector_ID
procedure
:: generalize
end type

25
26
27
28

interface Epetra_Vector ! constructors

module procedure duplicate,from_struct,constructor1
end interface

29
30

contains

31
32
33
34
35
36
37
38

type(Epetra_Vector) function from_struct(id)

type(FT_Epetra_Vector_ID_t) ,intent(in) :: id
from_struct%vector_id = id
from_struct%Epetra_MultiVector=Epetra_MultiVector( &
from_struct%alias_EpetraMultiVector_ID(from_struct%generalize()) &
)
end function

39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59

type(Epetra_Vector) function constructor1(BlockMap,zero_initial)

use ForTrilinos_enums ,only: &
FT_boolean_t,FT_FALSE,FT_TRUE,FT_Epetra_BlockMap_ID_t
use FEpetra_BlockMap ,only: Epetra_BlockMap
class(Epetra_BlockMap) ,intent(in) :: BlockMap
logical ,optional
,intent(in) :: zero_initial
integer(FT_boolean_t)
:: zero_out
type(FT_Epetra_Vector_ID_t)
:: constructor1_id
if (.not.present(zero_initial)) then
zero_out=FT_FALSE
elseif (zero_initial) then
zero_out=FT_TRUE
else
zero_out=FT_FALSE
endif
constructor1_id = Epetra_Vector_Create( &
BlockMap%get_EpetraBlockMap_ID(),zero_out &
)
constructor1 = from_struct(constructor1_id)
end function

11.3 Case Study: ForTrilinos and CTrilinos

60
61
62
63
64
65
66

type(Epetra_Vector) function duplicate(this)

type(Epetra_Vector) ,intent(in) :: this
type(FT_Epetra_Vector_ID_t) :: duplicate_id
duplicate_id = Epetra_Vector_Duplicate(this%vector_id)
duplicate = from_struct(duplicate_id)
end function

67
68
69
70
71

type(FT_Epetra_Vector_ID_t) function get_EpetraVector_ID(this)

class(Epetra_Vector) ,intent(in) :: this
get_EpetraVector_ID=this%vector_id
end function

72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92

type(FT_Epetra_Vector_ID_t) function alias_EpetraVector_ID(&

generic_id &
)
use iso_c_binding
,only: c_loc
use ForTrilinos_enums
,only: &
ForTrilinos_Universal_ID_t, FT_Epetra_Vector_ID
use ForTrilinos_table_man,only: CT_Alias
type(ForTrilinos_Universal_ID_t) ,intent(in) :: generic_id
type(ForTrilinos_Universal_ID_t) ,allocatable ,target :: alias_id
integer(c_int) :: status
type(error) :: ierr
allocate(alias_id,source= &
CT_Alias(generic_id,FT_Epetra_Vector_ID),stat=status &
)
ierr=error(status,FEpetra_Vector:alias_Epetra_Vector_ID)
call ierr%check_success()
alias_EpetraVector_ID=degeneralize_EpetraVector( &
c_loc(alias_id) &
)
end function

93
94
95
96
97
98
99

type(ForTrilinos_Universal_ID_t) function generalize(this)

use ForTrilinos_utils ,only: generalize_all
use iso_c_binding
,only: c_loc
class(Epetra_Vector) ,intent(in) ,target :: this
generalize = generalize_all(c_loc(this%vector_id))
end function

100
101
102
103
104
105
106
107
108
109
110

type(FT_Epetra_Vector_ID_t) function degeneralize_EpetraVector( &

generic_id) bind(C)
use ForTrilinos_enums
,only: &
ForTrilinos_Universal_ID_t,FT_Epetra_Vector_ID_t
use ,intrinsic :: iso_c_binding ,only: c_ptr,c_f_pointer
type(c_ptr)
,value
:: generic_id
type(FT_Epetra_Vector_ID_t) ,pointer :: local_ptr
call c_f_pointer (generic_id, local_ptr)
degeneralize_EpetraVector = local_ptr
end function

111
112
113
114

subroutine ctrilinos_delete_EpetraVector(this)
class(Epetra_Vector) ,intent(inout) :: this
call Epetra_Vector_Destroy(this%vector_id)

283

284

Mixed-Language Programming
115

end subroutine

116
117
118

end module

119

Figure 11.10. ForTrilinos Epetra_Vector wrapper excerpt.

EXERCISES
1. Expand the valid_kind_parameters() test of Figure 11.4 to include tests for
compiler support for the remaining interoperable types in the Fortran 2003 standard. See Metcalf et al. (2004) or the Fortran 2003 standard (a free draft version
is available on the Web).
2. Download the Trilinos solver framework from http://trilinos.sandia.gov. Write
Fortran 2003 a wrapper module for the Epetra_VbrMatrix class.

12 Multiphysics Architectures

When sorrows come, they come not single spies but in battalions.
William Shakespeare

12.1 Toward a Scalable Abstract Calculus

The canonical contexts sketched in Section 4.3 and employed throughout Part II were
intentionally low-complexity problems. Such problems provided venues for eshing
out complete software solutions from their high-level architectural design through
their implementation in source code. As demonstrated by the analyses in Chapter 3,
however, the issues addressed by OOA, OOD, and OOP grow more important as
a software packages complexity grows. Complexity growth inevitably arises when
multiple subdisciplines converge into multiphysics models. The attendant increase
in the scientic complexity inevitably taxes the hardware resources of any platform
employed. Thus, leading-edge research in multiphysics applications must ultimately
address how best to exploit the available computing platform.
Recent trends in processor architecture make it clear that fully exploiting the
available hardware on even the most modest of computing platforms necessitates
mastering parallelism. Even laptop computers now contain multicore processors,
and the highest-end machines contain hundreds of thousands of cores. The process
of getting a code to run efciently on parallel computers is referred to as getting a code
to scale, and code designs that facilitate scaling are termed scalable. The fundamental
performance question posed by this chapter is whether one can construct a scalable
ABSTRACT CALCULUS. The Sundance project (Long 2004) has already answered this
question in the afrmative for C++. The Morfeus project (Rouson et al. 2010) aims to
provide the rst demonstration of a similar capability in Fortran 2003. Whereas this
quest remains a subject of ongoing research, the current chapter presents progress
to date along with ideas for the path forward.
The holy grail of scalability is automatic parallelization of code without direct programmer intervention. Such automation might happen at the discretion of
an optimizing compiler or a linked library. In the absence of automation, parallelization typically happens explicitly at the hands of the programmer. We briey
explore parallelization by automatic means in Section 12.1.2, by compiler directives
285

286

Multiphysics Architectures

in Section 12.1.3, by wrapping a library in Section 12.1.4, and by intrinsic language

constructs in Section 12.1.5.
One aspect of the holy grail not likely to be addressed in the foreseeable future is algorithmic design. No parallelization tool of which we are aware has the
insight to alert a programmer to the availability of a more readily parallelized algorithm than the chosen one. One barrier to fully automated parallelization lies in
the difculty and domain specicity of dening useful high-level abstractions. Building application code atop an ABSTRACT CALCULUS helps in this regard. An ABSTRACT
CALCULUS facilitates the specication of a problems structure at a level high enough
to allow great variation in the underlying algorithms. The software expression of
the Burgers equation on page 203, for example, conveys the abstract notion that
differentiation must be performed without specifying how to perform it. The derivative implementation might be serial or parallel with any of several parallelization
strategies.
Consider the derivative implementations in Figures 9.2(c) and 9.3(e), which
require solving linear systems that yield 6th-order-accurate Pad estimates of the
derivatives at each grid point. Portions of the chosen LU factorization and back
substitution algorithms are inherently serial. For this reason, we switch to 2nd-orderaccurate central differences in Figure 12.1 to take advantage of greater opportunities
for parallelism. An exercise at the end of the chapter guides the reader in constructing a new main and a periodic_2nd_factory to replace the corresponding codes in
Figures 9.2(a) and Figures 9.2(e). All other modules from Figures 9.2 can be used
untouched with Figure 12.1.
The switch to central differences carries with it a drop in the asymptotic convergence rate and an associated likely drop in accuracy. The accuracy requirements of
the problem domain determine the acceptability of any accuracy losses on the given
grid. The size of the available hardware determines whether the accuracy can be
recovered through grid renement. The speed of the hardware and efciency of its
utilization determines whether such grid renement can be accomplished within acceptable solution time constraints. The problem domain thus inuences the outcome
of efforts to improve scalability.

12.1.1 Amdahls Law and Parallel Efciency

As suggested in Section 1.6, we advocate empirically driven optimization. As a useful
starting point, one can identify important bottlenecks by measuring the runtime share
occupied by each procedure in a representative run of the code. Table 12.1 proles
the runtime share for the dominant procedures in a serial run of the central difference
Burgers equation solver compiled without any optimizations. The run goes for 1,000
time steps to render initialization and cleanup costs negligible relative to the cost of
marching forward in time.

1
2
3

Figure 12.1
module periodic_2nd_order_module
use field_module ,only : field,initial_field
use kind_parameters ,only : rkind,ikind

12.1 Toward a Scalable Abstract Calculus

4
5
6

implicit none
private
public :: periodic_2nd_order, constructor

7
8
9
10
11
12
13
14
15
16
17
18
19
20
21

type ,extends(field) :: periodic_2nd_order

private
real(rkind) ,dimension(:) ,allocatable :: f
contains
procedure :: add => total
procedure :: assign => copy
procedure :: subtract => difference
procedure :: multiply_field => product
procedure :: multiply_real => multiple
procedure :: runge_kutta_2nd_step => rk2_dt
procedure :: x => df_dx
! 1st derivative w.r.t. x
procedure :: xx => d2f_dx2 ! 2nd derivative w.r.t. x
procedure :: output
end type

22
23

real(rkind) ,dimension(:) ,allocatable :: x_node

24
25
26
27

interface periodic_2nd_order
procedure constructor
end interface

28
29

real(rkind) ,parameter :: pi=acos(-1._rkind)

30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52

contains
function constructor(initial,num_grid_pts)
type(periodic_2nd_order) ,pointer :: constructor
procedure(initial_field) ,pointer :: initial
integer(ikind) ,intent(in) :: num_grid_pts
integer :: i
allocate(constructor)
allocate(constructor%f(num_grid_pts))
if (.not. allocated(x_node)) x_node = grid()
do i=1,size(x_node)
constructor%f(i)=initial(x_node(i))
end do
contains
pure function grid()
integer(ikind) :: i
real(rkind) ,dimension(:) ,allocatable :: grid
allocate(grid(num_grid_pts))
do i=1,num_grid_pts
grid(i) = 2.*pi*real(i-1,rkind)/real(num_grid_pts,rkind)
end do
end function
end function

53
54
55
56
57
58

real(rkind) function rk2_dt(this,nu, num_grid_pts)

class(periodic_2nd_order) ,intent(in) :: this
real(rkind) ,intent(in) :: nu
integer(ikind) ,intent(in) :: num_grid_pts
real(rkind)
:: dx, CFL, k_max

287

288

Multiphysics Architectures
59
60
61
62
63

dx=2.0*pi/num_grid_pts
k_max=num_grid_pts/2.0_rkind
CFL=1.0/(1.0-cos(k_max*dx))
rk2_dt = CFL*dx**2/nu
end function

64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82

function total(lhs,rhs)
class(periodic_2nd_order) ,intent(in) :: lhs
class(field) ,intent(in) :: rhs
class(field) ,allocatable :: total
type(periodic_2nd_order) ,allocatable :: local_total
integer
:: i
select type(rhs)
class is (periodic_2nd_order)
allocate(local_total)
allocate(local_total%f(size(lhs%f)))
do i=1, size(lhs%f)
local_total%f(i) = lhs%f(i) + rhs%f(i)
end do
call move_alloc(local_total,total)
class default
stop periodic_2nd_order%total: unsupported rhs class.
end select
end function

83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101

function difference(lhs,rhs)
class(periodic_2nd_order) ,intent(in) :: lhs
class(field) ,intent(in) :: rhs
class(field) ,allocatable :: difference
type(periodic_2nd_order) ,allocatable :: local_difference
integer
:: i
select type(rhs)
class is (periodic_2nd_order)
allocate(local_difference)
allocate(local_difference%f(size(lhs%f)))
do i=1, size(lhs%f)
local_difference%f(i) = lhs%f(i) - rhs%f(i)
end do
call move_alloc(local_difference,difference)
class default
stop periodic_2nd_order%difference: unsupported rhs class.
end select
end function

102
103
104
105
106
107
108
109
110
111
112
113

function product(lhs,rhs)
class(periodic_2nd_order) ,intent(in) :: lhs
class(field) ,intent(in) :: rhs
class(field) ,allocatable :: product
type(periodic_2nd_order) ,allocatable :: local_product
integer
:: i
select type(rhs)
class is (periodic_2nd_order)
allocate(local_product)
allocate(local_product%f(size(lhs%f)))
do i=1, size(lhs%f)

12.1 Toward a Scalable Abstract Calculus

114
115
116
117
118
119
120

local_product%f(i) = lhs%f(i) * rhs%f(i)

end do
call move_alloc(local_product,product)
class default
stop periodic_2nd_order%product: unsupported rhs class.
end select
end function

121
122
123
124
125
126
127
128
129
130
131
132
133
134

function multiple(lhs,rhs)
class(periodic_2nd_order) ,intent(in) :: lhs
real(rkind) ,intent(in) :: rhs
class(field) ,allocatable :: multiple
type(periodic_2nd_order) ,allocatable :: local_multiple
integer
:: i
allocate(local_multiple)
allocate(local_multiple%f(size(lhs%f)))
do i=1, size(lhs%f)
local_multiple%f(i) = lhs%f(i) * rhs
end do
call move_alloc(local_multiple,multiple)
end function

135
136
137
138
139
140
141
142
143
144
145

subroutine copy(lhs,rhs)
class(field) ,intent(in) :: rhs
class(periodic_2nd_order) ,intent(inout) :: lhs
select type(rhs)
class is (periodic_2nd_order)
lhs%f = rhs%f
class default
stop periodic_2nd_order%copy: unsupported copy class.
end select
end subroutine

146
147
148
149
150
151
152

function df_dx(this)
class(periodic_2nd_order) ,intent(in) :: this
class(field) ,allocatable :: df_dx
integer(ikind) :: i,nx, x_east, x_west
real(rkind) :: dx
class(periodic_2nd_order) ,allocatable :: local_df_dx

153
154
155
156
157
158
159
160
161
162
163
164

nx=size(x_node)
dx=2.*pi/real(nx,rkind)
allocate(local_df_dx)
allocate(local_df_dx%f(nx))
do i=1,nx
x_east = mod(i,nx)+1
x_west = nx-mod(nx+1-i,nx)
local_df_dx%f(i)=0.5*(this%f(x_east)-this%f(x_west))/dx
end do
call move_alloc(local_df_dx, df_dx)
end function

165
166
167
168

function d2f_dx2(this)
class(periodic_2nd_order) ,intent(in)
class(field) ,allocatable :: d2f_dx2

:: this

289

290

Multiphysics Architectures
169
170
171

integer(ikind) :: i,nx,x_east,x_west
real(rkind)
class(periodic_2nd_order) ,allocatable

:: dx
:: local_d2f_dx2

172
173
174
175
176
177
178
179
180
181
182
183
184

nx=size(this%f)
dx=2.*pi/real(nx,rkind)
allocate(local_d2f_dx2)
allocate(local_d2f_dx2%f(nx))
do i=1, nx
x_east = mod(i,nx)+1
x_west = nx-mod(nx+1-i,nx)
local_d2f_dx2%f(i) = &
(this%f(x_east)-2.0*this%f(i)+this%f(x_west))/dx**2
end do
call move_alloc(local_d2f_dx2, d2f_dx2)
end function

185
186
187
188
189
190
191
192

subroutine output(this)
class(periodic_2nd_order) ,intent(in) :: this
integer(ikind) :: i
do i=1,size(x_node)
print *, x_node(i), this%f(i)
end do
end subroutine

193
194

end module
Figure 12.1. Periodic, 2nd-order central differences concrete eld class.

Table 12.1. Runtime prole for central difference

Burgers solver procedures on an IBM smp system

Procedure

Runtime Share

d2f_dx2
df_dx
multiple
copy
product
total
difference

24.5%
24.0%
18.7%
15.1%
6.2%
5.8%
5.4%

Total

99.7%

The evidence in Table 12.1 suggests the greatest impact of any optimization
effort will come through reducing the time occupied by the seven listed procedures.
These procedures collectively occupy more than 99% of the runtime. First efforts
might best be aimed at the top four procedures, which collectively occupy a runtime
share exceeding 80%, or even just the top two, which occupy a combined share
near 50%.

12.1 Toward a Scalable Abstract Calculus

Amdahls Law quanties the overall speed gain for a given gain in a subset of a
process. In doing so, it also facilitates calculating the maximum gain in the limit as
the given subsets share of the overall completion time vanishes. Imagine breaking
some process into N distinct portions with the ith portion occupying Pi fraction of
the overall completion time. Then imagine ordering the portions such that the N th
portion subsumes all parts of the overall process that have xed costs and therefore
cannot be sped up. Dene the speedup of the ith portion as:
Si

toriginal
toptimized

(12.1)

where the numerator and denominator are the original and optimized completion
times, respectively, and where SN 1 by denition. Amdahls Law states that the
speedup of the overall process is:
 S)
 =
Sov erall (P,

PN +

1
N1 Pi

(12.2)

i=1 Si

 and S are vectors with i th elements Pi and Si , respectively.

where P
The previous paragraph makes no assumptions about whether the original completion time involves some optimization. Nor does it assume anything about the
optimization process. It does not even stipulate that the process in question involves
a computer! Amdahls Law applies as well to ones daily commute, for example, as
it does to scientic computing. A driver might optimize highway portions of the trip
by taking on a passenger and driving in a carpool lane, whereas nonhighway portions are likely to have xed costs. In scientic computing, the optimization might
involve algorithmic redesign, compiler optimizations, hardware improvements, or
parallelization. In the parallelization case, a speedup of 5 might involve going from
serial code to code running on ve execution units (e.g., cores) or it might involve
scaling from 10,000 processors to 50,000.
In addition to calculating the actual speedup in various scenarios, Amdahls
Law places limits on the theoretical maximum speedup in various scenarios. These
scenarios derive from taking the limit as various components of S grow without
bound. Allowing all components to grow yields the upper bound on all scenarios.
This ultimate bound is Smax 1/PN .
Smax sets an important limit on the potential utility of parallel hardware. To
expose this limitation, consider an optimization strategy limited to parallelization
alone. Distributing the execution across M hardware units, the problem-independent
maximum speedup such hardware can provide is M. With this observation, we dene
the parallel efciency:
Soverall
(12.3)
E
M
where E = 100% corresponds to the maximal use of the available hardware. When
Smax < M, it is impossible take full advantage of all available execution units. Thus,
efcient use of massively parallel systems requires very large Smax and correspondingly tiny PN , implying that every dusty corner of a code must scale. Any portion
that does not scale becomes the rate-limiting step.
After the procedures listed in Table 12.1, the next highest (unshown) as ranked
by their runtime share are compiler-generated, memory-management system calls.

291

292

Multiphysics Architectures

Each of these procedures occupies a share of 0.1% or lower, totaling 0.3%. Given
that the time occupied by system calls lies beyond an applications control and given
that reducing the number of calls would require signicant algorithmic redesign, it
seems reasonable to treat the system calls as a xed cost and take PN 0.003 and

Smax 333. Hence, it would be impossible to take full advantage of more than 333
execution units in parallel. With the recent advent of quad-socket, quad-core nodes,
the number of cores on even a 32-node cluster would exceed 333. At the time of this
writing, a relatively modest $150,000 research equipment grant is likely to garner
5121,024 cores, putting it out of reach for full utilization by a code with even a 0.3%
share of nonscalable execution.
This discussion neglects the overhead costs associated with parallel execution.
Such overhead might include, for example, the spawning and joining of threads
or communication between processes. Also, we considered only the ideal case of
Si i. With nite speedup on portions 1 through N 1, the Soverall might continue
to improve with increasing numbers of execution units beyond the aforementioned
limits. A code might then make effective use of greater than 250 execution units,
although Soverall Smax would still hold.
Finally, the parallel efciency denition in equation (12.3) applies to strong scaling, the case in which the problem size remains xed as the number of execution
units increases. The alternative, weak scaling describes the case where the problem
size grows proportionately to the number of execution units. Which type of scaling
matters more depends on the application domain.
Vision without Execution Is Hallucination.
Thomas Edison
Regarding code efciency, a viewpoint held without the backing of execution
data represents an unfounded belief. Proling procedures runtime share guides
developers toward the chief bottlenecks. Amdahls law determines the speedup
achievable by focusing on specic bottlenecks.

12.1.2 Automatic Parallelization: Compiler Technology

Chapter 1 argues that signicantly speeding up overall solution times requires optimizing execution times in ways that do not signicantly increase development times.
Guided by that principle, Sections 12.1.212.1.5 emphasize parallelization strategies
that minimally impact the programming process. The treatment progresses from
a strategy with no impact on programming but concomitantly less benet through
strategies with varying degrees of programmer burden but with greater control and
varying benet.
There is no free lunch; someone has to do the hard work to get an application
to scale, but much of that work can be shifted to library or compiler developers.
Along the way, it becomes clear that ABSTRACT CALCULUS lends itself naturally to
parallelization. With ABSTRACT CALCULUS, getting an application to scale reduces to
the problem of getting each operator in an overloaded expression to scale. Furthermore, the purely functional (intent(in)) requirements of Fortran-dened operators
help eliminate complex data dependencies that might otherwise inhibit parallelizing in those operations: Programmers and the compiler agree that no inadvertent

12.1 Toward a Scalable Abstract Calculus

updates on operands occur during operator evaluations (though operators might

overlap by reading the same memory location). These operator evaluations can take
place asynchronously, with synchronization occurring only at the end of the nal
assignment. ABSTRACT CALCULUS thus clears a straightforward path toward scalable
execution.
The simplest strategy leverages the automatic parallelization of code by an
optimizing compiler. Auto-parallelization generally comprises two categories of
techniques:
1. Multithreading: splitting a loop into multiple threads of execution or
2. Auto-vectorization: replacing scalar operations (performed on one array element at a time) with vector operations (performed on multiple elements at a
time) inside a loop.
Auto-vectorization requires supporting hardware. The growing of popularity of
single instruction multiple data (SIMD) processing brings auto-vectorization downmarket from the high-end platforms to the mainstream ones. Optimizing compilers
can now exploit vectorization capabilities on commodity hardware via libraries
such as the Mathematical Acceleration Subsystem (MASS) libraries.1 on the IBM
PowerPC family of processors.2
SIMD hardware comprises multiple processing elements capable of executing
the same operation on multiple data concurrently. Consider lines 130132 in the
user-dened multiplication operator implementation of Figure 12.1:
do i=1, size(lhs%f)
local_multiple%f(i) = lhs%f(i) * rhs
end do
A vectorizing compiler might unroll this loop by a factor of n, where n is the number of
oating point units on the target SIMD hardware. For instance a 16-byte-wide SIMD
unit can process four single-precision, oating-point values simultaneously, corresponding to n = 4. For double-precision values, n = 2. In any case, the transformed
loop would resemble:
do i=1, size(lhs%f), n
local_multiple%f(i:i+n-1) = lhs%f(i:i+n-1) * rhs
end do
!..residual loop processing..
where a SIMD unit can carry out the load, store, and multiplication on the array sections local_multiple%f(i:i+n-1)+ and lhs%f(i:i+n-1)+ as if operating
on a single pair of operand or pair of operands. The speedup attained via SIMD
instructions depends on n, which depends on the available vector length and the
oating-point unit length. On an IBM Power5+ machine, we have observed a speedup
of 2 on the multiple function with auto-vectorization. Similar speedups obtain on
all dened operators.

1 http://www-01.ibm.com/software/awdtools/mass/
2 http://tinyurl.com/2wmrjb7

293

294

Multiphysics Architectures

Vectorization requires that the array sections occupy contiguous memory. The
aforementioned example satises this condition because the f component, an
instance of periodic_2nd_order, contains an allocatable array and is thereby guaranteed to be contiguous. By contrast, pointer arrays may pose challenges in this
respect since they can be associated with noncontiguous array targets. In the case
of SIMD architectures, alignments can also pose challenges to compilers applying
vectorization techniques (Eichenberger et al. 2004).
The other popular auto-parallelization technique exploits symmetric multiprocessing (SMP) systems that is systems in which each processor has an identical view
of the memory. SMP computers (and SMP nodes on multinode, distributed clusters)
can run multiple threads to parallelize a loop. The advent of multicore chips bolsters the popularity of such multithreading. The technique requires a mechanism for
spawning and joining threads in shared memory as Section 12.1.3 explains. To maximize the speedup of a loop using all available cores, a compiler may automatically
parallelize the loop, that is, divide it into multiple segments and bind each segment
to an available core3 .
Considering the aforementioned loop from multiple, the concept of the autoparallelization can be explained using the following pseudo-code:
CHUNK_SIZE = size(lhs%f) / NUM_THREADS
! start parallel region
call create_thread(run_by_one_thread, 1, CHUNK_SIZE)
call create_thread(run_by_one_thread, CHUNK_SIZE+1, 2*CHUNK_SIZE)
...
call create_thread(run_by_one_thread, &
(NUM_THREADS-1)*CHUNK_SIZE+1, size(lhs%f))
! join threads & end of parallel region
... subroutine run_by_one_thread (lbound, ubound)
integer, intent( in ) :: lbound, ubound
do i=lbound, ubound
local_multiple%f(i) = lhs%f(i) * rhs
end do
end subroutine
Nothing fundamental differentiates between the automated parallelized loops executed at runtime by compiler-automated means versus the directive-based technique
we discuss next. For the simple loop used in multiple, the compiler can easily determine the loop is independent in execution order and therefore parallelizable. In
practice, however, many loops are not so amenable to simple analysis, and compilers
may have difculty removing data dependencies between loop iterations. In those
cases, the programmer usually has an edge in knowing the feasibility of parallelizing
a loop.

3 Although binding each thread to its own core avoids resource conicts and is allowable for a compiler

to do, the OpenMP multithreading technology discussed in Section 12.1.3 does not yet facilitate a
standard, portable mechanism for such bindings.

12.1 Toward a Scalable Abstract Calculus

We Dont Believe That... Joe the Programmer Should Have to Deal

with Parallelism in an Explicit Way.
Kunle Olukotun
Going parallel need not mean thinking parallel. Some parallelism can be achieved
by optimizing compilers without direct programmer intervention, sometimes with
results that outperform what a novice programmer can accomplish when given direct control. The next section demonstrates that greater programmer intervention
does not necessarily lead to greater performance than a sophisticated compiler
generates at least for modest numbers of cores.

12.1.3 Directive-Based Parallelization: OpenMP

The next step up in programmer involvement commonly involves multithreading via the OpenMP application programming interface (API). The OpenMP
API consists of compiler directives, library routines, and environment variables (Chapman et al. 2007). Multithreading follows the fork-join paradigm in which
a program forks into multiple threads of execution for some period of time before
joining these threads back into one. The join operation awaits completion of each
thread before continuing. In the most common approach, this happens multiple times
during a run, as depicted in Figure 12.2.
OpenMP supports a shared-memory programming model. Each thread has access to the same data unless the programmer explicitly species that each thread
have private copies of a given variable. The shared-memory viewpoint at the application programming level need not imply shared memory at the hardware level. The
operating system on the Silicon Graphics, Inc. (sgi) Altix4 supercomputer product
line, for example, provides programs a shared view of memory despite the memory
being distributed across multiple nodes. Even when the hardware and operating systems views of memory are distributed, it is possible to extend OpenMP to run on
distributed-memory platforms as with the Intel Corp.s Cluster OpenMP API5 .
Thread n

Master
Thread

Thread n

Master
Thread
Thread 3

Thread 3

Thread 2

Thread 2

Thread 1

Thread 1

n = OMP_NUM_THREADS ( )
Figure 12.2. Fork-join programming model: A typical multithreaded program forks and joins
multiple threads of execution many times throughout a given run.
4 http://www.sgi.com/products/servers/
5 http://software.intel.com/en-us/articles/cluster-openmp-for-intel-compilers/

295

296

Multiphysics Architectures

Most current compilers support OpemMP. Table 12.1 shows the runtime prole
for a serial, central-difference, Burgers-solver executable generated without automatic parallelization by the IBM XL Fortran compiler version 13.1 on an eight-core
platform running the AIX operating system. Since d2f_dx2 and df_dx combine to
occupy a 48.5% runtime share, optimizing these two procedures will have the greatest
impact on speeding up the code. The simplest approach for doing so is with OpenMP
bracing the loops in these procedures with OpenMP directives to the compiler. For
df_dx, the result is:
!$OMP parallel do private (x_east, x_west)
do i=1,nx
x_east = mod(i,nx)+1
x_west = nx-mod(nx+1-i,nx)
local_df_dx%f(i)=0.5*(this%f(x_east)-this%f(x_west))/dx
end do
!$OMP end parallel do
where the !$OMP OpenMP sentinels prompt an OpenMP-enabled compiler to interpret the trailing directives. A compiler without OpenMP capabilities ignores the
entire line, seeing the leading exclamation mark as the harbinger for a comment.
The parallel do/end parallel do pair simultaneously fork threads and delimit a
work-sharing region. At the beginning of this region, an OpenMP compiler forks
a number of threads equal to the OMP_NUM_THREADS environment variable6 . At the
end, the compiler joins all the threads into one. The !$OMP end parallel do directive implicitly synchronizes the computation, acting as a barrier that forces the
master thread to wait for all other threads to nish their work before closing the
parallel region. Finally, the private clause gives each thread its own copies of the
listed variables. These variables can be considered local to each thread.
When the number of threads exceeds the number of available execution units,
performance generally degrades. Hence, the maximum number of threads tested
on the aforementioned platform is eight. Figure 12.3 shows the performance of the
central-difference Burgers solver with loop-level OpemMP directives embedded in
df_dx and d2f_dx2 up to eight threads. The symbols labeled Amdahls law correspond to equation (12.2) with PN = 1 0.485 and Si = n, where n is the number of
threads. The speedup at n = 2 and n = 4 hugs Amdhals law closely, while the performance slips some at n = 8. Deviations from Amdahls law generally stem from costs
incurred for thread startup, synchronization, and any other overhead associated with
the OpenMP implementation.
Just as it is common to conate the shared-memory programming model with the
shared-memory hardware that often supports it, it is common to conate the forkjoin programming style with the OpenMP API that commonly supports it. In fact,
one can employ other styles with OpenMP, including the more prevalent parallel
programming style: Single Program Multiple Data (SPMD). With SPMD OpenMP,

6 The setting of environment variables is platform-dependent. In a Unix bash shell, for example, typing

export OMP_NUM_THREADS=8 at the command line sets the variable to 8.

12.1 Toward a Scalable Abstract Calculus

60
55
Amdhals law
Real execution

50

Wall time (s)

45
40
35
30
25
20
15

Number of threads

Figure 12.3. Scalability of a central difference Burgers solver using OpenMP directives.

one forks all threads at the beginning of program execution and joins them only at
the end of execution (Wallcraft 2000; Wallcraft 2002; Krawezik et al. 2002). Multiple
copies of the same program run simultaneously on separate execution units with the
work organized via explicit references to a thread identication number rather than
being shared out automatically by the compiler as with the aforementioned loop-level
directives.
Although there has been some work on automating the transformation of more
traditional OpenMP code to a SPMD style (Liu et al. 2003), manual SPMD OpenMP
programming requires considerably greater effort and care than loop-level OpenMP.
In the process, OpenMP loses some of its luster relative to the more common SPMD
API: the Message Passing Interface (MPI). The knowledge and effort required to
write SPMD OpenMP is comparable to that required for MPI. Furthermore, SPMD
OpenMP and MPI share the same motivation: Both scale better to high numbers
of execution units. Hence, when massive scalability drives the design, one more
naturally turns to MPI. The next section discusses a MPI-based solver.
12.1.4 Library-Based Parallelization: ForTrilinos and MPI
Programmers commonly describe MPI as the assembly language of parallel programming. Characteristic of this description, moving to MPI would typically be the
zenith of programmer involvement in the parallelization process. It would also be
very difcult for raw MPI code to adhere to the style suggestions set forth in Part I of
this book. Most MPI calls involve cryptic argument names and side effects, making
it challenging to write code that comments itself or code that gives visual cues to its
structure as recommended in Section 1.7.
Embedding MPI in an object-oriented library that supports domain-specic abstractions lowers the barrier to its use. The Trilinos project does this for various

297

Multiphysics Architectures

applied mathematics abstractions, including sparse linear and nonlinear solvers,

eigensolvers, and dozens of other useful services. When applications developers
extend or aggregate Trilinos classes in building solvers, Trilinos eliminates almost
all of the burden of direct interfacing with MPI.
Figure 12.5 lists a periodic_2nd_order implementation that aggregates an instance of the Epetra_Vector class presented in Chapter 11. The code in Figure 12.5
employs C preprocessor directives to differentiate between code that will be compiled on platforms with or without MPI present. This use of directives typies MPI
programming, where the availability of a library external to the language cannot be
guaranteed.
Figure 12.5 uses several ForTrilinos classes. The rst, Epetra_Map, describes the
distribution of data across the MPI processes. The second, Epetra_Vector, is described in Chapter 11. The third, Epetra_CrsMatrix, holds compressed-row-storage
matrices, providing a mechanism for performing sparse-matrix operations in distributed memory. The remaining modules provide utility classes, constants, and
procedures beyond the scope of the current discussion.
The periodic_2nd_order class in Figure 12.5(d) extends the abstract field
class and aggregates an Epetra_Vector object. Because all Epetra_Vector objects
in this case are constructed with the same Epetra_Map, Figure 12.5(d) stores the
map in a single object with module scope. See Chapter 8 for additional discussion
of the trade-offs associated with using module variables. The various operators in
Figure 12.5 delegate their operations to Epetra_Vector by invoking the appropriate
methods on their Epetra_Vector component.
Figure 12.5(e) lists a ForTrilinos-based Burgers-equation solver main program
employing ABSTRACT CALCULUS. That main program illustrates a hallmark of Trilinosbased solvers: The only MPI calls are for process startup (Mpi_Init), shutdown
(Mpi_Finalize), and timing (MPI_Wtime). Figure 12.4 shows the parallel efciency
of the ForTrilinos-based solver.
1X
X

0.8
Parallel efficiency

298

0.6

0.4

0.2
X

Strong scaling
Weak scaling

6
4
8
Number of processors

10

Figure 12.4. Parallel efciency of a ForTrilinos/MPI-based central difference Burgers solver:

strong scaling with 643 grid points and weak scaling with 643 M points for M MPI processes.
All simulations were run for 1,000 time steps.

12.1 Toward a Scalable Abstract Calculus

1
2
3
4
5
6

Figure 12.5 (a)

module field_module
use iso_c_binding, only: c_int, c_double
use FEpetra_Comm, only:Epetra_Comm
implicit none
private
public :: field, initial_field

7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24

type ,abstract :: field

contains
procedure(field_op_field) ,deferred :: add
procedure(field_eq_field) ,deferred :: assign
procedure(field_op_field) ,deferred :: subtract
procedure(field_op_field) ,deferred :: multiply_field
procedure(field_op_real) ,deferred :: multiply_real
procedure(real_to_real) ,deferred :: runge_kutta_2nd_step
procedure(derivative) ,deferred :: x ! 1st derivative
procedure(derivative) ,deferred :: xx ! 2nd derivative
procedure(output_interface) ,deferred :: output
generic :: operator(+)
=> add
generic :: operator(-)
=> subtract
generic :: operator(*)
=> multiply_real,multiply_field
generic :: assignment(=) => assign
procedure(force_finalize_interface) ,deferred :: force_finalize
end type

25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54

abstract interface
real(c_double) pure function initial_field(x)
import :: c_double
real(c_double) ,intent(in) :: x
end function
function field_op_field(lhs,rhs)
import :: field
class(field) ,intent(in) :: lhs,rhs
class(field) ,allocatable :: field_op_field
end function
function field_op_real(lhs,rhs)
import :: field,c_double
class(field) ,intent(in) :: lhs
real(c_double) ,intent(in) :: rhs
class(field) ,allocatable :: field_op_real
end function
real(c_double) function real_to_real(this,nu,grid_resolution)
import :: field,c_double,c_int
class(field) ,intent(in) :: this
real(c_double) ,intent(in) :: nu
integer(c_int),intent(in) :: grid_resolution
end function
function derivative(this)
import :: field
class(field) ,intent(in) :: this
class(field) ,allocatable :: derivative
end function
subroutine field_eq_field(lhs,rhs)
import :: field

299

300

Multiphysics Architectures
55
56
57
58
59
60
61
62
63
64
65
66
67
68

1
2
3
4
5
6
7

class(field) ,intent(inout) :: lhs

class(field) ,intent(in) :: rhs
end subroutine
subroutine output_interface(this,comm)
import :: field,Epetra_Comm
class(field) ,intent(in) :: this
class(Epetra_Comm), intent(in) :: comm
end subroutine
subroutine force_finalize_interface(this)
import :: field
class(field), intent(inout) :: this
end subroutine
end interface
end module
Figure 12.5 (b)
module field_factory_module
use FEpetra_Comm, only:Epetra_Comm
use iso_c_binding ,only :c_int
use field_module ,only : field,initial_field
implicit none
private
public :: field_factory

8
9
10
11
12

type, abstract :: field_factory

contains
procedure(create_interface), deferred :: create
end type

13
14
15
16
17
18
19
20
21
22
23
24

1
2
3
4
5
6
7
8

abstract interface
function create_interface(this,initial,grid_resolution,comm)
import :: field, field_factory ,initial_field,c_int,Epetra_Comm
class(Epetra_Comm),intent(in) :: comm
class(field_factory), intent(in) :: this
class(field) ,pointer :: create_interface
procedure(initial_field) ,pointer :: initial
integer(c_int) ,intent(in) :: grid_resolution
end function
end interface
end module
Figure 12.5 (c)
module periodic_2nd_order_factory_module
use FEpetra_Comm ,only:Epetra_Comm
use field_factory_module ,only : field_factory
use field_module ,only : field,initial_field
use periodic_2nd_order_module ,only : periodic_2nd_order
implicit none
private
public :: periodic_2nd_order_factory

9
10
11
12

type, extends(field_factory) :: periodic_2nd_order_factory

contains
procedure :: create=>new_periodic_2nd_order

12.1 Toward a Scalable Abstract Calculus

13

end type

14
15
16
17
18
19
20
21
22
23
24
25
26

1
2
3
4
5
6
7
8
9
10
11
12
13
14

contains
function new_periodic_2nd_order(this,initial,grid_resolution,comm)
use iso_c_binding, only:c_int
class(periodic_2nd_order_factory), intent(in) :: this
class(field) ,pointer :: new_periodic_2nd_order
procedure(initial_field) ,pointer :: initial
integer(c_int) ,intent(in) :: grid_resolution
class(Epetra_Comm), intent(in) :: comm
new_periodic_2nd_order=> &
periodic_2nd_order(initial,grid_resolution,comm)
end function
end module

Figure 12.5 (d)

module periodic_2nd_order_module
use FEpetra_Comm, only:Epetra_Comm
use FEpetra_Map!, only: Epetra_Map
use FEpetra_Vector!, only:Epetra_Vector
use FEpetra_CrsMatrix, only:Epetra_CrsMatrix
use ForTrilinos_enum_wrappers
use ForTrilinos_error
use field_module ,only : field,initial_field
use iso_c_binding ,only : c_double,c_int
use ForTrilinos_assertion_utility ,only: &
error_message,assert,assert_identical
implicit none
private
public :: periodic_2nd_order

15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30

type ,extends(field) :: periodic_2nd_order

private
type(Epetra_Vector) :: f
contains
procedure :: add => total
procedure :: subtract => difference
procedure :: assign => copy
procedure :: multiply_field => product
procedure :: multiply_real => multiple
procedure :: runge_kutta_2nd_step => rk2_dt
procedure :: x => df_dx
! 1st derivative w.r.t. x
procedure :: xx => d2f_dx2
! 2nd derivative w.r.t. x
procedure :: output
procedure :: force_finalize
end type

31
32
33
34

real(c_double) ,parameter
:: pi=acos(-1._c_double)
real(c_double) ,dimension(:) ,allocatable :: x_node
type(Epetra_Map)
,allocatable :: map

35
36
37
38

interface periodic_2nd_order
procedure constructor
end interface

301

302

Multiphysics Architectures
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75

contains
function constructor(initial,grid_resolution,comm) result(this)
type(periodic_2nd_order) ,pointer :: this
procedure(initial_field) ,pointer :: initial
integer(c_int) ,intent(in) :: grid_resolution
integer(c_int) :: i,j
class(Epetra_Comm), intent(in) :: comm
integer(c_int) :: NumGlobalElements
integer(c_int),dimension(:),allocatable :: MyGlobalElements
integer(c_int)
:: NumMyElements,IndexBases=1,status
real(c_double) ,dimension(:) ,allocatable :: f_v
type(error) :: ierr
NumGlobalElements=grid_resolution
allocate(this)
if (.not. allocated(x_node)) x_node = grid()
if (.not. allocated(map)) then
allocate(map,stat=status)
ierr=error(status,periodic_2nd_order: create map)
call ierr%check_allocation()
map = Epetra_Map(NumGlobalElements,IndexBases,comm)
end if
NumMyElements= map%NumMyElements()
allocate(MyGlobalElements(NumMyElements))
MyGlobalElements = map%MyGlobalElements()
allocate(f_v(NumMyElements))
forall(i=1:NumMyElements)f_v(i)=initial(x_node(MyGlobalElements(i)))
this%f=Epetra_Vector(map,zero_initial=.true.)
call this%f%ReplaceGlobalValues(NumMyElements,f_v,MyGlobalElements)
contains
pure function grid()
integer(c_int) :: i
real(c_double) ,dimension(:) ,allocatable :: grid
allocate(grid(grid_resolution))
forall(i=1:grid_resolution) &
grid(i)= 2.*pi*real(i-1,c_double)/real(grid_resolution,c_double)
end function
end function

76
77
78
79
80
81
82
83
84
85
86

subroutine copy(lhs,rhs)
class(field) ,intent(in) :: rhs
class(periodic_2nd_order) ,intent(inout) :: lhs
select type(rhs)
class is (periodic_2nd_order)
lhs%f = rhs%f
class default
stop periodic_2nd_order%copy: unsupported copy class.
end select
end subroutine

87
88
89
90
91
92
93

real(c_double) function rk2_dt(this,nu, grid_resolution)

class(periodic_2nd_order) ,intent(in) :: this
real(c_double) ,intent(in) :: nu
integer(c_int) ,intent(in) :: grid_resolution
real(c_double)
:: dx, CFL, k_max
dx=2.0*pi/grid_resolution

12.1 Toward a Scalable Abstract Calculus

94
95
96
97

k_max=grid_resolution/2.0_c_double
CFL=1.0/(1.0-cos(k_max*dx))
rk2_dt = CFL*dx**2/nu
end function

98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114

function total(lhs,rhs)
class(periodic_2nd_order) ,intent(in) :: lhs
class(field) ,intent(in) :: rhs
class(field) ,allocatable :: total
type(periodic_2nd_order) ,allocatable :: local_total
select type(rhs)
class is (periodic_2nd_order)
allocate(periodic_2nd_order::local_total)
local_total%f=Epetra_Vector(map,zero_initial=.true.)
call local_total%f%Update( &
1._c_double,lhs%f,1._c_double,rhs%f,0._c_double)
call move_alloc(local_total,total)
class default
stop periodic_2nd_order%total: unsupported rhs class.
end select
end function

115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131

function difference(lhs,rhs)
class(periodic_2nd_order) ,intent(in) :: lhs
class(field) ,intent(in) :: rhs
class(field) ,allocatable :: difference
type(periodic_2nd_order) ,allocatable :: local_difference
select type(rhs)
class is (periodic_2nd_order)
allocate(periodic_2nd_order::local_difference)
local_difference%f=Epetra_Vector(map,zero_initial=.true.)
call local_difference%f%Update(&
1._c_double,lhs%f,-1._c_double,rhs%f,0._c_double)
call move_alloc(local_difference,difference)
class default
stop periodic_2nd_order%difference: unsupported rhs class.
end select
end function

132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148

function product(lhs,rhs)
class(periodic_2nd_order) ,intent(in) :: lhs
class(field) ,intent(in) :: rhs
class(field) ,allocatable :: product
type(periodic_2nd_order) ,allocatable :: local_product
select type(rhs)
class is (periodic_2nd_order)
allocate(periodic_2nd_order::local_product)
local_product%f=Epetra_Vector(map,zero_initial=.true.)
call local_product%f%Multiply(1._c_double,lhs%f,rhs%f,0._c_double)
call move_alloc(local_product,product)
class default
stop periodic_2nd_order%product: unsupported rhs class.
end select
end function

303

304

Multiphysics Architectures
149
150
151
152
153
154
155
156
157
158

function multiple(lhs,rhs)
class(periodic_2nd_order) ,intent(in) :: lhs
real(c_double) ,intent(in) :: rhs
class(field) ,allocatable :: multiple
type(periodic_2nd_order) ,allocatable :: local_multiple
allocate(periodic_2nd_order::local_multiple)
local_multiple%f=Epetra_Vector(map,zero_initial=.true.)
call local_multiple%f%Scale(rhs,lhs%f)
call move_alloc(local_multiple,multiple)
end function

159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177

function df_dx(this)
class(periodic_2nd_order) ,intent(in) :: this
class(field) ,allocatable :: df_dx
type(Epetra_Vector) :: x
type(Epetra_CrsMatrix) :: A
type(error) :: err
real(c_double) ,dimension(:)
,allocatable :: c
real(c_double) :: dx
integer(c_int) :: nx
type(periodic_2nd_order), allocatable :: df_dx_local
integer(c_int),dimension(:),allocatable :: MyGlobalElements
integer(c_int),dimension(:),allocatable :: MyGlobalElements_diagonal
integer(c_int),dimension(:),allocatable :: NumNz
integer(c_int) :: NumGlobalElements,NumMyElements,i
integer(c_int) :: indices(2), NumEntries
real(c_double) ::values(2)
real(c_double),parameter :: zero =0.0
integer(c_int),parameter :: diagonal=1

178
179
180
181
182

! Executable code
nx=size(x_node)
dx=2.*pi/real(nx,c_double)
NumGlobalElements = nx

183
184
185
186
187
188

! Get update list and number of local equations from given Map
NumMyElements = map%NumMyElements()
call assert_identical( [NumGlobalElements,map%NumGlobalElements()] )
allocate(MyGlobalElements(NumMyElements))
MyGlobalElements = map%MyGlobalElements()

189
190
191
192
193

! Create an integer vector NumNz that is used to build the Epetra Matrix
! NumNz(i) is the number of non-zero elements for the ith global eqn.
! on this processor
allocate(NumNz(NumMyElements))

194
195
196
197

! We are building a tridiagonal matrix where each row has (-1 0 1)

! So we need 2 off-diagonal terms (except for the first and last eqn.)
NumNz = 3

198
199
200

! Create a Epetra_Matrix
A = Epetra_CrsMatrix(FT_Epetra_DataAccess_E_Copy,map,NumNz)

201
202
203

! Add rows one at a time

! Need some vectors to help

12.1 Toward a Scalable Abstract Calculus

204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231

! off diagonal values will always be -1 and 1

values(1) = -1.0/(2.0*dx)
values(2) = 1.0/(2.0*dx)
do i=1,NumMyElements
if (MyGlobalElements(i)==1) then
indices(1) = NumGlobalElements
indices(2) = 2
NumEntries = 2
else if(MyGlobalElements(i)==NumGlobalElements) then
indices(1) = NumGlobalElements-1
indices(2) = 1
NumEntries = 2
else
indices(1) = MyGlobalElements(i)-1
indices(2) = MyGlobalElements(i)+1
NumEntries = 2
end if
call A%InsertGlobalValues(&
MyGlobalElements(i),NumEntries,values,indices,err)
call assert( [err%error_code()==0_c_int] , &
[error_message(A%InsertGlobalValues: failed)] )
!Put in the diaogonal entry
MyGlobalElements_diagonal=MyGlobalElements+i-1
call A%InsertGlobalValues(MyGlobalElements(i), &
diagonal,[zero],MyGlobalElements_diagonal,err)
call assert( [err%error_code()==0_c_int] , &
[error_message(A%InsertGlobalValues: failed)] )
end do

232
233
234
235
236
237
238
239
240
241
242
243
244
245
246

!Finish up
call A%FillComplete()
!create vector x
x=Epetra_Vector(A%RowMap())
Call A%Multiply_Vector(.false.,this%f,x)
allocate(c(NumMyElements))
c=x%ExtractCopy()
!create vector of df_dx
allocate(periodic_2nd_order::df_dx_local)
df_dx_local%f=Epetra_Vector(map,zero_initial=.true.)
call df_dx_local%f%ReplaceGlobalValues(&
NumMyElements,c,MyGlobalElements)
call move_alloc(df_dx_local, df_dx)
end function

247
248
249
250
251
252
253
254
255
256
257
258

function d2f_dx2(this)
class(periodic_2nd_order) ,intent(in) :: this
class(field) ,allocatable :: d2f_dx2
type(Epetra_Vector) :: x
type(Epetra_CrsMatrix) :: A
type(error) :: err
real(c_double) ,dimension(:)
,allocatable :: c
real(c_double) :: dx
integer(c_int) :: nx
type(periodic_2nd_order) ,allocatable :: d2f_dx2_local
integer(c_int),dimension(:),allocatable :: MyGlobalElements,NumNz

305

306

Multiphysics Architectures
259
260
261
262
263

integer(c_int),dimension(:),allocatable :: MyGlobalElements_diagonal
integer(c_int) :: NumGlobalElements,NumMyElements,i
integer(c_int) :: indices(2),NumEntries
real(c_double) :: values(2),two_dx2
integer(c_int),parameter :: diagonal=1

264
265
266
267
268

! Executable code
nx=size(x_node)
dx=2.*pi/real(nx,c_double)
NumGlobalElements = nx

269
270
271
272
273
274

! Get update list and number of local equations from given Map
NumMyElements = map%NumMyElements()
call assert_identical( [NumGlobalElements,map%NumGlobalElements()] )
allocate(MyGlobalElements(NumMyElements))
MyGlobalElements = map%MyGlobalElements()

275
276
277
278
279

! Create an integer vector NumNz that is used to build the Epetra Matrix
! NumNz(i) is the number of non-zero elements for the ith global eqn.
! on this processor
allocate(NumNz(NumMyElements))

280
281
282
283

! We are building a tridiagonal matrix where each row has (1 -2 1)

! So we need 2 off-diagonal terms (except for the first and last eqn.)
NumNz = 3

284
285
286

! Create a Epetra_Matrix
A = Epetra_CrsMatrix(FT_Epetra_DataAccess_E_Copy,map,NumNz)

287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313

! Add rows one at a time

! Need some vectors to help
! off diagonal values will always be 1 and 1
values(1) = 1.0/(dx*dx)
values(2) = 1.0/(dx*dx)
two_dx2
=-2.0/(dx*dx)
do i=1,NumMyElements
if (MyGlobalElements(i)==1) then
indices(1) = NumGlobalElements
indices(2) = 2
NumEntries = 2
else if(MyGlobalElements(i)==NumGlobalElements) then
indices(1) = NumGlobalElements-1
indices(2) = 1
NumEntries = 2
else
indices(1) = MyGlobalElements(i)-1
indices(2) = MyGlobalElements(i)+1
NumEntries = 2
end if
call A%InsertGlobalValues( &
MyGlobalElements(i),NumEntries,values,indices,err)
call assert( [err%error_code()==0_c_int] , &
[error_message(A%InsertGlobalValues: failed)] )
!Put in the diaogonal entry
MyGlobalElements_diagonal=MyGlobalElements+i-1

12.1 Toward a Scalable Abstract Calculus

314
315
316
317
318

call A%InsertGlobalValues( MyGlobalElements(i) &

,diagonal,[two_dx2],MyGlobalElements_diagonal,err)
call assert( [err%error_code()==0_c_int] , &
[error_message(A%InsertGlobalValues: failed)] )
end do

319
320
321
322
323
324
325
326
327
328
329
330
331
332
333

!Finish up
call A%FillComplete()
!create vector x
x=Epetra_Vector(A%RowMap())
Call A%Multiply_Vector(.false.,this%f,x)
allocate(c(NumMyElements))
c=x%ExtractCopy()
!create vector of df_dx
allocate(periodic_2nd_order::d2f_dx2_local)
d2f_dx2_local%f=Epetra_Vector(map,zero_initial=.true.)
call d2f_dx2_local%f%ReplaceGlobalValues( &
NumMyElements,c,MyGlobalElements)
call move_alloc(d2f_dx2_local, d2f_dx2)
end function

334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353

subroutine output(this,comm)
class(periodic_2nd_order) ,intent(in) :: this
class(Epetra_Comm),intent(in) ::comm
integer(c_int) :: i,NumMyElements,NumGlobalElements
integer(c_int), dimension(:), allocatable :: MyGlobalElements
real(c_double), dimension(:), allocatable :: f_v
real(c_double), dimension(:), allocatable :: f
NumGlobalElements=map%NumGlobalElements()
NumMyElements=map%NumMyElements()
allocate(MyGlobalElements(NumMyElements))
MyGlobalElements=map%MyGlobalElements()
allocate(f_v(NumMyElements))
f_v=this%f%ExtractCopy()
allocate(f(NumGlobalElements))
call comm%GatherAll(f_v,f,NumMyElements)
do i=1,NumGlobalElements
if (comm%MyPID()==0) write(20,(2(E20.12,1x))) x_node(i),f(i)
enddo
end subroutine

354
355
356
357
358
359

1
2
3
4
5
6

subroutine force_finalize(this)
class(periodic_2nd_order), intent(inout) :: this
call this%f%force_finalize
end subroutine
end module

Figure 12.5 (e)

program main
#include "ForTrilinos_config.h"
#ifdef HAVE_MPI
use mpi
use FEpetra_MpiComm, only:Epetra_MpiComm
#else

307

308

Multiphysics Architectures
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30

use FEpetra_SerialComm, only:Epetra_SerialComm

#endif
use ForTrilinos_utils, only : valid_kind_parameters
use iso_c_binding, only : c_int,c_double
use field_module ,only : field,initial_field
use field_factory_module ,only : field_factory
use periodic_2nd_order_factory_module,only: periodic_2nd_order_factory
use initializer ,only : u_initial,zero
implicit none
#ifdef HAVE_MPI
type(Epetra_MpiComm) :: comm
#else
type(Epetra_SerialComm) :: comm
#endif
class(field), pointer :: u,half_uu,u_half
class(field_factory), allocatable :: field_creator
procedure(initial_field) ,pointer :: initial
real(c_double) :: dt,half=0.5,t=0.,t_final=0.6,nu=1.
real(c_double) :: t_start,t_end
integer(c_int) :: tstep
integer(c_int), parameter :: grid_resolution=(64**3)/4
integer(c_int)
:: MyPID, NumProc
logical
:: verbose
integer :: rc,ierr

31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61

if (.not. valid_kind_parameters()) &

stop C interoperability not supported on this platform.
#ifdef HAVE_MPI
call MPI_INIT(ierr)
t_start=MPI_Wtime()
comm = Epetra_MpiComm(MPI_COMM_WORLD)
#else
call cpu_time(t_start)
comm = Epetra_SerialComm()
#endif
allocate(periodic_2nd_order_factory :: field_creator)
initial => u_initial
u => field_creator%create(initial,grid_resolution,comm)
initial => zero
half_uu => field_creator%create(initial,grid_resolution,comm)
u_half => field_creator%create(initial,grid_resolution,comm)
do tstep=1,20 !2nd-order Runge-Kutta:
dt = u%runge_kutta_2nd_step(nu ,grid_resolution)
half_uu = u*u*half
u_half = u + (u%xx()*nu - half_uu%x())*dt*half ! first substep
half_uu = u_half*u_half*half
u = u + (u_half%xx()*nu - half_uu%x())*dt ! second substep
t = t + dt
if (comm%MyPID()==0) print *,timestep=,tstep
end do
if (comm%MyPID()==0) write(10,*) u at t=,t
#ifdef HAVE_MPI
t_end= MPI_Wtime()
#else
call cpu_time(t_end)

12.1 Toward a Scalable Abstract Calculus

62
63
64
65
66
67
68
69
70
71
72

#endif
if (comm%MyPID()==0) write(10,*) Elapsed CPU time=,t_end-t_start
call u%output(comm)
call half_uu%force_finalize
call u_half%force_finalize
call u%force_finalize
call comm%force_finalize
#ifdef HAVE_MPI
call MPI_FINALIZE(rc)
#endif
end program
Figure 12.5. ForTrilinos-based Burgers solver: (a) eld, (b) eld_factory, (c) periodic_2nd_factory, (d) periodic_2nd_order, and (3) main program.

12.1.5 Intrinsic Parallelization: Fortran 2008 Coarrays

Our journey toward scalable execution concludes with Fortran 2008 coarrays.7 As
compared to the ways in which we used automatic, multithreaded, and distributed
parallel programming models in the preceding sections, coarrays beautifully blend
simplicity, power, and portability. Originally designed as a language extension to
Fortran 95 by Numrich and Reid (1998), and named by analogy with covariant indices in tensor calculus (Numrich 2005), coarrays aim to help Fortran programmers
convert their serial applications to efcient parallel applications with ease.
Reid (2005) summarized the coarray language design as follows:
The coarray programming model is designed to answer the question What is the smallest
change required to convert Fortran into a robust and efcient parallel language?. Our
answer is a simple syntactic extension. It looks and feels like Fortran and requires Fortran
programmers to learn only a few new rules. These rules are related to two fundamental
issues that any parallel programming model must resolve, work distribution and data
distribution.

Based on the SPMD programming model, a coarray application replicates itself

across multiple execution units. The number of replications is set either at compile
time or at program start up time. Each replication is called an image and has its own
local data. All images execute asynchronously until they encounter a user-dened
synchronization point.
With the exception of coarrays, all data objects declared within the program are
local to each image and can be referenced only within that image. References to
non-local data are represented syntactically by surrounding co-indices with square
brackets as opposed to the usual parentheses that bracket array indices. This minimalistic manner of stipulating communication between images produces signicantly
simpler code than MPI and yet has the full potential to rival MPIs performance on
large, parallel applications. One expects nearly identical MPI and coarray performance with compilers that support coarrays by using MPI under the hood as the
open-source Gnu Fortran (gfortran) developers plan to do.8 Coarray performance
7 The Fortran 2008 standard was published in October 2010. The nal committee draft can be found

at ftp://ftp.nag.co.uk/sc22wg5/N1801-N1850/N1830.pdf.
8 Cite Google archive of comp.lang.fortran

309

310

Multiphysics Architectures

can even exceed that of MPI on platforms where a compiler can take advantage of
special hardware support for the coarray programming model. Cray compilers and
hardware exhibit this behavior (Numrich 2005).
Coarrays thus provide a powerful duo of simplicity and platform-agnosticism.
Whereas syntax greatly enhances the scalability of development-time processes, operating at an abstraction level slightly above other parallel programming models
enhances the scalability of runtime processes. The language standard body moved to
this higher abstraction level by using the image terminology without tying the term
to a specic technology. Hence, one compiler development team might implement
its coarray support in MPI, whereas another development team might implement it
in a hardware-specic instruction set. A third could use OpenMP. And with current
trends toward increasing single-socket parallelism with multicore chips, the greatest
benet might ultimately derive from mapping coarray syntax to a mixed model that
distributes work locally via multithreading (e.g., with OpenMP) and communicating
nonlocally via message passing (e.g., with MPI).
Consider again the 2nd-order, central-difference Burgers equation solver. The
following factors arise in attempting to get a supporting ABSTRACT CALCULUS to scale
on parallel systems using coarrays:
Data distribution.
Work load distribution, or operational model.
Necessary synchronization.

Next we address these factors in the order listed.

One naturally envisions evenly distributing the grid values and the operations
on those values across all images. We therefore embed a coarray component in the
periodic_2nd_order class. Figure 12.6 illustrates the design of that class. Figure 12.7
shows the implementation.9 As with our other Burgers equation solvers, field
remains abstract. A concrete periodic_2nd_order extends field and stores a reference to the grid values in its global_f coarray component. As shown at line 11 in
Figure 12.7, the following syntax declares an allocatable coarray:
real(rkind) ,allocatable :: global_f(:)[:]
The same component could be represented with the alternate syntax:
real(rkind),dimension(:),codimension[:],allocatable :: global_f
Allocations of global_f take the form shown at line 55 in Figure 12.7 and repeated
here:
allocate (this%global_f(local_grid_size)[*])

9 At the time of this writing, no fully compliant Fortran 2008 compilers have been released. Most

compilers with partial compliance have not reached a sufcient maturity in terms of features and
robustness to fully test the coarray code in this section. The GNU Fortran compiler is transitioning
towards supporting the coarray syntax for single-image runs. All code in this section has been
checked for syntactical correctness by compiling it with the development version of GNU Fortran
4.6.0 20100508 (experimental) installed via macports on a Mac OS X 10.6 platform. Where the
compiler appears to be in error, we have consulted the draft Fortran 2008 standard for guidance.

12.1 Toward a Scalable Abstract Calculus

Periodic_2nd_Order
- global_f : Coarray<real>
<<constructor>>
+ construct()
<<operator>>
+ add() : Field
+ subtract() : Field
+ multiply() : Field
<<assignment>>
+ assign()
<<compute>
+ x() : Field
+ xx() : Field
<<query>>
+ output()

{sync all}

{sync all}
Field
- f : Array<real>
<<operator>>
+ add() : Field
+ subtract() : Field
+ multiply_field() : Field
+ multiply_real() : Field
<<assignment>>
+ assign()
+ copy()
+ state() : Array<real>

Figure 12.6. Class diagram of periodic_2nd_order class using coarray grid eld. Methods
constuct() and assign() require synchronization as noted by constraints sync.

where the asterisk represents the number of images that will be set at compile-time
or runtime. The language standard precludes any more explicit specication of the
number of images to be allocated.
When allocated, global_f behaves as though it consists of many rank-one arrays,
each afliated with an image. An image can access its own global_f array using
regular array syntax. It can also access the global_f array on another image using
an image index syntax that encloses the index in square brackets. Computing the
spatial derivatives x() and xx(), for example, requires cross-image data access at
lines 141, 144, and 148 in Figure 12.7.
Considering the second of the aforementioned factors, the work load distribution highlights a critically signicant distinction between the field interface and
its implementation in periodic_2nd_order in the context of a coarray ABSTRACT
CALCULUS: Each periodic_2nd_order type-bound function, such as the operators +
and * and the differentiation methods x() and xx(), returns a field instance rather
than a periodic_2nd_order one. (Figures 12.8 denes the field class.) This circumvents Fortrans prohibition against functions returning objects that contain coarray
components.10
10 Fortrans prohibition against allocatable coarray components inside function addresses perfor-

mance concerns. The standard mandates automatic deallocation of allocatable components of a

function result after the results use. Deallocating allocatable coarrays requires synchronizations
across all images, which leads to segment ordering difculties. In our 2nd-order eld case, it would
also be superuous to reallocate and copy the global mesh since it doesnt change at all.

311

312

Multiphysics Architectures

More importantly, studying the time integration expressions reveals that none
of the supporting operations involve updates on global_f until the nal assignment. Thus, none require synchronization and their results can be stored locally
in a field. This principle can be captured using the following operational model:
While periodic_2nd_order holds a reference to the global eld, all its operators are
evaluated concurrently on all images, storing temporary results locally. Nonlocal updates need happen only after all of an expressions local operations have completed,
making the results ready for assignment back to the global_f.
The third factor, synchronization, refers to barrier points in the executing code.
By denition, all images have to reach a given barrier before executing any subsequent code. In general, a synchronization is needed after a coarray is updated
on one image and before it is accessed from another image. Based on our operational model, periodic_2nd_order assignments require synchronization. The only
other place that needs synchronization is construct(), where the allocation and
initialization of the global_f occurs. The following call is the most common form of
synchronization:
sync all
Another form of synchronization occurs implicitly upon allocation of an allocatable
coarray.
Lines 4575 in Figure 12.7 show the user-dened periodic_2nd_order constructor. This procedure could be written as a subroutine due to the aforementioned
restriction on functions returning objects containing a coarray component. The
num_images() intrinsic function employed in the constructor queries the total number of images in the run. The constructor uses that value to compute the local grid
size assuming the eld is evenly distributed on all images. Function this_image()
queries the image index of the executing image and uses that value to calculate the
initial eld values.
Figure 12.7
1
2
3
4

module periodic_2nd_order_module
use kind_parameters ,only : rkind, ikind
use field_module ,only : field

5
6
7
8
9
10
11
12
13
14
15
16
17
18
19

implicit none
private
public :: periodic_2nd_order, initial_field
type periodic_2nd_order
private
real(rkind), allocatable :: global_f(:)[:]
contains
procedure :: construct
procedure :: assign
=> copy
procedure :: add
=> add_field
procedure :: multiply
=> multiply_field
procedure :: x
=> df_dx
procedure :: xx
=> d2f_dx2
procedure :: output

12.1 Toward a Scalable Abstract Calculus

20
21
22
23

generic
generic
generic
end type

:: assignment(=) => assign

:: operator(+)
=> add
:: operator(*)
=> multiply

24
25

real(rkind) ,parameter :: pi=acos(-1._rkind)

26
27
28
29
30
31
32

abstract interface
real(rkind) pure function initial_field(x)
import :: rkind
real(rkind) ,intent(in) :: x
end function
end interface

33
34

contains

35
36
37
38
39
40
41
42
43

subroutine output (this)

class(periodic_2nd_order), intent(in) :: this
integer(ikind) :: i
do i = 1, size(this%global_f)
print *, (this_image()-1)*size(this%global_f) + i, &
this%global_f(i)
end do
end subroutine

44
45
46
47
48
49

subroutine construct (this, initial, num_grid_pts)

class(periodic_2nd_order), intent(inout) :: this
procedure(initial_field) ,pointer :: initial
integer(ikind) ,intent(in) :: num_grid_pts
integer :: i, local_grid_size

50
51
52

!<-- assume mod(num_grif_pts, num_images()) == 0

local_grid_size = num_grid_pts / num_images()

53
54
55

! set up the global grid points

allocate (this%global_f(local_grid_size)[*])

56
57

this%global_f(:) = grid()

58
59
60
61

do i = 1, local_grid_size
this%global_f(i) = initial(this%global_f(i))
end do

62
63

sync all

64
65
66
67
68
69
70
71
72
73
74

contains
pure function grid()
integer(ikind) :: i
real(rkind) ,dimension(:) ,allocatable :: grid
allocate(grid(local_grid_size))
do i=1,local_grid_size
grid(i) = 2.*pi*(local_grid_size*(this_image()-1)+i-1) &
/real(num_grid_pts,rkind)
end do
end function

313

314

Multiphysics Architectures
75

end subroutine

76
77
78
79
80
81
82
83
84
85
86

real(rkind) function rk2_dt(this,nu, num_grid_pts)

class(periodic_2nd_order) ,intent(in) :: this
real(rkind) ,intent(in) :: nu
integer(ikind) ,intent(in) :: num_grid_pts
real(rkind)
:: dx, CFL, k_max
dx=2.0*pi/num_grid_pts
k_max=num_grid_pts/2.0_rkind
CFL=1.0/(1.0-cos(k_max*dx))
rk2_dt = CFL*dx**2/nu
end function

87
88
89
90
91

! this is the assignment

subroutine copy(lhs,rhs)
class(periodic_2nd_order) ,intent(inout) :: lhs
class(field) ,intent(in) :: rhs

92
93
94
95
96

! update global field

lhs%global_f(:) = rhs%state()
sync all
end subroutine

97
98
99
100
101

function add_field (this, rhs)

class(periodic_2nd_order), intent(in) :: this
class(field), intent(in) :: rhs
class(field), allocatable :: add_field

102
103
104
105

allocate (add_field)
add_field = rhs%state()+this%global_f(:)
end function

106
107
108
109

function multiply_field (this, rhs)

class(periodic_2nd_order), intent(in) :: this, rhs
class(field), allocatable :: multiply_field

110
111
112
113

allocate (multiply_field)
multiply_field = this%global_f(:)*rhs%global_f(:)
end function

114
115
116
117
118
119
120

function df_dx(this)
class(periodic_2nd_order), intent(in) :: this
class(field) ,allocatable :: df_dx
integer(ikind) :: i,nx, me, east, west
real(rkind) :: dx
real(rkind), allocatable :: tmp_field_array(:)

121
122
123

nx=size(this%global_f)
dx=2.*pi/(real(nx,rkind)*num_images())

124
125

allocate(df_dx,tmp_field_array(nx))

126
127

me = this_image()

128
129

if (me == 1) then

12.1 Toward a Scalable Abstract Calculus

130
131
132
133
134
135
136
137
138

west = num_images()
east = 2
else if (me == num_images()) then
west = me - 1
east = 1
else
west = me - 1
east = me + 1
end if

139
140
141

tmp_field_array(1) = &
0.5*(this%global_f(2)-this%global_f(nx)[west])/dx

142
143
144

tmp_field_array(nx) = &
0.5*(this%global_f(1)[east]-this%global_f(nx-1))/dx

145
146
147
148
149

do i=2,nx-1
tmp_field_array(i)=&
0.5*(this%global_f(i+1)-this%global_f(i-1))/dx
end do

150
151
152

df_dx = tmp_field_array
end function

153
154
155
156
157
158
159

function d2f_dx2(this)
class(periodic_2nd_order), intent(in) :: this
class(field) ,allocatable :: d2f_dx2
integer(ikind) :: i,nx, me, east, west
real(rkind) :: dx
real(rkind), allocatable :: tmp_field_array(:)

160
161
162

nx=size(this%global_f)
dx=2.*pi/(real(nx,rkind)*num_images())

163
164

allocate(d2f_dx2,tmp_field_array(nx))

165
166

me = this_image()

167
168
169
170
171
172
173
174
175
176
177

if (me == 1) then
west = num_images()
east = 2
else if (me == num_images()) then
west = me - 1
east = 1
else
west = me - 1
east = me + 1
end if

178
179
180
181

tmp_field_array(1) = &
(this%global_f(2)-2.0*this%global_f(1)+this%global_f(nx)[west])&
/dx**2

182
183
184

tmp_field_array(nx) =&
(this%global_f(1)[east]-2.0*this%global_f(nx)+this%global_f(nx-1))&

315

316

Multiphysics Architectures
185

/dx**2

186
187
188
189
190
191

do i=2,nx-1
tmp_field_array(i)=&
(this%global_f(i+1)-2.0*this%global_f(i)+this%global_f(i-1))&
/dx**2
end do

192
193
194
195

d2f_dx2 = tmp_field_array
end function
end module
Figure 12.7. Denition of periodic 2nd-order central difference eld class contains a coarray
global eld. Operation results are stored locally using eld type.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20

Figure 12.8
module field_module
use kind_parameters ,only : rkind, ikind
implicit none
private
public :: field
type :: field
real(rkind), allocatable :: f(:)
contains
procedure :: add
=> total
procedure :: subtract
=> difference
procedure :: multiply_field => product
procedure :: multiply_real => multiple
procedure :: assign
=> assign_field_f
procedure :: copy
=> copy_filed
procedure :: state
=> field_values
generic
:: operator(+)
=> add
generic
:: operator(-)
=> subtract
generic
:: operator(*)
=> multiply_real,multiply_field
generic
:: assignment(=) => assign, copy
end type

21
22

contains

23
24
25
26
27
28

function field_values (this)

class(field), intent(in) :: this
real(rkind), allocatable :: field_values(:)
field_values = this%f
end function

29
30
31
32
33
34

subroutine assign_field_f (lhs, rhs)

class(field), intent(inout) :: lhs
real(rkind), intent(in) :: rhs(:)
lhs%f = rhs
end subroutine

35
36
37
38
39

subroutine copy_filed (lhs, rhs)

class(field), intent(inout) :: lhs
class(field), intent(in) :: rhs
lhs%f = rhs%f

12.2 Case Studies: Multiphysics Modeling

40

end subroutine

41
42
43
44
45
46
47
48

function total(lhs,rhs)
class(field) ,intent(in) :: lhs
class(field) ,intent(in) :: rhs
class(field) ,allocatable :: total
allocate (total)
total%f = lhs%f + rhs%f
end function

49
50
51
52
53
54
55
56

function difference(lhs,rhs)
class(field) ,intent(in) :: lhs
class(field) ,intent(in) :: rhs
class(field) ,allocatable :: difference
allocate (difference)
difference%f = lhs%f - rhs%f
end function

57
58
59
60
61
62
63
64

function product(lhs,rhs)
class(field) ,intent(in) :: lhs
class(field) ,intent(in) :: rhs
class(field) ,allocatable :: product
allocate(product)
product%f = lhs%f * rhs%f
end function

65
66
67
68
69
70
71
72
73

function multiple(lhs,rhs)
class(field) ,intent(in) :: lhs
real(rkind) ,intent(in) :: rhs
class(field) ,allocatable :: multiple
allocate(multiple)
multiple%f = lhs%f * rhs
end function
end module
Figure 12.8. Denition of eld type that stores operation results from periodic_2nd_order
class.

12.2 Case Studies: Multiphysics Modeling

This section presents several of the case studies from which we culled the patterns
that appear in this book. In addition to inspiring some of the patterns, these cases
demonstrate the broad spectrum of physics subdisciplines to which the design principles in this book apply. Fluid dynamics lies at the core of each problem. Characteristic
of multiphysics modeling, we couple numerous other phenomena to the ows under
study. These phenomena include quantum vortices, magnetic induction and diffusion, electromagnetic wave propagation, and solid particle transport. We also vary
the representation of the uid ow itself from the most common, a deterministic
continuum mechanics approximation (the Navier-Stokes and the magnetic induction equations), to the most exotic, a discrete vortex lament approximation (the
superuid equation and Biot-Savart law), and nally an inherently probabilistic
particle representation (the lattice Boltzmann equation).

317

318

Multiphysics Architectures

The following sections briey describe the physics, the software architecture, and
the simulation results. We completed most of the related work before the availability
of Fortran 2003 compilers. As such, some of the symbols in the associated UML
diagrams represent constructs or relationships we emulated without direct language
support namely the interface implementation (inheritance) relationship symbol and
the abstract class notation. The emulation techniques build upon the programming
styles outlined by Decyk et al. (1997b, 1997a, 1998) and Akin (2003) and supported
by the memory management infrastructure described by Rouson et al. (2006).

12.2.1 Quantum Turbulence in Superuid Liquid Helium

Section 4.3.2 provides an overview of the superuid phenomenology. At sufciently high driving velocities, the quantum vortices described in that section
become a complicated tangle often referred to in the literature as quantum turbulence (Vinen and Niemela 2002). Despite the microscopic difference between this
turbulent state and that of a classical uid, numerous laboratory experiments have
demonstrated that the macroscopic statistics closely resemble classical turbulence.
The aim of the simulations was to investigate a widely hypothesized explanation
for this similarity: vortex locking, the bundling of quantum vortices colocated

Figure 12.9. Quantum vortices and classical vortices in turbulent superuid liquid helium.
Image courtesy of Morris (2008). For a color version of this image, please visit
http://www.cambridge.org/Rouson.

12.2 Case Studies: Multiphysics Modeling

and aligned with their normal-uid counterparts so that the overall two-uid mixture moves in tandem (Morris et al. 2008). Figure 12.9 demonstrates this bundling,
colocation, and alignment.
The Navier-Stokes equations govern the normal uid velocity u:
ut = u u p + 2 u + f
u = 0

(12.4)
(12.5)

where p is the hydrostatic pressure, is the kinematic viscosity, and f is the mutual
friction force between the superuid and normal uid. The total material velocity is u + v. Morris et al. (2008) solved a transformed system of two equations: one
corresponding to the Laplacian of one component of equation (12.4) and a second
corresponding to the parallel component of the curl of equation (12.4). Choosing
the second component in the Cartesian system {x1 , x2 , x3 } and dening the vorticity
u and the nonlinear product H u yields:
2

( H) + 2 H2 + 2 ( 2 u2 )
u2 =
t
x2

2 = ( H)2 + 2 2
t

(12.6)
(12.7)

where the mutual friction force was dropped under the assumption that it inuences
the normal uid motion much less than it does the massless superuid vortex lines.
The advantage of equations (12.6)(12.7) over (12.4) lies in the global conservation of
energy (when = 0) as well as in the reduction of the number of dependent variables
from four to two.11
Section 4.3.2 presented the superuid evolution equations (4.7)(4.8), which we
repeat here for convenience:

(S r) dS

v=
(12.8)
4
S r3
dS
= v S (u v)  S [S (u v)]
dt

(12.9)

Section 4.3.2 denes each of the quantities in the latter two equations.
Morris et al. (2008) solved equations (12.6)(12.7) on the semiopen, cubical domain
xj [0, 2 ), j = 1, 2, 3 with periodic boundary conditions in each direction.
Figure 12.10 shows the class model emulated by the solver of Morris et al. (2008).
At the heart of this class diagram are two physics abstractions: a Fluid class
that encapsulates equations (12.6)(12.7) and a Tangle class that encapsulates
equations (4.7)(4.8). The depicted Integrand abstract class species the interface
physics abstractions must satisfy in order to be integrated over time by the marching
algorithm of Spalart et al. (1991). Appendix A describes that scheme. An exercise
at the end of the current chapter asks the reader to write ABSTRACT CALCULUS versions of the governing equations and to enhance the class diagram to show the public
methods and the likely private data that would support these expressions.
11 The differentiations required to go from equation (12.4) to equation (12.7) destroy information

that can be recovered by solving the original equations averaged over the two directions of
differentiation: x1 and x3 . See Rouson et al. (2008b) for more detail.

319

320

Multiphysics Architectures
Integrand

Superfluid

Tangle

Fluid
1
2
Field

Figure 12.10. Quantum turbulence solver architecture.

Two patterns appear in Figure 12.10. The Superuid class embodies the PUPPETEER pattern, whereas the Integrand class embodies an ABSTRACT CALCULUS for
time integration. That gure also presents several additional opportunities for using
patterns. Whereas the Field class embodies an ADT calculus for spatial differentiation, use of the ABSTRACT CALCULUS pattern would increase the exibility of the
design by replacing the concrete Field with an abstract class. To facilitate this, one
might also add an ABSTRACT FACTORY and a FACTORY METHOD to the design. Furthermore, the time-integration process could be made more exible via a STRATEGY and
a supporting SURROGATE. Another exercise at the end of the chapter asks the reader
to add these elements to the design.
In addition to the time advancement role the Superuid PUPPETEER plays, it also
shuttles information between the Fluid and the quantum vortex Tangle. Specically,
the Superuid queries the Tangle for the locations of each vortex lament grid point.
The Superuid queries the Fluid for the local velocities at each location vortex point
and passes these velocities to the Tangle for use in equation (4.8). Likewise, in a fully
consistent simulation that accounts for the inuence of the quantum vortices on the
normal uid motion, the Superuid would query the Tangle for the mutual friction
vectors at each vortex point and pass these to the Fluid for incorporation into the
Navier-Stokes equation (12.4).
In many applications, the Superuid might also referee the competing stability
and accuracy requirements of its puppets. A likely algorithm for doing so would
involve polling the puppets for their suggested time steps based on their private evolution equations and state vectors at a given point in time. The Superuid would then
choose the smallest answer so as to satisfy the accuracy and stability requirements
of each puppet. It might be useful to design an abstract POLLSTER to conduct such
polls when passed a pointer array of poll respondents. To facilitate this, each physics
abstraction could extend a Respondent type, and a polymorphic poll() procedure
could be written to collect the responses and set the time step.

12.2 Case Studies: Multiphysics Modeling

12.2.2 Lattice-Boltzman Biouid Dynamics

This section briey introduces the lattice Boltzmann equation (LBE) method and
an applies it to a biouid engineering problem: blood ow around a stent. The
Boltzmann equation originates from the kinetic theory of gases with the two assumptions: binary collisions and a lack of correlation between a molecules velocity
and its position (Huang 1963). The Boltzmann equation follows:
f
+ ef = (f )
t

(12.10)

with the primary variables f (x, e, t) representing the particle probability distribution
functions, x describing the particles locations, e denoting the their velocities, and 
containing collisional information. For incompressible ow, the collision term can
be simplied using Bhatnagar-Gross-Krook (BGK) model (Bhatnagar et al. 1954)
as (f f eq )/ , where f eq is the particle distribution function at the state of thermal
equilibrium and is the relaxation time. We restrict ourselves to the BGK model in
this section.
The LBE method borrows ideas from cellular automata uids (Frisch et al. 1986).
A regular lattice covers the physical space domain. Particles are distributed in each
lattice with probability distribution function f . This approach can also be viewed as a
nite-difference approximation to the Boltzmann equation (He and Luo 1997). Particles travel to neighbor lattices along different directions and collide with each other.
Figure 12.11 shows the lattice geometry and velocity vectors in the two-dimensional
nine-speed D2Q9 model(He and Luo 1997). A rst-order-accurate discretization in
time leads to:
1
eq
fi (x + ei t, t + t) fi (x, t) = (fi (x, t) fi (x, t))

(i = 0, 1, , 8)

(12.11)

where the particle velocities ei is given by:

(0, 0)
ei = c(cos(i 1)/2, sin(i 1)/2)

2c(cos(2i 9)/2, sin(2i 9)/2)

e2

e5

e6
e0

e1

e3
e8
e7
Figure 12.11. The lattice of D2Q9 model.

e4

i=0
i = 1, 2, 3, 4
i = 5, 6, 7, 8

(12.12)

321

322

Multiphysics Architectures

where here c = x/t is the lattice velocity. The equilibrium form of the distribution
function is approximated as:


1
1
1
eq
2
fi = i 1 + 2 ei u + 4 (ei u) 2 u u
(12.13)
cs
2cs
2cs

where cs = c/ 3 is the LBE sound speed (He and Luo 1997), the weighting functions 0 = 4/9, i = 1/9 for i = 1, 2, 3, 4, and i = 1/36 for i = 5, 6, 7, 8. The uid density
and velocity u is obtained by:
=

8

i=0

fi ,

u =

8


ei fi

(12.14)

i=0

This completes the mathematical model.

The LBE method has been applied to many engineering and science problems
including turbulence (Cosgrove et al. 2003), porous media ow (Manz et al. 1999),
multiphase ows (Boyd et al. 2004) and physiological ows (Artoli et al. 2006). One
of the advantages of using the LBE method is that data communications between
lattices are always local, which makes the method extremly efcient for large-scale
parallel computations. Since LBE method is derived based on kinetic equations in
mesoscopic scales, it is also very useful for as a scale-bridge for multiscale simulations.
The motivation for the work described here stems from the problems in the
use of stents in aneurysms. Despite stents ability to reduce blood ow into an an
aneurysm sac, complications occur that require a deeper understanding of the uid
dynamics. Xu and Lee (2008) used D2Q9 LBE method to perform a simulation of
ows in aneurysms with and without a transverse stent obstacle on the upper wall to
analyze the ow inside the artery. Won et al. (2006) investigated problems encountered during and after stentgraft treatment, including stentgraft migration, stentgraft
folding, cerebral ischemia and mechanical failure. The porosity and size of the stent
and the difculty in implementation to complex vessel geometry are notable problems. The modication of the ow eld also causes coagulation in the aneurysm
leading to its permanent occlusion after treatment(Lieber et al. 1997). Therefore,
minimal ow changes by the disease treatment and the ease of implementation to
the patient-specic geometry are very important.
As shown in Figure 12.12, different shapes of obstacles situated before an
aneurysm were used. The parent vessel is 4mm in diameter, and the diameter of
the sac of the aneurysm is 5mm. Because the vessel diameter is larger than 1mm
(i.e., it falls in the region of high strain rates), non-Newtonian effect associated with
complicated shear-strain relationships are small. The ow is therefore assumed to be
Newtonian uid with a dynamic viscosity of 3.5106 m2 /s and a density of 1060kg/m3 .
For simplicity, the vessel wall is assumed rigid. At the entry of the artery, a pulsatile
inlet velocity was used. Figure 12.13 compares contours of the vorticity (the curl of
the velocity) between nonstented and different stent (obstacle) shapes. It shows that
the rectangular stents with height-to-width ratio H/d = 0.2 resulted in the largest
vorticity reduction,
Figure 12.14 shows the class model emulated in the LBE solver. A Fluid class
implements the blood characteristics by encapsulating equation 12.11. The streamand-collide time marching algorithm is implemented in the Integrand abstract class.

12.2 Case Studies: Multiphysics Modeling

da=5 mm
Stent
d=4 mm

40 mm

Figure 12.12. The computational domain of the aneurysm model.

Vorticity

t=0.3T

1E-06
0
1E-06
1.2E-06
1.4E-06
1.6E-06
1.8E-06
2E-06
2.2E-06
2.4E-06
2.6E-06
2.8E-06
3E-06

No stent

t=0.3T

t=0.3T

Triangular stent
(0.1dH_0.1dL)
t=0.3T

t=0.3T

Rectangular stent
(0.1dH_0.1dL)
t=0.3T

t=0.3T

Rectangular stent
(0.1dH_0.3dL)

t=0.3T

Half circular stent

(0.1dH_0.2dL)

Triangular stent
(0.2dH_0.1dL)

Half circular stent

(0.1dH_0.4dL)

Rectangular stent
(0.2dH_0.1dL)

t=0.3T

Half circular stent

(0.2dH_0.4dL)

t=0.3T

Rectangular stent
(0.2dH_0.3dL)

t=0.3T

Half circular stent

(0.2dH_0.6dL)

Figure 12.13. Vorticity contours inside the sac of aneurysm under the effects of different shapes
of stent. For a color version of this image, please visit http://www.cambridge.org/Rouson.

Using the basic two-step stream-and-collide algorithm, equation 12.11 can be

discretized into two equations as:
1
eq
Collisionstep :fi (x, t) = fi (x, t) (fi (x, t) fi (x, t))

StreamingStep :fi (x + ei t, t + t) = fi (x, t)

(12.15)
(12.16)

The lattice information is stored in the Grid class, including the additional boundary
lattice information used to treat the complicated geometries. The information in the
Grid class is hidden from Fluid class and can only be accessed from the Field class,
which represents the particle distribution functions in nine directions.
With this code structure, the solver can be extended to three-dimensional code
by simply modifying the Field and Grid classes to 3D domains while keeping the

323

324

Multiphysics Architectures
Integrand

Fluid
1
1
Field
1
1
Grid

Figure 12.14. Class diagram of LBE solver.

Integrand

Blood
1
1

Plasma

RBC
1

1
1

Plasma_Field

RBC_Field
1
1
Grid

Figure 12.15. Class diagram of a multicomponent LBE solver.

Fluid and Integrand classes the same. One could also develop a general solver for
both 2D and 3D ows by replacing the Field and Grid classes with ABSTRACT FACTORY
class and adding FACTORY METHOD to 2D or 3D application. For the bioengineering
problems we discussed in this section, the blood is treated as a Newtonian uid due to
the large diameter of the artery is. When people are more interested in better understanding the blood characteristics, detailed numerical simulations will be needed to
study the interactions among the blood plasma, blood cells, and elastic vessels. Due
to the different properties of blood plasma and blood cells, a multicomponent solver
would be needed for such a study. Figure 12.15 shows a class diagram that extends
the single-component LBE solver to a multicomponent solver. The RBC class represents the red blood cells in the blood plasma. The PUPPETEER pattern implemented
in the Blood class facilitated the extension.

12.2 Case Studies: Multiphysics Modeling

12.2.3 Particle Dispersion in Magnetohydrodynamics

Magnetic elds often provide the only practical means for pumping, stirring, and
otherwise controlling the motions of molten metals (Mulcahy 1991; Lu et al. 2002).
In one important class of industrial applications, semisolid metals processing, the
medium so manipulated consists of solid nucleation sites dispersed throughout a
liquid metal casting. The mechanical properties of semisolids render them highly
manipulable, offering the prospect of manufacturing so-called near-net-shape products that require little ling or other materially wasteful nal preparations. Semisolids
might well be found also in natural settings near solid/liquid phase boundaries such
as the inner edge of the earths liquid mantle.
Particle dispersions in purely hydrodynamical ows that is those without
magnetic eld effects often exhibit preferential concentration, wherein particles
concentrate in certain ow regions. Less is known about the particle dispersion in
magnetohydrodynamic turbulent ow. One can investigate this behavior in semisolid
metals by solving the magetohydrodynamics (MHD) equations coupled to a set of Lagrangian equations that describe the motion of immersed particles. MHD combines
the Navier-Stokes equations with Maxwells equations in the form of the magnetic
induction equation (Davidson 2001). The incompressible, homogeneous MHD ow
can be described by the incompressible MHD equations:
ui,i = 0

(12.17)

bi,i = 0

(12.18)

ext
t ui = P,i + Hi + ui,jj + bj bi,j + Bext
j bi,j + Bi,j bj

(12.19)

ext
t bi = uj bi,j + bj uj,i uj Bext
i,j + Bj ui,j + bii

(12.20)

where ui and bi are the components of velocity and magnetic uctuation elds,
is the magnitude of
is the kinematic viscosity, is the magnetic diffusivity, Bext
i
externally applied magnetic eld, P = p/ + 1/2ui ui is the modied pressure, is the
uid density, p is the uctuating pressure, Hi = i,j,k uj k , i,j,k is the alternating unit
symbol, and i = i,j,k uk,j is the vorticity. Assuming particles follow the drag law of
Stokes (Stokes 1851), the equations of motion for each particle are:
dr
=v
dt
dv
1
= [u(r, t) v]
dt
p
p

2p a2
9

(12.21)
(12.22)
(12.23)

where r and v are the particle position and velocity, receptively; u(r, t) is the undisturbed uid velocity in the neighborhood of the particle; and p , a and p are the
particles hydrodynamic response time, radius, and material density, respectively.
Equation (12.22) balances particle inertia (as characterized by its mass 4p a3 /3)
against the molecular momentum exchange integrated over the particle surface (as
characterized by the transport coefcient ). The undisturbed uid velocity, u(r, t), interpolated to particle location, r, estimates the appropriate freestream velocity in the

325

326

Multiphysics Architectures

Y
Z

Figure 12.16. Planar projection of inertial particles in liquid-metal MHD turbulence under
the inuence of an externally applied magnetic eld orthogonal to the viewing plane:
fastest (red), slowest (blue). For a color version of this image, please visit http://
www.cambridge.org/Rouson.

neighborhood of the particle. The dependence on p indicates that denser particles

respond more slowly to changes in the neighboring uid velocity. The dependence
of the dynamic viscosity, , and particle radius a indicates that smaller particles in more viscous uids more rapidly track changes in the local uid velocity.
The particles can be treated as passive scalar when the particle volume fraction
is low.
Rouson et al. (2008a) simulated the particle dispersion in the initially isotropic
MHD turbulence under the inuence of a spatially constant external magnetic eld
at low magnetic Reynolds number. In MHD turbulence, the particles tend to cluster
near the large-strain regions where both magnetic led and the uid velocity eld
appear strong uctuations. Figure 12.16 shows the particle positions projected onto
the plane which is perpendicular to the direction of magnetic force. The apparent
voids driven by the large strain represent approximately cylindrical evacuated regions
extending the entire length of the problem domain, thin sheets of particle clusters
surrounding the voids are observed.
The class model shown in Figure 12.17 emulates the solver coupling MHD and
Lagrangian particle equations. The Magnetouid class represents the MHD physical abstraction and encapsulates equations (12.17)(12.20). The Cloud class is the
abstraction of inertial particles and encapsulates equations (12.21)(12.23). The time
marching algorithm of Spalart et al. (1991) is implied in the Integrand abstract class.
The PUPPETEER pattern is implemented in two classes: Semisolid class and Magnetouid class, whereas the ABSTRACT CALCULUS pattern appears in the Integrand
class.
12.2.4 Radar Scattering in the Atmospheric Boundary Layer
Remote sensing observations via lidar and millimeter-wavelength radar are revealing
the ne-scale features of turbulence in the earths atmosphere with unprecedented
detail. The length scales at which researchers have recently observed scattered return
signals in clear atmospheres, however, pose challenges for both the theoretical

12.2 Case Studies: Multiphysics Modeling

Integrand

Semisolid

Cloud

Magnetofluid

Magnet

Fluid
1
2

1
3
Field

Figure 12.17. Semisolid MHD solver architecture.

understanding and computational exploration (Linde and Ngo 2008). Theoretically,

the standard statistical treatments of turbulence posit a spectrum of eddy sizes
that cuts off above the lengths required to scatter radar at the observed frequencies (Tatarskii 1961; Atlas et al. 1966; Doviak and Zrnic 1984). Computationally,
the broadness of the spectrum of eddy sizes makes direct numerical simulation (DNS)
of all dynamically relevant scales infeasible on even the largest and fastest computers.
To gain insights while rendering the problem tractable, Morris et al. (2010)
recently performed simulations of the turbulent Ekman layer, a model problem
relevant to the dynamics of the atmosphere boundary layer (ABL) near the earths
surface. In the atmosphere, the Ekman layer is the vertical region over which the
wind transitions from its prevailing geostrophic direction imposed by atmospheric
pressure imbalances to its near-surface direction strongly inuenced by the earths
rotation and friction at the ground. Figure 12.18 shows the geometry and coordinatesystem conguration imagined in an Ekman-layer simulation. In the case Morris and
colleagues studied, = /2; hence, the angular velocity vector  points along the x3
direction. The simulation domain is a small rectangle with one boundary containing
the x1 -x2 ground plane in Figure 12.18.
Morris et al. (2010) solved a dimensionless form of equations (12.4)(12.5):
ut = u u p
u = 0

1
1
e (u eG ) + 2 u
Ro
R

(12.24)
(12.25)

where the velocity u has components u1 , u2 , and u3 in the x1 , x2 , and x3 coordinate

directions, which Morris et al. refer to as the streamwise, wall-normal, and spanwise
directions, respectively. In equations (12.4)(12.5), all variables are rendered nondimensional by normalizing them by the geostrophic wind speed G, the computational
domain height , and the uid density . The dimensionless parameters R G/

327

328

Multiphysics Architectures

x2

x3
x1

Figure 12.18. Geophysical coordinate system used for deriving the governing equations: x1 ,
x2 , and x3 point to the East (into the page), the North, and the vertical directions, respectively.

and Ro G/(2) are the Reynolds and Rossby numbers; is the kinematic viscosity;  is the angular velocity of the system; and p is the hydrodynamic pressure plus
appropriate potentials.
The unit vectors associated with the system rotation and the geostrophic velocity
are e and eG , respectively. In equation (12.24), eG = ex1 , and e = ex2 . These choices
generate Coriolis forces solely in the horizontal (x1 -x3 ) plane and an effective driving
uniform pressure gradient in the negative x3 direction.
Figure 12.19 depicts the physics of interest in simulating the ABL for purposes of
understanding radar return signals. Figure 12.19(a) demonstrates streamwise velocity
component contours characteristic of vertical plumes that spread horizontally away
from the ground. Figure 12.19(b) demonstrates the high mechanical energy dissipation rates that surround the depicted plumes. Dissipative eddies make up the
ow structures with the shortest length and time scales and therefore represent
the most likely candidates for scattering short-wavelength electromagnetic waves.
Morris et al. demonstrated highly non-Gaussian statistics of the dissipation and the
corresponding much greater likelihood of the presence of the smallest-scale features
than would otherwise be estimated based on average dissipation rates.
A useful next step in understanding the anomalous return signals would be to
predict variation of quantities known to cause scattering. In the scattering mechanism identied by Tatarskii (1961), this principally includes the relative humidity.
Additionally, the relevant plume dynamics would likely be inuenced by other scalar
thermodynamic variables such as the absolute temperature. A complete study would
then involve predicting the return signals of the simulated scattering elds.
Figure 12.20 diagrams a solver architecture for simulating scalar transport in the
ABL. The depicted Atmosphere PUPPETEER aggregates zero or more scalar quantities
that would presumably be simulated via a scalar advection/diffusion equation of the
form:
1
(12.26)
t = u + 2
Sc
where is the transported scalar and Sc G/  is the dimensionless parameter
characterizing the relative importance of bulk transport and molecular diffusion with
a diffusion coefcient . In cases where these scalars actively inuence the air ow
for example, temperature variations that impart momentum to the air via buoyancy

12.2 Case Studies: Multiphysics Modeling

0.40

0.30

0.20

0.10

0.00
1.6

y
x

1.8

2.0
8

2.2

3.250645360-14

2.4

2.6

16

12

17.66129438678

(a)
0.40

0.30

0.20

0.10

0.00
1.6

y
x

1.8
z

2.0

0.0001

2.2
Preido-Dispation
0.001
0.01

4.98622067E-06

2.4

2.6

0.1
0.819025337696

(b)
Figure 12.19. Instantaneous contours in a vertical slice of an Ekman layer: (a) wind velocity component, and (b) pseudo-dissipation. For a color version of this image, please visit
http://www.cambridge.org/Rouson.

forces the Atmosphere abstraction would mediate the communication of coupling

terms in the manner described in Chapter 8.
The multiphysics solver discussions in Sections 12.2.112.2.3 neglected the impact of boundary condition specication on the solver architectures. The simplicity of
the geometry and boundary conditions in those cases facilitated the implicit inclusion
of boundary constraints on a Field into the the Field ABSTRACT CALCULUS abstraction
itself. Ultimately, realistic simulation of the ABL problem is likely to require studying a sufciently broad a range of boundary condition types, values and geometries
as to greatly increase the complexity of the Field class should it subsume all of the
related possibilities. In such cases, Shalloway and Trott (2002) recommend, Find
what varies and encapsulated it. The Boundary class in Figure 12.20 thus encapsulates the boundary condition values, types, and geometry to insulate the Field class
from the associated complexities. In the aforementioned rectangular box geometry, a
Field would aggregate six boundaries, each corresponding to a face of the box. More
complicated ground topographies might include additional boundaries by immersed

329

330

Multiphysics Architectures

Integrand

Atmosphere

Fluid

Scalar

Field
1
6..
Boundary
Figure 12.20. Atmospheric boundary layer class diagram.

boundary methods (Mittal 2005) or by mapping the rectangular box to a deformed

but topologically consistent shape.
Morris et al. employed a modied version of the procedural, serial Fortran 77
solver developed by Bernard et al. (1993). As a rst step toward updating that solver,
we have wrapped several of the Bernard et al. (1993) procedures using the techniques
Chapter 2 outlined for wrapping legacy Fortran. The resulting architectural exibility
eases the inclusion of new physics and alternative numerics and opens up the possibility of progressive replacement of procedures with scalable numerical algorithms from
Trilinos, leading to faster simulations with increased resolution of the ow features
of interest. The completion of this new architecture remains an ongoing effort.

12.3 The Morfeus Framework

Morfeus is the Multiphysics Object-oriented Recongurable Fluid Environment for
Unied Simulations. The acronym is one transliteration for the name of the ancient Greek god more commonly translated to English as Morpheus. Morpheus
had the ability to appear in peoples dreams while taking on many different forms.
The Morfeus project aims to bring this sort of exibility to scientic software design via patterns. Morfeus also aims to be the rst such pattern-based framework
implemented in object-oriented Fortran 2003. It thereby provides a straightforward
migration path to OOD patterns and OOP for Fortran programmers, for most of
whom both OOP and patterns present signicant new conceptual leaps.
Morfeus consists of an open-source repository12 of interfaces (abstract classes)
and example concrete implementations of those interfaces in the context of
12 As of this writing, the Trilinos project leadership has agreed to host Morfeus in a repository for

terminal packages, that is, applications that build atop Trilinos packages but on which Trilinos does

12.3 The Morfeus Framework

multiphysics applications. A suggested Morfeus use case involves end-users substituting their own concrete classes for the Morfeus examples and writing applications
that access those concrete classes only through the interfaces Morfeus publishes. The
determination of what concrete implementation gets invoked would then happen at
runtime via dynamic polymorphism. Morfeus provides only the structure through
which user-dened methods get invoked on user-dened objects. Morfeus thus exemplies a framework in the manner Gamma et al. (1995) contrast with applications
and libraries: An application is any code that has a main program, whereas a library
is code that user code calls, and a framework is code that calls user code.
Early candidates for inclusion in Morfeus include a tensor eld ABSTRACT CALCULUS, including scalar, vector, and dyadic tensor eld interfaces. Each of these
will publish generic operators and corresponding deferred bindings for arithmetic
and spatial differentiation, integration, and arithmetic. These will be based on the
coordinate-free programming concepts of Grant, Haveraaen, and Webster (2000),
and will be developed in collaboration with the second of those three authors. In
Fortran syntax, the coordinate-free programming paradigm involves writing differential operations on tensors in forms such as .grad.,.div., and .curl. without
specifying a coordinate system, deferring to a lower-level abstraction the resolution
of these forms into coordinate-specic calculations based on the selection of a manifold in which the tensors are embedded. That manifold could provide a metric tensor
in the case of general curvilinear coordinates or scale factors in the special case of
orthogonal coordinates (Aris 1962). Figure 12.21 depicts several tensor eld interfaces. That gure species no relationships between the abstract classes that describe
the interfaces. An exercise at the end of the chapter asks the reader to discuss the
pros and cons of various approaches to relating those interfaces based on the design
principles discussed in this book.
The problem decomposition described in Figure 12.21 carries with it at least two
scalability benets. First, it naturally lends itself to a purely functional, side-effectfree programming style that supports asynchronous expression evaluation as noted in
Section 12.1.2. This becomes increasingly important as massively parallel hardware
platforms evolve toward communication dominating computation in the run-time
budget. Synchronization implies communication. A design patterns that lights a
path to highly asynchronous computation thus lights a path to massively parallel
computation. In this sense, scalable design leads directly to scalable execution.
Second, Numrich (2005) demonstrated a beautiful, formal analogy between array indices and contravariant tensor indices, and likewise between coarray indices
and covariant tensor indices. This analogy naturally generates a powerful notation
for expressing parallel algorithms for matrix and vector algebra. Moreover, the analogy supports very direct mapping of objects onto distributed memory via co- and
contravariant partitioning operators. All of this enables an exceptionally direct translation of parallel mathematical algorithms into compilable, efcient code that closely
mirrors the original mathematical syntax. This makes coarray Fortran 2008 an obvious candidate for inclusion into an ABSTRACT CALCULUS. We envision a scenario in
which continuum tensor calculus expressions written at the highest abstraction levels
not, in turn, depend. The rst release of the Morfeus source is expected to occur around the time
of publication of this book.

331

332

Multiphysics Architectures
Physics
+ t(Physics) : Physics
+ L(Physics) : Physics
+ N(Physics) : Physics
Field
<<operator>>
+ grad(Field) : VectorField
+ total(Field,Field) : Field
+ difference(Field,Field) : Field
+ product(Field,Field) : Field
+ laplacian(Field) : Field

{ t = L(this) + N(this) }
Manifold
metric : Dyad
grid : Array<Points>

Dyad

VectorField
<<operator>>
+ grad(VectorField) : Dyad
+ div(VectorField) : Field
+ curl(VectorField) : VectorField
+ dot(VectorField,VectorField) : Field
+ total(VectorField,VectorField) : VectorField
+ difference(VectorField,VectorField) : VectorField

Figure 12.21. Abstract classes in the Morfeus framework.

get resolved into discrete, tensor algebra expressions at the lowest abstraction levels,
and that these tensor algebra calculations get translated very directly into coarray
syntax via the analogy with covariant tensor forms.
Another early candidate for inclusion in Morfeus will be a Physics abstraction
that publishes an abstract interface for an evolution equation by dening a public time
derivative method, t(), as in Table 3.1. For Runge-Kutta schemes, t() represents
the RHS function evaluation in equations of the form 4.1. For semi-implicit time
advancement, wherein it proves useful to advance implicit terms differently from
explicit ones, the default implementation for t() might simply sum the linear and
nonlinear differential operators L and N, respectively, as indicated in Figure 12.21.
Appendix A.5.2.4 provides a description of the semi-implicit algorithm used in
several of the case studies described in Section 12.2.1.
Another candidate would be a Puppeteer abstraction that denes several default
methods for polling puppets for information that requires coordination such as time
step selection, cross-coupling terms and Jacobian construction. Additional design
patterns described in Part II of the current text also provide likely candidates either
on their own or incorporated into the other patterns. For example, the STRATEGY
and SURROGATE patterns could be incorporate into an ABSTRACT CALCULUS to provide exibility in the decision which discrete approximations will be employed to
approximate the continuous forms.
Between design and execution lies development. In this context, scalable development refers to practices for managing the complexities of programming writ large:
multideveloper, multiplatform, multilanguage, multipackage, multiuser software
construction. Scalable development methodologies must address build automation,
documentation, repository management, issue tracking, user feedback, and testing.

12.3 The Morfeus Framework

Hosting Morfeus on the Trilinos terminal application repository brings with it professional software engineering solutions to each of these problems. These solutions
include:
Nightly automated multi-platform builds via CMake,13
Nightly automated regression testing via the CMake tool CTest,
Automated e-mail notication of test failures with links to an online test matrix

dashboard constructed by the CMake tool CDash,

State-of-the-art repository management and version control via Git,14
User, developer, and leadership mailing lists managed by Mailman,15
Web-based issue tracking via Bugzilla,16
Automated hypertext markup of documentation by Doxygen.17

We expect that each of these technologies will contribute to an end-to-end, ubiquitous scalability that is, scalable design leading to scalable development leading
to scalable execution with an inclusive vision of domain-specic abstractions that
disruptively reduce design and development times and thereby bring more scientic
software developers into the once-protected domain of massively parallel execution.
[T]he Most Dangerous Enemy of a Better Solution Is an Existing
Codebase That Is Just Good Enough.
Eric S. Raymond
The Morfeus framework aims to increase the exibility of multiphysics software architectures by encouraging the use of patterns. In particular, it aims to
demonstrate a scalable ABSTRACT CALCULUS that enables application developers
to operate at an abstraction level characteristic of their application domain: tensor
calculus. Moreover, the promise of coarray Fortran 2008 suggests the possibility
of writing tensor-algebraic approximations to the tensor calculus in ways that
translate directly into coarray syntax via analogies with covariant tensors. Such
an approach could represent a disruptive technology if it meets the challenge of
overcoming attachments to working codes that get the job done for the small
subset of researchers with the skills and access to use them.

EXERCISES
1. The switch from 6th-order Pad differences to 2nd-order central differences can
be accomplished with minimal code revision by designing a new concrete factory periodic_2nd_factory to replace periodic_6th_factory in Figure 9.1(e).
Write this new factory and the revise the main program of Figure 9.1(a) to use
your new factory instead of periodic_6th_factory.
2. Calculate the cumulative, overall theoretical maximum speedups that could result
from progressive optimization of the procedures in Table 12.1.
13
14
15
16
17

http://www.cmake.org
http://git-scm.com/
http://www.gnu.org/software/mailman/
http://www.bugzila.org
http://www.doxygen.org

333

334

Multiphysics Architectures

3. Progressively augment the type-bound procedures in periodic_2nd_order module in Figure 12.1 with loop-level (parallel do) directives. Use the Fortran
cpu_time() and system_time() procedures to measure cumulative speedup on
1, 2, and 4 cores for progressive parallelization of the rst four procedures listed
in Table 12.1.
4. Write ABSTRACT CALCULUS versions of the governing equations (12.6)(12.9) as
they would appear inside a time derivative method on the Integrand class of
Figure 12.10. Enhance the class diagram in that gure to include the public methods that appear in your expressions. Further enhance that gure with private data
in each class that could support the calculations you specied.
5. Replace the Field class in Figure 12.10 with an ABSTRACT FACTORY and a FACTORY
METHOD capable of constructing a periodic_6th_order concrete product. Also
add a STRATEGY to the Integrand class diagram to facilitate swapping between
2nd- and 4th-order accurate Runge-Kutta algorigthms.
6. Discuss the pros and cons of choosing aggregation or composition versus
inheritance to relate the tensor eld abstractions in Figure 12.21.

APPENDIX A

Mathematical Background

This chapter provides the mathematical background necessary for the material in this
book. The level of the treatment targets seniors and beginning graduate students in
engineering and the physical sciences.

A.1 Interpolation
A.1.1 Lagrange Interpolation
Solving the quantum vortex problem described in Chapters 4 and 5 requires estimating the 3D uid velocity vector eld u(x, t) at each quantum vortex mesh point S.
Without loss of generality, we consider here interpolating a scalar eld u(x) that can
be interpreted as a single component of u at a given instant of time. We estimate the
desired values with 3D linear Lagrange polynomial interpolation. Given a collection of data points u(x0 , y0 , z0 ), u(x1 , y1 , z1 ), . . . , u(xk , yk , zk ) across a 3D space, each
Lagrange polynomial in a series expansion of such polynomials evaluates to unity
at the location of one datum and zero at the locations of all other data. In 1D, this
yields:
k

j (x) =
i=0,i=j

(x x0 ) (x xj1 )(x xj+1 ) (x xk )

x xi
=
.
xj xi (xj x0 ) (xj xj1 )(xj xj+1 ) (xj xk )

(A.1)

When the problem is well resolved in space, it can be preferable for both accuracy
and efciency reasons to use only the nearest-neighbor end points of the smallest
interval (in 1D) or corners of the smallest cube (in 3D) that brackets the point of
interest. Each polynomial for the 1D case then reduces to a single factor of the form:
j (x) =

(x xi )
.
(xj xi )

(A.2)

where xi is an endpoint of the interval between xi and xj . This facilitates writing the
series:
u(x) =

1


uj j (x)

(A.3)

j=0

335

336

Mathematical Background
(i +1,j +1,k +1)

(i,j +1,k+1)

(i +1,j,k +1)

(i,j,k+1)

(x,y,z)

(i,j,+1,k)

(i,j,k)

(i +1,j+1,k)

(i +1,j,k)

Figure A.1. 3D Lagrange interpolation in uniform grid.

In 3D, the j (x, y, z) can be obtained by multiplying 1D Lagrange polynomials:

j (x, y, z) = j (x)j (y)j (z)

(A.4)

In the book, we only use the nearest-neighbor points. Thus:

u(x, y, z) =

8


uj j (x, y, z)

(A.5)

j=0

where:
j (x, y, z) =

(x xi ) (y yi ) (z zi )
(xj xi ) (yj yi ) (zj zi )

(A.6)

For the uniform grid as shown in Fig. A.1, the distances between neighbor points in
x, y, and z direction are all equal to , and we have:
u(x, y, z) =

(x xi,j,k )(y yi,j,k )(z zi,j,k )

ui,j,k
3
(x xi+1,j,k )(y yi+1,j,k )(z zi+1,j,k )
+
ui+1,j,k
3
(x xi,j+1,k )(y yi,j+1,k )(z zi,j+1,k )
+
ui,j+1,k
3
(x xi+1,j+1,k )(y yi+1,j+1,k )(z zi+1,j+1,k )
+
ui+1,j+1,k
3
(x xi,j,k+1 )(y yi,j,k+1 )(z zi,j,k+1 )
+
ui,j,k+1
3
(x xi+1,j,k+1 )(y yi+1,j,k+1 )(z zi+1,j,k+1 )
+
ui+1,j,k+1
3
(x xi,j+1,k+1 )(y yi,j+1,k+1 )(z zi,j+1,k+1 )
+
ui,j+1,k+1
3
(x xi+1,j+1,k+1 )(y yi+1,j+1,k+1 )(z zi+1,j+1,k+1 )
+
3
ui+1,j+1,k+1

(A.7)

A.2 Linear Solvers

which provides a good approximation given well-resolved data. Higher-order

Lagrange polynomials offer formally better estimates but suffer from the possibility
of oscillations related to the lack of control of the polynomial derivatives. These
oscillations can be particularly accurate near the edges of the domain over which the
polynomial is dened. All polynomials ultimately diverge at a rate determined by
the order of the polynomial, and such divergence renders higher-order polynomials
inaccurate near domain boundaries.

A.2 Linear Solvers

A.2.1 Gaussian Elimination
Gaussian elimination solves a system of linear algebraic equations in the form:
Ax = b

(A.8)

when the determinant of matrix A does not vanish. When det A = 0, the system is
ill-posed and the equations have no unique solution. The basic Gaussian elimination
procedure includes two steps. The rst step, forward elimination, transfers the system
of equations into an upper triangular form. The second step solves the triangular
system using back substitution.
Consider the following equations:

a11
a21
..
.

a12
a22
..
.

..
.

a1n
a2n
..
.

an1

an2

ann

x1
x2
..
.

xn

b1
b2
..
.

(A.9)

bn

The augmented matrix of the above linear system is the matrix [A| b]:

a11
a21
..
.
an1

a12
a22
..
.
an2

..
.

a1n
a2n
..
.
ann

 b1

 b2


 .
 ..

 b
n

(A.10)

This system retains the same solution under certain elementary operations, including
row interchanges, multiplication of a row by a nonzero number, and replacement of
one row with the sum of two rows. Gaussian elimination uses these operations to
transform the equation system into upper triangular form. In the rst step, one
chooses an equation with a nonzero x1 coefcient as the pivot equation and
swaps this equations coefcients and RHS with those in the rst row of the augmented matrix. Upon switching, one zeroes out the x1 coefcients listed below the
pivot equation by multiplying the rst row of the augmented matrix by ai1 /a11 and
subtracting the result from the ith row for all i = 2, 3, , n.
A similar process based on choosing an equation with a nonzero x2 coefcient
as the pivot and swapping that equation with the second equation zeros ultimately
produces zeros below the diagonal in the second column. This process continues until

337

338

Mathematical Background

it produces an equivalent upper triangular system:

a11
0
..
.

a12
a22
..
.

..
.

a1n
a2n
..
.

ann

 b1
 
 b
 2
 .
 ..

 b
n

(A.11)

where primes denote values that have been modied by the Gaussian elimination
process. Upon obtaining the upper triangular form, the back substitution process
proceeds as follows: the bottom equation immediately yields:
xn = bn /ann .

(A.12)

The penultimate equation then yields:

xn1 = (bn1 ann xn )/an1

(A.13)

and so forth, working up the equation listing until the rst equation has been solved.
The accuracy of Gaussian elimination can deteriorate by round-off errors accumulated through the many algebraic operations. Approximately n3 /3 multiplications
and n3 /3 additions are required to solve n equations using Gaussian elimination. The
accuracy also depends on the arrangement of the equation system. Rearranging the
equation in order to put the coefcients with largest magnitude on the main diagonal
tends to improve accuracy. This rearrangement process is called partial pivoting.
Figure A.2 shows corresponding pseudocode.

1
2

input: A(n,n), b(n)

output: x(n)

Figure A.2

3
4
5
6
7
8
9
10
11

!transform matrix A to upper diagonal form

for i=1 to n-1 do
!find the pivot row with the maximum value in main diagonal
for j=i+1 to n do
if absolute(A(j,i))>absolute(A(i,i)) then
swap row i and row j of matrix A and vector b
end if
end for

12
13
14
15
16
17
18
19
20

!eliminate all terms in the ith colume of the rows below the pivot row
for j=i+1 to n do
for k=i to n do
factor=A(j,i)/A(i,i)
A(j,k)=A(j,k)-factor*A(i,k)
end for
end for
end for

21
22
23
24

!back substitute
x(n) = b(n)/A(n,n)
for i=n-1 to 1 do

A.2 Linear Solvers

25
26
27
28
29
30

known=0.0
for j=i+1 to n do
known=known+A(i,j)*x(j)
end for
x(i) = (b(i) - known)/A(i,i)
end for
Figure A.2. Pseudocode of Gaussian elimination with partial pivoting.

Chapter 3 demonstrates the application of Gaussian elimination to a linear system that arises in solving a n heat conduction problem, wherein one obtains a vector
of temperature values T n+1 at time tn+1 from the temperature vector T n at a previous
time step tn = tn+1 t for some time step t according to:


t 2 1 n
T
T n+1 = I

(A.14)

where I is the identity matrix with unit values along the diagonal and zeros elsewhere.
Given a uniformly spaced grid overlaid on the computational domain, applying the
central difference formula of Section A.5.1 to the RHS of equation (A.14) leads to
the linear algebraic equations in the form of (A.8), where:

A=

1+

2t
(x)2
t 2
(x)

..
.
0

t
(x)2
1 + 2t 2
(x)

..
.
0

..
.
0

..
.

0
..
.

0
..
.

t
(x)2

T1n+1
T2n+1
..
.

1 + 2t 2
(x)
t 2
(x)

t 2
(x)
1 + 2t 2
(x)

(A.15)

x=

n+1
Tm1
n+1
Tm
n
T1 + t 2 Tchip
(x)

T2n

..
b=
.

Tm1
n + t T
Tm
2 air

(A.16)

(A.17)

(x)

where m is the number of discrete node points at which one seeks n temperatures and
x is the uniform space between each node. Matrix A is diagonally dominant, in the
sense that the magnitude of each main diagonal element exceeds that of the other elements in any given row. Diagonal dominance eliminates the need for partial pivoting.
Figure (A.3) provides a Fortran subroutine implementation of Gaussian elimination
without partial pivoting useful for solving the above equations. Figure (A.4) presents
an equivalent C++ code.

339

340

Mathematical Background

When A is tridiagonal as in equation (A.15), one can simplify the Gaussian

elimination procedures and reduce the required storage by exploiting the zero coefcients. Thomas (1949) suggested this method. Using the Thomas algorithm, one
transforms the tridiagonal matrix in equation (A.15) to an upper triangular form by
replacing the diagonal coefcients with:

di = 1 +

2t
(x)2

2
t
(x)2
1 + 2t 2
(x)

i = 2, 3, , m

(A.18)

and replacing the vector b by:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31

T2

t
(x)2
1 + 2t 2
(x)

= T2n +


Tm

t
(x)2
1 + 2t 2
(x)

n
= Tm
+

T1n +

n
Tm
+

t
(x)2


Tchip

t
(x)2

(A.19)


Tair

(A.20)

Figure A.3
module linear_solve_module
use kind_parameters ,only : rkind,ikind
implicit none
contains
function gaussian_elimination(lhs,rhs) result(x)
real(rkind) ,dimension(:,:) ,allocatable ,intent(in) :: lhs
real(rkind) ,dimension(:) ,allocatable ,intent(in) :: rhs
real(rkind) ,dimension(:)
,allocatable :: x,b ! Linear system:
real(rkind) ,dimension(:,:) ,allocatable :: A
! Ax = b
real(rkind)
:: factor
real(rkind) ,parameter :: pivot_tolerance=1.0E-02
integer(ikind)
:: row,col,n,p ! p=pivot row/col
n=size(lhs,1)
b = rhs ! Copy rhs side to preserve required intent
if ( n /= size(lhs,2) .or. n /= size(b)) &
stop gaussian_elimination: ill-posed system
allocate(x(n))
A = lhs
! Copy lhs side to preserve required intent
!______ Gaussian elimination _______________________________________
do p=1,n-1
! Forward elimination
if (abs(A(p,p))<pivot_tolerance) &
stop gaussian_elimination: use pivoting
do row=p+1,n
factor=A(row,p)/A(p,p)
forall(col=p:n) A(row,col) = A(row,col) - A(p,col)*factor
b(row) = b(row) - b(p)*factor
end do
end do
x(n) = b(n)/A(n,n) ! Back substitution
do row=n-1,1,-1
x(row) = (b(row) - sum(A(row,row+1:n)*x(row+1:n)))/A(row,row)

A.2 Linear Solvers

32
33
34

end do
end function
end module
Figure A.3. Gaussian elimination procedure.

Ti = Tin +
where d1 = 1 +
according to:

2t
.
(x)2

t
(x)2
1 + 2t 2
(x)

n
Ti1


Tm
dm

Figure A.4(a)
#ifndef GAUSSIAN_ELIMINATION_H_
#define GAUSSIAN_ELIMINATION_H_ 1

3
4
5

#include <exception>
#include "mat.h" // 2-D array

6
7
8
9

struct gaussian_elimination_error : public std::exception {

virtual gaussian_elimination_error() throw () {}
};

10
11

crd_t gaussian_elimination (const mat_t & lhs, const crd_t & rhs);

12
13

1
2

#endif
Figure A.4(b)
#include "gaussian_elimination.h"
#include <cmath>

3
4
5
6
7

crd_t gaussian_elimination (const mat_t& lhs, const crd_t& rhs)

{
const real_t pivot_tolerance=0.01;
const int n = lhs.rows();

8
9
10
11
12
13

if ( n != lhs.cols() || n != rhs.size()) {
std::cerr << "gaussian_elimination: ill-posed system"
<< std::endl;
throw gaussian_elimination_error();
}

14
15
16
17

// copy parameters to preserve lhs and rhs

crd_t b = rhs;
mat_t A = lhs;

18
19
20
21

(A.21)

The unknowns can be solved from back substitution

n+1
=
Tm

i = 3, 4, , m 1

//----- Gaussian elimination ----// forward elimination

for (int p = 0; p < n-1; ++p) {

(A.22)

341

342

Mathematical Background
if (fabs(A(p,p)) < pivot_tolerance) {
std::cerr << "gaussian_elimination: use pivoting"
<< std::endl;
throw gaussian_elimination_error();
}
for (int row = p+1; row < n; ++row) {
real_t factor = A(row,p) / A(p,p);

22
23
24
25
26
27
28
29

for (int col = p; col < n; ++col) {

A(row,col) -= A(p,col)*factor;
}

30
31
32
33
34
35

36

b.at(row) -= b.at(p)*factor;

37

// Back substitution
crd_t x(n);

38
39
40

x.at(n-1) = b.at(n-1) / A(n-1, n-1);

41
42

for (int row = n-2; row >=0; --row) {

real_t sum = 0.0;

43
44
45

for (int col = row+1; col < n; ++ col) {

sum += A(row,col) * x.at(col);
}
x.at(row) = (b.at(row) - sum) / A(row, row);

46
47
48
49
50
51
52

}
return x;

53

Figure A.4. Gaussian elimination in C++. (a) Header le. (b) C++ code.

Tin+1

Ti +

t
T n+1
(x)2 i+1

di

i = m 1, m 2, , 1

(A.23)

Figure (A.5) presents a Fortran implementation of the Thomas algorithm.

A.2.2 LU Decomposition
LU decomposition functions improves upon Gaussian elimination process by factoring the coefcient matrix into a form that can be reused with multiple RHS b vectors.
The operation count for solving the resulting factored system is lower than the count
for Gaussian elimination. This approach proves useful when a linear system must be
solved at each time step of a differential equation solver and the coefcient matrix
stays constant over time. Such is the case for equations (A.57) and (A.58).
LU decomposition factors any nonsingular coefcient matrix A (that is one with
nonzero determinant) into the product of a lower triangular matrix L and an upper
triangular one U. When no pivoting is required, the factorization takes the simplest

A.3 Nonlinear Solvers

form: A = LU. With pivoting, LU decomposition takes the form PA = LU, where P
is a permutation matrix. Permutation matrices have one unit entry in each row and
each column with zeros elsewhere. The multiplication of PA produces a new matrix
by interchanging the rows in matrix A so that PA has an LU decomposition.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28

Figure A.5
function thomasTimes(lhs,rhs)
real(rkind) ,dimension(:,:) ,allocatable ,intent(in) :: lhs
class(field)
,intent(in) :: rhs
type(field)
,allocatable :: tomasTimes
real(rkind) ,dimension(:)
,allocatable :: x,b ! Linear system:
real(rkind) ,dimension(:,:) ,allocatable :: A
! Ax = b
real(rkind)
:: factor
integer(ikind)
:: row,n
n=size(lhs,1)
b = rhs%node
! Copy rhs side to preserve required intent
if ( n /= size(lhs,2) .or. n /= size(b)) &
stop thomasTimes: ill-posed system
allocate(x(n))
A = lhs
! Copy lhs side to preserve required intent
!______ Establish upper triangular matrix _______________________
do row=2,n
factor=A(row,row-1)/A(row-1,row-1)
A(row,row) = A(row,row) - factor*A(row-1,row)
b(row) = b(row) - factor*b(row-1)
end do
!______ Back supstitution _______________________________________
x(n) = b(n)/A(n,n) ! Back substitution
do row=n-1,1,-1
x(row) = (b(row) - A(row,row+1)*x(row+1))/A(row,row)
end do
allocate(thomasTimes)
thomasTimes%node = x
end function
Figure A.5. Thomas algorithm procedure.

The following steps summarize the LU decomposition method with pivoting:

1. Obtain the permutation matrix P, lower triangular matrix L, and upper triangular
matrix U such that A = PLU.
2. Calculate Pb, which interchanges the rows of vector b and replaces the RHS
of (A.8) so PAx = Pb.
3. Solve Ly = Pb for y using forward substitution where y = Ux.
4. Solve Ux = y for solution x using back substitution.

A.3 Nonlinear Solvers

A.3.1 Newtons Method in 1D
The Newtons method, also known as the Newton-Raphson method, approximates
the root of f (x) = 0. Expanding a 1D, real function f (x), in a Taylor series about a

343

344

Mathematical Background
f (x)

Xn

Xn+1

Figure A.6. One iteration of Newtons method (dashed line is the tangent line of f [x] at xn )

reference point xn :
f (xn + x) = f (xn ) + f  (xn )x +

(A.24)

where x = xn+1 xn and f  is the derivative of f with respect to x, Newtons method

proceeds from one approximation, xn , to the next, xn+1 , by truncating the series at
the second term and setting f (xn+1 ) = f (xn + x) = 0, where:
x = xn+1 xn =

f (xn )
.
f  (xn )

(A.25)

Figure A.6 illustrates this process. One nds the root of f (x) by repeating this process until |x| is sufciently small. Starting sufciently close to the desired root
of a smooth function, Newtons method converges faster than some competing
methods and it can diverge for poor initial guesses or ill-behaved functions. One
pathological case occurs when the initial guess lies at a stationary point that is, a
point with a horizontal tangent line, in which case the iteration step (A.25) requires
division by 0.

A.3.2 Newtons Method in Multiple Dimensions

One can readily extend the 1D Newtons method to multiple dimensions. Given
m nonlinear equations f(x) = 0, where f and x have components fi and xi for i =
1, 2, , m, one represents successive approximations to the roots as:
xn+1 = xn + J1 (xn )f(xn )

(A.26)

where the Jacobian matrix J generalizes the 1D derivative f  . The ij th element of J is:
Jij =

fi (x)
i = 1, 2, , m j = 1, 2, , m.
xj

(A.27)

A.4 Partial Differential Equations

An approximate solution to the nonlinear equations obtains from iterating on

equation (A.26), starting from an appropriate initial guess. Much like the pathological stationary point in the 1D case, the multidimensional Newton method diverges
with singular or nearly singular Jacobian matrices.

A.4 Partial Differential Equations

A.4.1 The Heat Equation
Throughout Part I of this text, we discussed heat conduction in a solid within which
energy conservation arguments applied throughout a simply connected volume V
bounded by a closed surface S, neglecting other energy forms, leads to:



edV = q dS
(A.28)
t V
S
where , e, and q are the material density, thermal energy per unit mass, and surface
heat ux, respectively, and where the LHS and RHS integrals are volume and surface
integrals, respectively. The LHS of equation (A.28) represents the total thermal
energy in the volume, whereas the RHS represents the net heat transfer into the
volume at the boundary.
One can combine the two sides of equation (A.28) into a single term by applying
Gausss divergence theorem to convert the RHS of equation (A.28) to a volume
integral so that:


( [e) + q] dV = 0
(A.29)
t
V
Because there is nothing special about the volume V, equation (A.29) holds
for any arbitrary volume, including one shrunk to within an innitesimally small
neighborhood of a point. This forces the integrand to vanish at all points:

(e) + q = 0
t

(A.30)

Writing the energy, e cT, in terms of the specic heat c and temperature T, and writing Fouriers law of heat conduction q kT, where k is the thermal conductivity,
yields:
T
= 2 T
(A.31)
t
where k/(c) is the thermal diffusivity. Equation (A.31) is the heat equation.
We now present the analytical solutions for the 1D heat equation problem solved
numerically throughout Part I of this text. In equation 1.1, the heat equation is
subjected to the initial condition T(x, 0) = Tair and the boundary condition T(0, t) =
Tchip and T(Ln , t) = Tair . If T(0, t) = T(Ln , t) = 0, the heat equation can be solved
by the method of separation variables and yields the general solution:
T(x, t) =

n=


Bn sin

n=0

where:
2
Bn =
Ln

n n2 2 t/L2n
xe
Ln

Ln

Tair sin
0

n
xdx
Ln

(A.32)

(A.33)

345

346

Mathematical Background

In order to solve equation 1.1 subjected to nonzero boundary condition, we assume

T = T1 + T2 , where T1 satises the nal steady state, that is, the point in time after
which the temperature becomes independent of time. Therefore, one can determine
T1 from:
2 T1
=0
(A.34)
x2
subject to the boundary conditions:
T1 (0) = Tchip T1 (Ln ) = Tair

(A.35)

The solution is T1 = A + Bx where A = Tchip and B = (Tair Tchip )/Ln . The solution
T1 does not satisfy the initial condition. In order to have T(x, 0) = T1 (x) + T2 (x, 0) =
Tair , T2 must satisfy the initial condition:
T2 (x, 0) = Tair Tchip

Tair Tchip
x
Ln

(A.36)

T2 also must satisfy the heat equation and the boundary conditions:
T2 (0, t) = 0, T2 (Ln , t) = 0

(A.37)

According to equation A.32, the solution of T2 will be:

T2 (x, t) =

n=


Bn sin

n=0

where:
2
Bn =
Ln

Ln 

n n2 2 t/L2n
xe
Ln


Tair Tchip
n
Tair Tchip
x sin
xdx
Ln
Ln

(A.38)

(A.39)

Thus, the analytical solution of the temperature in equation 1.1 is T(x, t) = T1 (x) +
T2 (x, t).
A.4.2 The Burgers Equation
Of much younger vintage, the Burgers equation at rst appears more daunting due
to its nonlinearity:
(A.40)
ut + uux = uxx .
Nonetheless, as mentioned in Section 4.3.3, the Burgers equation can be transformed
into the heat equation and solved exactly for the case of periodic boundary conditions
employed in the text. Section 4.3.3 provides the resulting analytical solution. Here we
briey sketch the analytical solution process. The initial condition is u(x, 0) = A sin(x)
and the boundary condition is u(x + n , t) = u(x, t) = 0 for all integers n. As described
in Section 4.3.3, the transformed Burgers equation is:
t = xx .

(A.41)

where u = 2x /. Integrating the initial condition (x, 0) yields:

1

(x, 0) = e 2

A sin(x)dx

= e 2 A cos(x)

(A.42)

A.5 Numerical Analysis

347

Using the method of separation of variables, we obtain the general solution of

Eq A.41:

cn en t cos(nx)
2

(A.43)

n=

where:
cn =

1
2

1
2

(x, 0) cos(nx)dx
1

e 2 A cos x cos(nx)dx

(A.44)

The above integration can be expressed by Bessel function In as (Benton and Platzman 1972):
cn = (1)n In (

A
)
2

(A.45)

Hence, the solution is:

u=

n=

n2 t sin(nx)
n=1 ncn e

n2 t cos(nx)
c0 + 2 n=
n=1 cn e

(A.46)

A.5 Numerical Analysis

A.5.1 Finite Differences
All of the numerical PDE solvers in this book employ the semidiscrete approach,
wherein one rst discretizes the spatial domain by replacing the the spatial derivatives with difference approximations. This results in a set of ODEs with continuous
dependence on time. One then discretizes the time domain by choosing a numerical quadrature that marches the equations forward in time over a sequence of nite
steps.
In order to represent the spatial derivatives using a nite difference approximation, one overlays a grid, or mesh, on the continuous problem domain. Figure (A.7)
shows a typical nite-difference stencil: The layout of the grid points employed in
approximating x- and y-direction derivatives at a point (x0 , y0 ). One can dene a
nite-difference approximation to an x-direction derivative of a function f (x, y) at a
point (x0 , y0 ) as:
f
f (x0 + x, y0 ) f (x0 , y0 )

x
x

(A.47)

When x is sufciently small, (A.47) approximates f /x closely. A Taylor-series

expansion of f about (x0 , y0 ) motivates equation (A.47):


f 
2 f 
f (x0 + x, y0 ) = f (x0 , y0 ) +
x + 2 
x (x0 ,y0 )
x 

(x0 ,y0 )

x2
+
2

(A.48)

348

Mathematical Background
(x0, y0+y)

(x0x, y0)

(x0+x,y0)

(x0, y0)

(x0, y0y)

Figure A.7. A nite-difference stencil at the point (x0 , y0 ).

The nite-difference approximation to f /x derives from rearranging equation (A.48) and combining it with equation (A.47) so that:


f
2 f 
f 
+
=

x
x (x0 ,y0 ) x2 

(x0 ,y0 )

x
+
2

(A.49)

where the second and subsequent RHS terms represent the truncation error, which is
the difference between the f /x and f /x. The truncation error behaves asymptotically like its leading term, which is usually specied with the order of (O) notation,
so the truncation error behaves like:
e


f
f 

= O(x)

x (x0 ,y0 ) x

(A.50)

Rening the mesh by shrinking x decreases the truncation error proportionately

with the rst power x. Equation (A.47) is therefore referred to as a rst-order,
forward difference.
There are an innite number of nite-difference approximations to f /x with
different orders of truncation error. Jumping backward to (x0 x, y0 ) and forming
the Taylor series expansion about (x0 , y0 ) yields:


f 
2 f 
x + 2 
f (x0 x, y0 ) = f (x0 , y0 )
x (x0 ,y0 )
x 

(x0 ,y0 )

x2
+
2

(A.51)

A.5 Numerical Analysis

from which we can extract the rst-order, backward difference formula:


f (x0 , y0 ) f (x0 x, y0 )
f 


x (x0 ,y0 )
x

f 
=
+ O(x)
x (x0 ,y0 )

(A.52)
(A.53)

The leading-order truncation error for this approximation equals that of the forward
difference formula.
Higher-order nite-difference approximations result from combining Taylor series expansions of f at different points. For example, one derives the second-order,
central difference formula by subtracting equation (A.51) from equation (A.48) and
rearranging so that:

f (x0 + x, y0 ) f (x0 x, y0 )
f 

(A.54)

x (x0 ,y0 )
x
which has a leading-order error of order O[(x)2 ]. We leave the derivation of the
precise form of this error as an exercise for the reader. Other combinations of Taylor
series produce approximations to higher-order derivatives. Adding equation (A.51)
to equation (A.48) gives a second-order-accurate approximation to the second partial
derivative of f with respect to x:

2 f 
f (x0 + x, y0 ) 2f (x0 , y0 ) + f (x0 x, y0 )

(A.55)

2
x 
(x)2
(x0 ,y0 )

which has a leading-order error of order O[(x)2 ].

One arrives at corresponding nite-difference estimates of derivatives with respect to other variables, such as y, similarly. For example, the rst-order, forward
nite-difference approximation to f /y is:

f 
f (x0 , y0 + y) f (x0 , y0 )
(A.56)

y (x0 ,y0 )
y
which has a leading-order error of order O(y), and likewise in the z direction.
Higher-order nite difference methods result from canceling additional error
terms. To avoid broadening the stencil, compact schemes can be constructed with
special care. These generally express the derivative approximation implicitly, necessitating the solution of a linear system of equations. The solution to this system yields
the derivative approximation at all points in the domain simultaneously. The classical Pad schemes take this form (Moin 2001) for the fourth-order rst derivative:



f 
f 
f 
3
+
+4
=
(A.57)
(fx +x fxi x )



x (xi +x) x (xi x)
x xi x i
and fourth-order second derivative:



1 2 f 
10 2 f 
1 2 f 
+

 +

12 x2 
12 x2 
12 x2 
(xi x)

xi

(xi +x)

fxi +x 2fxi + fxi x

x2

both of which yield linear systems with tridiagonal coefcient matrices.

(A.58)

349

350

Mathematical Background

A.5.2 Numerical Methods for Differential Equations

As indicated in Chapter 4, the governing equations in most multiphysics simulations
can be described as:

 x, t)), x , t (0, T]
 U(
U(x, t) = (
t

(A.59)

with boundary and initial conditions:

 U)
 = C(
 x, t), x 
B(

(A.60)

 x, 0) = U0 (x), x 
U(

(A.61)

 {U1 , U2 , ..., Un }T is the problem state vector; x and t are coordinates in the
where U
 {1 , 2 , ..., n }T is a vector-valued operator
space-time domain  (0, T]; and 
that couples the state vector components via a set of governing ordinary-, partial-, or
 U)
 typically
integro-differential equations. The space domain  is bounded by , B(
 and its derivatives, and C
 species
represents linear or nonlinear combinations of U
the values of those combinations on .
Numerical solution of (A.59) on a digital computer requires discretizing space
and time. Adopting the aforementioned semidiscrete approach and applying the
spatial nite difference methods of the previous section results in a system of ODEs
of the form:
d 
 V)

V = R(
(A.62)
dt
 contains the discretized values of U
 on the grid and the vector function
where V

R represents the discrete approximation to the RHS of (A.59) derived via nite
differences or another suitable approximation formalism.
Considering the following single, rst-order ODE simplies the presentation of
approaches to solving (A.62):
dy
= f (y)
(A.63)
dt
The numerical methods discussed next estimate the solution yn+1 at time tn+1 =
tn + t based on a know solution at tn .

A.5.2.1 Euler Method

One approach to deriving numerical time advancement schemes involves expanding
the solution y in a Taylor-series forward in time from tn to tn+1 :



d3 y 
t 2
t 3
dy 
d2 y 
+ 3
+
(A.64)
yn+1 = yn +
t + 2 

dt t=tn
2
6
dt 
dt 
t=tn

t=tn

According to equation (A.63), dy/dt|tn = f (yn ). Using this information and truncating equation (A.64) after two RHS terms produces the forward Euler method:
yn+1 = yn + tf (yn ),

(A.65)

A.5 Numerical Analysis

which is also known as the explicit Euler method because the unknown quantity yn+1
appears only on the LHS.
Whereas studying the truncation error addresses the asymptotic behavior of a
numerical method for innitesimal t, ensuring accurate solutions in practice requires also considering the methods behavior at nite t. In the most pathological
cases, errors that stay under control at sufciently small t might explode at larger
t. Ideally, numerical time-integration algorithms would remain stable whenever
the exact solution is stable, and unstable only when the exact solution is unstable.
One gains some understanding of a time advancement schemes stability
properties by studying its behavior for a simple, model problem:
dy
= y,
dt

(A.66)

where R + iI and where R and I are the real and imaginary parts of . Since
equation (A.66) admits an exponential solution, a sufcient condition for a stable
exact solution is R 0. This bound thus holds also for the problems of interest in
studying the stability of time advancement schemes.
Applying explicit Euler method (A.65) to the model problem (A.66) yields:
yn+1 = yn + tyn

(A.67)

Thus, the solution at time tn can be written as:

yn = y0 (1 + t)n = y0 (1 + R t + iI t)n = y0 n

(A.68)

where = (1 + R t + iI t) is called the amplication factor. The Euler explicit

scheme will be stable if the solution yn is bounded when n becomes large. The
necessary condition for Euler explicit method to be stable is | | 1, so that:
|1 + R t + iI t| = (1 + R t)2 + 2I (t)2 1

(A.69)

Therefore, the Euler explicit method is conditionally stable in the sense that the time
step must be small enough to satisfy (A.69).
The expression just to the left of the inequality sign in equation (A.69) is a
polynomial in t. All explicit time-advancement schemes generate stability criteria
that place polynomial bounds on the time step. As such, explicit schemes suffer
from the same problem that aficts high-order polynomial interpolation schemes
as mentioned in Section A.1.1: All polynomials diverge at innity. This precludes
unconditional stability because any polynomial will violate the inequality in (A.69)
outside some nite region of the complex- plane. As we demonstrate next, implicit
methods circumvent this restriction by generating stability bounds based on rational
polynomials that is, ratios of polynomials. Some implicit algorithms even offer
unconditional stability: They are stable for all problems for which the exact solution
is stable.
The implicit Euler scheme derives from expanding y in a Taylor series about the
time tn+1 :



t 2
t 3
d3 y 
dy 
d2 y 
yn+1 = yn
3
+
(A.70)
t + 2 

dt t=tn+1
2
6
dt 
dt 
t=tn+1

t=tn+1

351

352

Mathematical Background

Implicit Euler truncates the series after two terms:

yn+1 = yn + tf (yn+1 ),

(A.71)

which is also called the backward Euler method. The additional stability afforded
by implicit schemes comes at the cost of requiring iterative solutions when f (yn+1 ) is
nonlinear. This additional complexity generally manifests in increased solution costs
relative to explicit methods.
Applying the implicit Euler scheme to the model problem (A.66) leads to:
yn+1 = yn + tyn+1 ,

(A.72)

which admits a solution at time tn given by:

yn = n y0 ,

(A.73)

where the amplication factor is the rational polynomial = 1/(1t). The method
is stable if, and only if, | | 1, which always holds when the exact solution is stable
(R 0). Therefore, the implicit Euler method is unconditionally stable.
A.5.2.2 Trapezoidal Method
An alternative strategy for developing time-advancement algorithms integrates
equation (A.63) over the interval [yn , yn+1 ] to obtain:
 tn+1
yn+1 = yn +
f (y) dt
(A.74)
tn

A large family of schemes derive from applying various Lagrange polynomial

approximations to f (y) in A.74 using the f values at the m + 2 time steps
{tnm , tnm+1 , . . . , tn , tn+1 } as the nodal data:
n+1


f [y(t)]

fi i (t),

(A.75)

i=nm

where the Lagrange functions, i , are as dened in Section A.1.1. Substituting

equation (A.75) into the integral in equation (A.74) yields:
n+1


yn+1 = yn +

ci

ci fi ,

(A.76)

i=nm
tn+1
tn

i (t) dt.

(A.77)

In the degenerate case of a single quadrature point, the only available interpolating
polynomial is a constant. Choosing the point tn , for example, generates the polynomial n = 1, which produces the explicit Euler formula (A.65). Choosing instead
tn+1 generates the polynomial n+1 = 1 and the implicit Euler formula (A.65). This
general machinery outputs a time-integration algorithm in the form of coefcient
values, {ci |i = n m, ..., n + 1}, for each choice of quadrature points and resulting
interpolating polynomials.

A.5 Numerical Analysis

n +1

0
tn1

tn2

tn

tn+1

Figure A.8. Linear Lagrange polynomials for the case in which the f values at tn and tn+1 are
used to derive a numerical quadrature scheme.

Figure A.8 displays the case in which the two quadrature points are the previous
time step, tn , and the next one, tn+1 . With these points, the interpolating polynomials
are the two lines shown: n (t) and n+1 (t). In this simple case, one can evaluate the
integral (A.77) by graphical inspection:
yn+1 = yn +

t
[f (yn ) + f (yn+1 )],
2

(A.78)

which is called the trapezoidal rule. The trapezoidal rule provides the timeintegration mechanism in the Crank-Nicolson method, which couples trapezoidal time integration with central difference spatial approximation for the heat
equation (1.1). In the 1D case, the Crank-Nicolson method is thus:
Tjn+1

= Tjn +

t
2

n+1
n+1
2Tjn+1 + Tj1
Tj+1

(x)2

n 2T n + T n
Tj+1
j
j1

(x)2

"
(A.79)

where subscripts indicate location on a uniformly distributed 1D spatial grid and

superscripts indicate time step. The Crank-Nicolson method is an unconditionally
stable, implicit method.
Additional choices of quadrature points lead to multistep methods. When the
point at tn+1 is included, equation (A.75) generates implicit schemes called AdamsMoulton methods. When the point at tn+1 is dropped, equation (A.75) generates
explicit schemes called Adams-Bashforth methods.

A.5.2.3 Runge-Kutta Method

A third approach to deriving time-integration algorithms employs intermediate evaluations of f at points between tn and tn+1 . These schemes are generally called
Runge-Kutta methods. The approximate solution at tn can then be calculated at once
as a weighted sum of these f values. Because an equivalent estimate can also be
constructed via intermediate predictions for y between tn and tn+1 and successive
corrections to ultimately arrive at the approximation at tn+1 , these schemes are also
called predictor-corrector methods.

353

354

Mathematical Background

We illustrate this sections approach with the second-order, Runge-Kutta (RK2)

method:
yn+1 = yn + 1 k1 + 2 k2

(A.80)

where k1 = tf (yn ) and k2 = tf (yn + k1 ). The values of 1 , 2 , and determine
the methods truncation error. A common trait among Runge-Kutta methods is
that they truncate the Taylor series expansion of the solution precisely at the term
corresponding to the order of accuracy of the method. In other words, the leadingorder error term precisely equals the rst omitted term in the Taylor expansion of the
exact solution. According to the Taylor series expansion given by equation (A.64),
second-order accuracy requires retaining three terms:

(t)2 df 
f (yn )
(A.81)
yn+1 = yn + tf (yn ) +
2 dy t=tn
Comparing the estimated solution in the RK2 formula (A.80) to the truncated Taylor
expansion (A.81) determines the the RK2 constants. To facilitate a direct comparison, one substitutes the Taylor series expansion of k2 about yn equation (A.80). This
leads to:

df 
yn+1 = yn + (1 + 2 )tf (yn ) + 2 (t)2
f (yn )
(A.82)
dy t=tn
Comparing equation (A.82) to equation (A.81) and matching the coefcients of the
similar terms gives:
1 + 2 = 1
1
2 =
2

(A.83)

Therefore, 1 = 1 1/(2), and we have arrived at a one-parameter family of RK2

methods, each member of which corresponds to a different value for . Thus:
k1 = tf (yn ),
k2 = tf (yn + k1 ),


1
1
yn+1 = yn + 1
k1 +
k2 .
2
2

(A.84)
(A.85)
(A.86)

A predictor-corrector form of RK2 reads:

t
f (yn ),
2
yn+1 = yn + tf (yn+1/2 ),

yn+1/2 = yn +

(A.87)
(A.88)

corresponding to = 0.5.
A.5.2.4 Semi-implicit Methods
To take advantage of the stabilizing effect of implicit schemes in the presence of
diffusive terms that impose severe time step restrictions while avoiding iterative solution of nonlinear equations, it is attractive to advance linear and nonlinear terms

A.5 Numerical Analysis

in different manners. Spalart et al. (1991) published a semi-implicit algorithm that

exemplies this approach. Their method advances linear terms implicitly and nonlinear and inhomogeneous terms explicitly. For a solution vector U and differential
equation system:
Ut = L(U) + N(U) R(u)

(A.89)

where L contains all linear terms and N contains all nonlinear and inhomogeneous
terms, the Spalart et al. algorithm takes the form:
U = Un + tn+1 {L[1 U + 1 Un ] + 1 N(Un )}




(A.90)


U = U + tn+1 {L[2 U + 2 U ] + 2 N(Un ) + 1 N(U )}

(A.91)

Un+1 = U + tn+1 {L[3 Un+1 + 3 U ] + 3 N(U ) + 2 N(U )}

(A.92)

where primes denote substep values between the nth and (n+1)th time steps, and
tn+1 = tn+1 tn . For N, this approach resembles an explicit Euler substep followed
by two second-order Adams-Bashforth substeps N1 . For L, this approach resembles
trapezoidal integration.
Third-order accuracy can be achieved by matching the Taylor series:
t 2  Ri
Uj
2
Uj
3

Ui (t + t) = Ui + tRi (U) +

(A.93)

j=1

t 3   2 Ri
3
Uj Uk + O(t 4 )
6
Uk Uj
3

=+

(A.94)

j=1 k=1

This condition generates 11 unknown coefcients that must satisfy 17 equations.

By further requiring that the length of the substeps be the same for L as for N,
Spalart, Moser, and Rogers (1991) reduce this system to eight equations in eight unknowns. Unfortunately, they also reported that the system appears to be insoluble.
They therefore sacriced one constraint, ultimately achieving third-order accuracy
for N and second-order accuracy for L. In subsequent work, they have typically used
the values:
{1 , 2 , 3 } {4/15, 1/15, 1/6}

(A.95)

{1 , 2 , 3 } {4/15, 1/15, 1/6}

(A.96)

{1 , 2 , 3 } {8/15, 5/12, 3/4}

(A.97)

{1 , 2 } {17/60, 5/12}

(A.98)

When we use their scheme in this book, we use these values.

A.5.2.5 Stability of PDE solvers

Multiphysics simulations typically involve coupled, multidimensional, nonlinear
PDEs in complicated geometries with complex boundary conditions. Proving the
1 Adams-Bashforth schemes are multistep schemes in that they require time values from multiple

previous steps in order to move forward. See Moin (2001).

355

356

Mathematical Background

stability of numerical approximations in such contexts requires a great degree

of mathematical sophistication and is often best approached by starting from a
discretization strategy that supplies a rich mathematical formalism. Most such discretizations fall under the heading of weighted residual methods (WRM), including
spectral and nite element methods and hybridizations thereof (Canuto, Hussaini,
Quarteroni, and Zang 2006; 2007; Karniadakis and Sherwin 2005). In the interest of
simplicity of presentation, we do not pursue such strategies in this text.
Both the nite difference approximations employed in this book and the WRM
involve laying a grid over the spatial domain and marching the solution through
time in discrete steps. The grid spacing, x, characterizes the spatial discretization,
whereas the step size, t, characterizes the temporal discretization. In multiphysics
modeling specically applications that couple uid dynamics to other physical phenomena stability arguments can often be couched in terms of limitations on the
speed with which information propagates. Information propagation occurs either by
organized motion of bulk matter (ow) or by random, disorganized motion of individual molecules (diffusion). The ow speed u characterizes the ow. A diffusion
coefcient say, the thermal diffusivity, , in the case of thermal energy transport
typically characterizes diffusion.
Combining x, t, u, and leads to two dimensionless parameters that generally
determine the stability of PDE discretizations:
CFL ut/x
t/(x)2

(A.99)
(A.100)

The rst of these is the Courant-Freidrichs-Lewy number and characterizes the

stability of explicit time advancement of ow equations. The second parameter characterizes the stability of explicit time advancement of diffusion equations. The stability
criteria take the form of inequalities that restrict the values of these parameters to
remain below values of order unity. Although the derivation of specic stability
bounds is beyond the scope of this discussion, such bounds are enforced in most of
the solvers throughout this book.
Since varies inversely with the square of the grid spacing, the stability bounds
on tend to be more restrictive than the limits on the CFL number. Implicit time
advancement circumvents such restrictions. Because most multiphysics simulations
involve nonlinear processes, however, fully implicit time advancement requires
costly iterations. It is common therefore to use semi-implicit time advancement,
marching the nonlinear terms forward explicitly to avoid iteration, while simultaneously marching the implicit terms forward implicitly to relax the time step restrictions.
The multiphysics simulations discussed in Sectioin 12.2.1 all employ the semi-implicit
time advancement algorithm of Spalart, Moser, and Rogers (1991).

APPENDIX B

Unied Modeling Language Elements

This appendix summarizes the Unied Modeling Language (UML) diagrammatic

notation employed throughout this book along with the associated terminology and
brief denitions of each term. We consider the elements that appear in the ve types
of UML diagrams used in the body of the current text: use case, class, object, package,
and sequence diagrams. At the end of the appendix, we give a brief discussion of
Object Constraint Language (OCL), a declarative language for describing rules for
UML models.

B.1 Use Case Diagrams

A use case is a description of a systems behavior as it responds to an outside request
or input. It captures at a high level who does what for the system being modeled. Use
cases describe behavior, focusing on the roles of each element in the system rather
than on how each element does its works.
A use case diagram models relationships between use cases and external
requests, thus rendering a visual overview of system functionality. Figure B.1
reexamines the n heat conductor analyzer diagram from Figure 2.6, adding notations
to identify the elements of the use case diagram.
Use case diagrams commonly contain the following elements:
1. Actors: people or external systems that interact with the system being modeled.
Actors live outside the system and are the users of the system. Typically actors
interact with the system through use cases. In UML, actors are drawn as stick
gures. In the n analyzer system example, system architect, thermal analyst,
and numerical analyst are actors.
2. Use cases: abstractions that represent sets of sequences of actions that the system performs to produce measurable results to actors. A use case describes the
actions the system takes in response to a request from an actor. It captures the
essence of interactions between an actor and the system. In UML, a use case is
drawn as an ellipse, with its name printed in the center of the ellipse. The n
analyzer example includes three use cases: specify problem, design numerical
method, and predict heat conduction. All three use cases have actors associated
with them.
357

358

Unied Modeling Language Elements

fin analyzer

association

specify
problem
actor
system
architect

design
numerical
method
numerical analyst

predict heat
conduction
system boundary

thermal analyst
use case

Figure B.1. Use case diagram for a n heat conduction analysis system.

3. Associations: connections between use cases and actors. An association is established when an actor interacts with the system through a use case that describes
the interaction. An association is drawn using a solid line connecting an actor and a use case. Optionally an arrow can be added to indicate the direction
of interaction, that is, who initiates the interaction. In the n analyzer system, ve associations are identied with the following pairs of actor and use
case: system architectspecify problem, numerical analystspecify problem, numerical analystdesign numerical method, thermal analystdesign
numerical method, and thermal analystpredict thermal conduction.
4. The system boundary: shown as a rectangular box to indicate the scope of the
system. All the use cases are inside the system boundary, and all actors are
outside the system boundary.
5. Other relationships: relationships between actors and between use cases. Actor
generalization and use case relationships, such as include, extend, and
generalization, are also common in practice (Booch et al. 1999).
A use case diagram can also include packages to group different modeling elements
into chunks. The UML package diagram provides a convenient way to depict very
general collections of model elements. The elements need not be object-oriented. For
example, Figure 6.1 depicts a package containing a class and a stand-alone procedure
that accepts an instance of that class as an argument. Because package diagrams see
minimal use in the current text, appearing only in Figure 6.1, we go into no greater
depth here.
In our n analyzer example, the interaction drawn between the system architect
and the n analyzer system depicts use of the system to specify the heat conduction
problem. This usage is depicted by specify problem use case. The output from this
problem specication including the production of a set of physical parameters such
as tolerances, boundary conditions, initial heat conditions, and so on becomes the

B.2 Class Diagrams

input for another actor, the numerical analyst. Note the numerical analyst has two
use case associations with the n analyer. First, she receives data from the n analyzer
through the specify problem use case. This is depicted by the numerical analyst
specify problem association. Then she develops a numeric model based on these data
from the n analyzer. The latter interaction is depicted by the numerical analyst
design numerical method association. Similarly, the thermal analyst receives the
numeric model produced by the numerical analyst and uses this model to predict the
n heat transfer performance.
Use cases diagrams produce a static view of a system. Applications of use case
diagrams in modeling typically take two forms (Booch et al. 1999):
1. Actor-centric: model the context of a system. This modeling method places an
emphasis on actors and their roles, focusing on the environment surrounding
the system. It starts with identifying actor and organizing actors according to
their roles, followed by pinpointing use cases and their associations with actors
based on the actors behaviors and requirements on the system. Once the use
case diagram is completed, all use cases are associated with one or more actors.
Our n analyzer example follows this approach.
2. System behavior-centric: model the requirements of a system. Much like specifying a contract on the system functionalities, this approach focuses on the system
behaviors. It begins with identifying actors and gathering their requirements to
the system. Use cases are developed based on these requirements. Additional
use cases may be developed based on additional system requirements, variants
to existing use cases, and so on. (Booch et al. 1999) has provided an example
of validation system for credit card transactions to illustrate this approach. A
detect card fraud use case is an addition to the use case diagram due to security requirements. Similarly, they introduced a manage network outage use
case to the diagram due to a requirement that the system provide reliable and
continuous operations.
In each case, use case diagrams aid in visualizing and documenting the system
behavior. They can also aid in documenting the business requirements. Ideally, these
diagrams enhance designers, developers, and domain experts understanding of the
system before and during construction.

B.2 Class Diagrams

Class denitions comprise the most fundamental building blocks in any objectoriented application. One can model a systems classes using class diagrams, the most
common UML diagram. In particular, class diagrams can convey class attributes and
operations as well as the interactions and relationships among different classes. Class
diagrams provide a static, structural overview of a modeled system. The remainder
of this section covers the two key elements for class diagrams: representations of
classes and the relationships among classes.
B.2.1 Classes
A class is an abstraction for a set of objects that share the same attributes,
operations, and relationships to others. UML represents a class graphically using a

359

360

Unied Modeling Language Elements

Class name

visibility

Astronaut
attribute
greeting : String
<< constructor >>

stereotype

+ astronaut(String) : Astronaut
<< process >>
+ greet () : String

operations

Figure B.2. Class diagram for astronaut class.

three-compartment box. In Figure B.2, we redraw the Astronaut class from Figure 2.2
to illustrate the following elements of a class:
Name: The top compartment holds the class name, Astronaut (Figure 2.2). The

name is the only mandatory part of a class diagram. In diagrams that contain
large collections of classes (hundreds or thousands in some cases), it is common
to omit details such as attributes and operations so readers can focus on the
relationships between classes.
Attributes: As shown in Figure B.2, the middle compartment of a class is used to
show the class attributes. These are the Fortran-derived type components or the
C++ data members. The attributes are normally drawn in one of two ways:
1. By showing attribute names only.
2. By specifying attribute names followed by their data types, separating the
two by a colon. Additionally the diagram can provide default attribute
values.
Our example in Figure B.2 uses the second method. The Astronaut class contains
one attribute, greeting, of type String.
Operations: An operation is a service a class provides. Common Fortran
operations include type-bound procedures, dened operators, and dened
assignments. Common C++ operations include virtual member functions and
overloaded operators. As shown in Figure B.2, the bottom compartment of
a class box displays the operations. Similar to attributes, operations can be
specied with one of two methods:
1. By showing operation names only.
2. By specifying operation names accompanied with their function signatures,
including the procedure name and argument types followed by a colon and
the return data type, if any.
Our Astronaut example employs the latter method to reveal more information
about the operations. Two operations are shown in the Astronaut class: an overridden constructor astronaut(), which takes a string and returns a constructed
Astronaut object; and the procedure greet(), which produces String.

B.2 Class Diagrams

As with attributes, the list of operations of a class can be omitted from the
diagram if the emphasis of the design is placed on other aspects of the system
such as the interactions between classes.
Responsibilities: Optionally, one can detail a classs responsibilities in an additional compartment drawn below that of the operations. The responsibility part
of a class describes the contracts or obligations of the class (Booch et al. 1999).
In addition to these basic elements, there are some further notations commonly
employed to specify additional information on classes:
Visibility: Visibility refers to the accessibility of attributes or operations by other

classes. It falls into one of the following three states:

1. Public: Denoted by a plus sign (+), this state implies that code outside
the class can access the attribute or operation. In our Astronaut class, the
operations astronaut() and greet() are publicly accessible.
2. Private: Denoted by a minus sign (), this state indicates the attribute or
operation is not visible to code outside the class that contains the attribute
or operation. The attribute greeting in Astronaut class is private, preventing
direct access from outside the Astronaut class.
3. Protected: Denoted by a pound, or number, sign (#), an attribute or class
declared as protected is visible only to the class and its direct subclass. C++
supports the declaration of protected attributes and operations. Fortran does
not.1
Stereotypes: Graphically represented using names enclosed by guillemets (),
a stereotype in UML is one of extensibility mechanisms that allow derivation of
new and domain-specic modeling elements from existing ones. Our Astronaut
class exhibits one common usage of stereotypes in a class diagram: organizing
operations. constructor is routinely used in UML to denote that operations
immediately following the stereotype are overloaded constructors, and similarly
process marks the roles of the subsequent operations.
Abstract classes: UML distinguishes an abstract class from a concrete one by
putting the abstract class name in a bold italic font. The Integrand class in our
ABSTRACT CALCULUS pattern is an example of an abstract class (see Figure 6.1).
Interfaces: UML distinguishes an interface from a class by enclosing the interface
in a circle instead of a box. Interface diagrams are more relevant to languages
that have an explicit interface construct, for example, Java. In Fortran and C++,
one typically models interfaces as abstract classes. For clarity, we use the abstract
class notation in all class diagrams in this book.
Abstract operations: An important functionality of an abstract class is to specify
the services its subclasses must provide. Each specication usually takes the
form of an abstract operation that is, an operation signature dened without
concrete implementations. These are Fortran deferred procedures bindings or
C++ pure virtual member functions. An abstract operation is indicated by italic
font. In the abstract class Integrand, shown in Figure 6.1(b), operations add,
multiply and t are all abstract.

1 The protected keyword in Fortran denotes a module variable that is visible but not directly

modiable from outside a module.

361

362

Unied Modeling Language Elements

Template classes: A template class is a parameterized class. C++ template classes

and Fortran parameterized derived types are both template classes. In UML,
a template class is drawn as an ordinary class with an additional dashed box
in the upper-right corner of the class. This additional box shows the template
parameters used to instantiate an object of this class. As shown in Figure 2.3(a),
the Array class in Chapter 2 is a template class, with Element being the sole
template parameter.
Datatypes: As one of the classiers available in UML, a datatype represents
a set of data values, including intrinsic data types and enumerations. The graphical representation of a datatype is similar to a stereotype. Figure 2.3(b) shows
a common usage of datatypes in UML modeling: dening data types used in
template classes. In the example, data types of Integer, Character, Real, and
Boolean can be used as for the template parameter Element.

B.2.2 Relationships
In UML, a relationship is abstracted as a connection between two elements. In class
diagrams, all relationships are represented lines connecting classes. The following
three kinds of relationships are identied as the most important relationships in
OOD: dependency, generalization, and association.

B.2.2.1 Dependency
A dependency is a relationship between two classes wherein one class depends on
the behaviors of the other. Most often this indicates a use relationship between two
classes: We say class A depends on class B when B is used either as a passed parameter
or as a local variable for the operations of A. One draws dependencies in UML using
a dashed line connecting two classes with an arrow pointing to the class on which the
other class depends.
Derived from Figure 3.4, Figure B.3 emphasizes the class dependencies of the
heat conduction system given in Section 3.1. We only draw the dependencies used in

Integrable_conductor
<<constructor>>
integrable_conductor(Problem, Procedure)

Field
Differentiator

<<constructor>>
field(Problem, Procedure)

Figure B.3. Class dependencies in heat conduction class diagram.

Problem

B.2 Class Diagrams

the heat conduction system, omitting other relationships between classes. In the
system, both the Integrable_conductor class and Field class have a dependency
on the Problem class. The class denitions for Integrable_conductor and Field
also highlight the cause of their dependencies on Problem: The constructions of
Integrable_conductor and Field require the passing of a Problem object. From the
corresponding code in Figure 3.2 and Figure 3.3, we can further see that the Fields
construction is directly impacted by the implementation (or implementation changes)
of Problem because the Field derives the nodal values, the stencil, and the system boundary conditions from the Problem attributes, whereas the dependency on
Problem for constructing an Integrable_conductor object is pertinent to that of its
T_field component. In this sense, we can say Field has a stronger dependency on
Problem.

B.2.2.2 Generalization and Realization

In a class diagram, a generalization refers to a relationship between a base class
(the parent) and one that extends it (the child). The generalization relationship is
also known as inheritance or an is-a relationship. In a generalization, the parent represents the general properties of a class (i.e., attributes and operations),
whereas the child represents a specialized extension of the parent. In application
code, a child type is always substitutable for its parent type in procedure calls,
but not the reverse. In UML, a generalization is graphically rendered as a solid
line connecting the parent and the child, with an open arrowhead pointing to the
parent.
Closely related to a generalization, a realization in a class diagram is a
relationship between an interface and a class (an implementor) that realizes the
operations specied by the interface. Since abstract classes in C++ and Fortran
model interfaces, a realization relationship is established when a class extends an
abstract class. Consequently, a realization relationship automatically replaces a generalization relationship when the base class is abstract. The interface consistency
in a realization is reafrmed by the fact that both C++ and Fortran mandate the
implementor to keep the same procedure interfaces when implementing the inherited operations. In UML, a realization relationship is rendered as a dashed line
connecting the interface and its child, with an open arrowhead pointing to the
interface.
Figure B.4 re-illustrates the generalizations and realizations used in the class
diagram of a timed Lorenz system that employs SURROGATE and STRATEGY patterns.
Comparing to the original diagram in Figure 7.1, one may notice the omission of
other relationships between classes except generalizations and realizations.
In this system, classes Euler_Integrate and Runge_Kutta_Integrate are the
actual strategies used in time integration for Lorenz and Time_Lorenz, respectively.
Both Euler_Integrate and Runge_Kutta_Integrate are concrete implementations of
abstract class Strategy, thus both have a realization relationship with Strategy as
shown in the gure. For the class hierarchy tree started with Surrogate as the root,
the relationships between the parent/child pair can be said as follows: Integrand
realizes Surrogate, Lorenz realizes Integrand with concrete implementation, and
Timed_lorenz generalizes Lorenz.

363

364

Unied Modeling Language Elements

realization

surrogate

integrand

lorenz

strategy

euler_integrate

generalization
timed_lorenz

runge_kutta_integrate

Figure B.4. Class generalizations and realizations in SURROGATE and STRATEGY patterns for a
timed Lorenz system.

B.2.2.3 Association
An association represents a link or connection between two classes. Graphically,
associations add visual cues to the structural relationships between classes in a class
diagram. In UML, an association is drawn as a solid line connecting two classes.
Beyond the basic form, a few adornments supply additional information when
applied to association relationships:
Name and direction: An association name (usually a verb) is used to describe

the nature of the association. An additional direction can be supplied, as a

big solid triangle, to indicate the action ow. Figure B.5 simplies the class
diagram of Figure 9.1 with only association relationships depicted. As an
implementation of FieldFactory, the concrete class Periodic6thFactory constructs a Periodic6hOrder Field object through its implementation of create()
method. Exploiting the inheritance relationship between Periodic6thFactory
and FieldFactory, however, the method create() returns a reference to type Field.
Role: Specic roles can be given to the two participants in an association relationship. The role name is normally written as a plain text adjacent to the end
of an association line connected to the class.
Multiplicity: Multiplicity is a way to specify rules or restrictions on the number
of instances of a class in an association. It usually answers the question of how
many class instances are to participate in one association. The number can be
specied as exact (1), within a range (1..5), or many (1..*).
Figure B.6 highlights the association between HermeticField and HermeticFieldPointer used in Figure 10.7. The association relationship is a pointer to target
relationship, HermeticField being a target and HermeticFieldPointer being a
pointer. In the class diagram, the role for HermeticField is specied as a target; we omitted the specication for HermeticFieldPointer since its name is
self-explanatory. The diagram further species that for each (1) HermeticField
object, many (1..*) HermeticFieldPointer objects can be created and associated
with the same HermeticField object.

B.2 Class Diagrams

Field

FieldFactory

re
associations

tu

rn

Periodic6thOrder

Periodic6thFactory
constructs

direction

Figure B.5. Association relationships in

solver.

name
ABSTRACT FACTORY

patterns applied to a Burgers

multiplicity

HermeticFieldPointer

1 *

1
HermeticField
target

Role name

Figure B.6. HermeticField and HermeticFieldPointer class association and adornments.

Aggregation: Aggregation is a special kind of association that can be best de-

scribed as a whole-part, has-a, or contains relationship between two

classes. An aggregation relationship can involve more than two classes, with
one class (the whole) being a collection or a container of the rest of the classes
(the parts). In UML, an aggregation is depicted as an open diamond attached
to the class that behaves as the whole. The PUPPETEER pattern in Chapter 8
employs an aggregation relationship between the puppeteer and its three puppets: air, cloud, and ground. Figure B.7 reproduces the original class diagram as
shown in Figure 8.2, except for the added adornment labeling the relationship
as aggregation.
Composition: Composition is a variant of aggregation with added semantics.
Compared to an aggregation, a composition relationship identies a stronger
ownership between the whole and its parts for the association. In plain
language, a composition best describes a relationship in which one class (the
whole/composite) is composed of many others (the parts/components). Very
often the lifetime of the parts coincides with that of the whole in a composition
relationship. In comparison, an aggregation has a much looser ownership that
allows the parts to live independently from the whole. In addition, the whole is

365

366

Unied Modeling Language Elements

Puppeteer
aggregation

Ground

Air

Cloud

Figure B.7. In the PUPPETEER pattern, puppeteer is an aggregation of air, cloud, and ground.

also responsible for the disposition (e.g., creation and destruction) of its parts
in a composition2 . In UML, a composition is represented using a lled diamond
attached to the whole.

B.3 Object Diagrams

An object diagram is a graphical representation of instances of the classes found in a
class diagram, their states, and the relationships between them at a given time. Much
like its corresponding class diagram, an object diagram provides a static snapshot of
the system being modeled. However, a correlated set of object diagrams can provide
insights into how the system is expected to evolve over time.
In general, an object diagram is less abstract compared to its companion class
diagram. Parallel to the roles of classes and relationships in a class diagram, objects
and links are the key elements in an object diagram. Figure B.8 replots the object
diagram of Figure 2.12, showing that the n object of type Conductor is composed
of an object diff of type Differentiator. The diagram elements are as follows:
Objects: An object is an instance of a class. Whereas a class represents an

abstraction of similar things, an object gives a concrete instance of one. Similarly to a class, an object is also depicted using a box in UML. However, an object
can be readily distinguished from a class by the fact that the text contained in
the name compartment is underlined. Also the text is usually composed of two
components: an object name followed by a class name, separated by a colon (:).
Sometimes an object can be anonymous, that is, the object name before the
colon is omitted. This usually characterizes a temporary object returned from a
function call. In rare situations, one can also render objects without associated
class abstractions. These objects are called orphans, and readers can refer to
Booch et al. (1999) for more details.
In addition to the name compartment, an object can have a compartment
showing attributes and their values. Collectively they represent the particular
2 It can be confusing to decide whether an aggregation or a composition relationship should be used

in a design. The difference between them is subtle, and the design errors caused by making a wrong
choice are normally very minor.

B.4 Sequence Diagrams

class name

object name

fin:Conductor

object

diff

attribute

composition relationship
link

diff:Differentiator
diff_matrix = (/1.,2.,1./)
attribute values

Figure B.8. An object diagram that shows the object n of class Conductor is composed of an
object diff of class Differentiator.

state that the object is in. Figure B.8 contains two objects: the heat conductor n
and the differentiator diff. Their respective types are Conductor and Differentiator. Further we note that n contains a private component, named diff, whereas
n is in its initial state with its attribute diff_matrix being assigned to a specic
set of values.
Links: Analogous to an object being an instance of a class, a link is an instance
of an association. As with associations, links show relationships between objects
in an object diagram. In UML, a link is rendered using a line connecting two objects, just like an association in class diagrams. In addition, adornments used on
associations, that is, names, multiplicities, roles, and aggregations/compositions,
can also be applied to links. In Figure B.8, one can observe the fact that n is a
composition of diff.
Both the object diagrams and the class diagrams described in B.2 belong to the
structural modeling category. Both types of diagrams are commonly used to capture
the static view of a system. In the next section, we introduce sequence diagrams that
focus primarily on modeling the dynamic aspects of the system.

B.4 Sequence Diagrams

A sequence diagram is a kind of interaction diagram that depicts the interactions
between different objects and the ordering of these interactions. It is commonly
used to model the dynamic aspects of a system to show the details of an operation
involving a set of objects. Objects, messages (with data exchanges), and sequences
of actions are the key elements of an interaction diagram. A sequence diagram
emphasizes the time ordering among processes during an operation.

367

368

Unied Modeling Language Elements

Table B.1. Types of actions that are commonly used in an interaction
Type of action

description

call

Invokes an operation on an object. An object may send a message to itself,

resulting in the local invocation of an operation.
Provides a result to the caller.
Transmits a signal to an object.
Instantiates an object explicitly. Createevents are designated in UML
diagrams via stereotypes.
Eliminates an object explicitly. An object may initiate a self-destruction.
Destroy events are designated in UML diagrams via stereotypes.

return
send
create
destroy

B.4.1 Interactions and Messages

In UML, an interaction is comprised of a set of objects and a set of messages
exchanged among the objects. A message is a specication of a communication and
the information exchange between two objects. Occasionally an object can also send
a message to itself, for example an invocation of a recursive function call to itself.
In concrete terms, a message in an interaction can usually be mapped into an action
that involves only two objects. The receipt of a message is normally considered an
instance of an event during which an action is taken and a state change results as a
consequence. Also the data passed between the two objects are an inherent part of the
message. Even in a complex operation that involves multiple objects, it is always possible to decompose the whole process into sequences of actions, each only involving
two objects. These ultimate actions are drawn as messages in UML diagrams.
In UML, a message is depicted using a straight line between the two objects with
an arrow indicating the direction of invocation. In the case a self-call, that is, an object
sends a message to itself, a connecting line (either curved or with straight segments)
is used to show that the object is both the sender and the receiver. There are a few
kinds of particular actions that are commonly used, as described in Table B.1.
Taking again the heat conduction example used in Chapter 3, in particular the
forward Euler computation of fin + fin%t()*dt as shown in Figure 3.5(a), we
illustrate the use of UML interactions via a small section from that gure.
Figure B.9 details the call sequences during the computation of the second-order
spatial derivative on the T_eld component of n. As can be traced clearly by the
sequence numbers encoded in the messages, the object n initiates the computation
by sending a message xx() to T_eld, followed by the creation and initiation of a
temporary object by T_eld using the action create . Next a laplacian() call to
stencil is made by the temporary. The last action shown in the diagram is stencil
returning the laplacian back to the temporary eld object. For brevity, we omit the
obvious steps needed to return the computed value of xx() all the way back to n in
the diagram.

B.4.2 Sequence Diagrams

Sequence diagrams are amongst the most common interaction diagrams used in
UML. Communication diagrams are another popular kind of interaction diagram.

B.4 Sequence Diagrams

object

sequence number

message

1 : xx()

fin:integrable_conductor

T_field:field
action call

link
1:1 <<create>>

action create

1.2 : laplacian()

:field

stencil:differentiator

1:3: return

action return

Figure B.9. Interactions among objects during the computation of the 2nd-order spatial
derivative for forward Euler method.

Figure B.9 is a communication diagram, and it contains all the elements normally
found in a communication diagram. Semantically a communication diagram and a
sequence diagram are equivalent in that one can convert from one type of diagram
to the other without losing any information.
We choose to use a sequence diagram in our examples based on two factors.
First sequence diagrams are more explicit in revealing the time ordering among
different actions during an operation. In a complex operation, particularly when
many function calls are made between the same couple of objects at different times,
the sequences of calls can become cluttered in a communication diagram, but can
always be shown clearly in a sequence diagram. The second factor is that the sequence
diagram has analogues in other elds. For example, the Message Sequence Chart
(MSC) used in the telecommunication industry is nearly identical to a sequence
diagram.
We redraw the interactions of Figure B.9 using a sequence diagram in
Figure B.10. By comparing these two gures, one can observe that the vertical axis
in a sequence diagram denotes the time. The higher a message is drawn, the earlier
the event occurs.
Next, every object in the sequence diagram has an object lifeline: a dashed
vertical line showing the time period during which the object exists. Most objects
will exist during the whole operation. So they are aligned at the top of the diagram,
with their lifelines extended from the top to the bottom of the diagram. However,
objects can also be brought into life by dynamic creation create.
Last, every action is drawn with a focus of control, a tall thin rectangle, to indicate
the duration of the action. Usually the top of a rectangle is aligned with a message
that starts the action. The bottom of the rectangle indicates the completion of an
action, and may be aligned with another message, for example, a return message.
A focus of control may also provide hints as to the duration of the action: A short
focus of control usually implies a rapid completion of an action. Compared to the

369

Unied Modeling Language Elements

fin:integrable_conductor

T_field:field

stencil:differentiator

xx ()
<<create>>
time

370

:field
laplacian()

lifeline

focus of control

Figure B.10. Sequence diagram showing operations during the computation of the 2nd-order
spatial derivative for forward Euler method.

communication diagram, we can see that a sequence diagram places greater emphasis
on the time ordering of messages.

B.5 The Object Constraint Language

The Object Constraint Language (OCL), rst developed in 1995 during a business modeling project within IBM, is a language based on mathematical set theory
and predicate logic but is designed to be understood by people who are neither
mathematicians nor computer scientists (Warmer and Kleppe 2003). OCL alleviates
problems with incompleteness, informality, and imprecision that arise in models written in UML. In combination with UML, OCL expressions can make models more
precise and detailed. OCL appears at rst to be a constraint language, but in UML 2,
one can use OCL to write not only the constraints but also any information about the
elements in the UML diagrams, such as the initial values, derivation rule, or denition
of an instance. This makes OCL a query language as well. The current section focuses
on three main OCL expressions, or rules: those that add initial values, invariants,
and pre-/postconditions.

B.5.1 The Context of OCL Expression

An OCL expression augments a UML diagram with information about the model.
The link between the UML diagrams and the OCL expression is the context of OCL
expression, which species the element in the UML diagrams for which the OCL
expression is dened. The contents of the context are usually the class, interface,
datatype, component, or operation that are dened in the UML diagrams. When the
OCL expression is incorporated in the UML diagram, such as in the Figure 10.6, the
context denition is shown by a dotted line that links the element in a UML diagram
and the OCL expression. When the OCL expression is given in a separate text le,

B.5 The Object Constraint Language

the context is denoted by the keyword context followed by the name of the element
in UML diagram. OCL expressions depend on the context of the expression and can
have many different functions; for example, the expression can be an initial value
when the context is an attribute.

B.5.2 Initial Value Rule

Initial value rules add information regarding the initial values of attributes and
association ends in the UML diagram. The context of an initial value rule is the UML
diagram element that holds the relevant attributes or association ends. For example,
Figure 10.5 shows the HermeticField class has a temporary attribute for which the
default initial value is always true. In OCL, an initial value of temporary can be
dened as follows:
context : HermeticField :: temporary
init: true
B.5.3 Specifying Object Invariants
An OCL expression can be added to a UML diagram to specify an invariant that is
a constraint for an object and must be true for the object throughout its complete
life. The context of the invariants can be a class, an interface, or a type in the UML
diagram. Any attribute of the context can be used in invariant expressions. For
example, because the GuardedField object dened in Figure 10.6 must be guarded
against premature nalization, the guarded state can be formalized in the following
OCL expression:
context : GuardedField
inv: depth > 1
B.5.4 Adding Pre- and Postconditions to Operations
Preconditions and postconditions are boolean expressions that constrain operations.
A precondition of an operation must be true in order for an operation to execute. A
postcondition must be true after an operation terminates; otherwise, the operation
has failed. The pre- and postconditions only describe the applicability and the effect
of the operation without specifying the implementation.
The context of pre- and postconditions is the class name that holds the operation,
followed by a double colon with the operation name, parameters of the operation,
and the returned type. The operation name, parameters of the operation, and the
returned type are usually dened in the UML diagram. The OCL expression of preand postconditions follows the key words pre and post each return a boolean
value:
context: Field :: d_dx(direction : Integer) : Field
pre: self.fourier.allocated() xor self.physical.allocated
post: self.temporary implies
(not self.fourier.allocated()) and (not self.physical.allocated())

371

372

Unied Modeling Language Elements

Field
N : Integer
fourier : Array<complex>
physical : Array<real>
temporary : Boolean
<< constructor>>
+ field_(Array<real>) : Field
<<process>>
+ d_dx(direction : Integer) : Field
+ operator(+) (right : Field) : Field

Figure B.11. Field class diagram of the turbulence solver of Rouson et al. (2006).

where self refers explicitly to the contextual instance and can be omitted when the
contextual instance is obvious. In this example shown in Figure B.11, self refers to
the instance of Field class for which the operation d_dx is called. In the precondition, xor is the exclusive or operation, which evaluates to true only when exactly one
operand is true. The precondition species that the contextual instance must have
allocated values in either Fourier space or physical space but not both. The implies
operation in the postcondition indicates that the OCL expression is true when both
boolean operands before and after implies are true. If the boolean operand before implies is false, the OCL expression is always true. The postcondition states
that neither the Fourier nor physical space values of the contextual instance can be
allocated if the contextual instance is a temporary object.

Bibliography

Akin, E. (2003). Object-Oriented Programming via Fortran 90/95. Cambridge University Press,
Cambridge.
Alexander, C. (1979). The Timeless Way of Building. Oxford University Press, New York.
Alexander, C., S. Ishikawa, and M. Silverstein (1977). A Pattern Language: Towns, Buildings,
Construction, Volume 2 of Center for Environmental Structure Series. Oxford University
Press, New York.
Alexander, C., M. Silverstein, S. Angel, S. Ishikawa, and D. Abrams (1975). The Oregon
Experiment. Oxford University Press, New York.
Allan, B., R. Armstrong, D. E. Bernholdt, F. Bertrand, K. Chiu, T. L. Dahlgren, K. Damevski,
W. R. Ewasif, T. G. W. Epperly, M. Govindaraju, D. S. K. J. A. Kohl, M. Krishnan, G. Kumfert, J. W. Larson, S. Lefantzi, M. J. Lewis, A. D. Malony, L. C. McInnes, J. Nieplocha,
B. Norris, S. G. Parker, J. Ray, S. Shende, and T. L. W. andn S. Zhou (2006). A component architecture for high-performance scientic computing. Intl. J. High Perform. Comput.
Appl. 20(2), 163202.
Allan, B., B. Norris, W. R. Elwasif, R. C. Armstrong, K. S. Breuer, and R. M. Everson
(2008). Managing scientic software complexity with Bocca and CCA. Scientic Programming 16(4), 315327.
Aris, R. (1962). Vectors, Tensors, and the Basic Equations of Fluid Mechanics. Dover
Publications. Mineola, NY.
Artoli, A., A. Hoekstra, and P. Sloot (2006). Mesoscopic simulation of systolic ow in the
human abdominal aorta. Journal of Biomechanics 39, 837884.
Atlas, D., K. R. Hardy, and K. Naito (1966). Optimizing the radar detection of clean air
turnbulence, Journal of Applied Meterology 5(4), 450460.
Balay, S., K. Buschelman, V. Eijkhout, W. Gropp, D. Kaushik, M. Knepley, L. C. McInnes,
B. Smith, and H. Zhang (2007). Petsc users manual. Technical Report ANL-95/11 Rev.
2.3.3, Argonne National Laboratory.
Barker, V. A., L. S. Blackford, J. Dongarra, J. D. Croz, S. Hammarling, M. Marinova,
J. Waniewsky, and P. Yalamov (2001). LAPACK95 Users Guide. Society for Industrial
and Applied Mathematics.
Barrett, D. (1998). Polylingual systems: An approach to seamless interoperability. Ph. D. thesis,
University of Massachusetts, Amherst, MA.
Barton, J. and L. R. Nackman (1994). Scientifc and Engineering C++: An Introduction with
Advanced Techniques and Examples. Addison-Wesley, New York.
Beck, K. and W. Cunningham (1987). Using pattern languages for object-oriented programs. In OOPSLA-87 Workshop on the Specication and Design for Object-Oriented
Programming.
Benton, E. R. and G. W. Platzman (1972). A table of solutions of the one-dimensional burgers
equation. Quarterly of Applied Mathematics 30, 195212.
373

374

Bibliography
Berger, M. J. and P. Colella (1989). Local adaptive mesh renement for shock hydrodynamics.
J. Comput. Phys 82, 6484.
Bernard, P. S., J. M. Thomas, and R. A. Handler (1993). Vortex dynamics and the production
of Reynolds stress. Journal of Fluid Mechanics 253, 385419.
Bhatnagar, P., E. Gross, and M. Krook (1954). Model for collision processes in gases. Phys.
Rev. 94, 511525.
Blackford, L. S., J. Choi, A. Cleary, E. DAzevedo, J. Demmel, I. Dhillon, J. Dongarra,
S. Hammarling, G. Henry, A. Petitet, K. Stanley, D. Walker, and R. C. Whaley (1997).
ScaLAPACK Users Guide. Society for Industrial and Applied Mathematics.
Blackford, L. S., J. Demmel, J. Dongarra, I. Duff, S. Hammarling, G. Henry, M. Heroux,
L. Kaufman, A. Lumsdaine, A. Petitet, R. Pozo, K. Remington, and R. C. Whaley (2002).
An updated set of basic linear algebra subprograms (blas). ACM Trans. Math. Soft. 28(2),
135151.
Blilie, C. (2002). Patterns in scientic software. Computers in Science and Engineering 4(4),
4853.
Booch, G., J. Rumbaugh, and I. Jacobson (1999). The Unied Modeling Language User Guide.
Addison-Wesley, New York.
Boyd, J., J. Buick, J. Cosgrove, and P. Stansell (2004). Application of the lattice boltzmann
method to arterial ow simulation: Investigation of boundary conditions for complex arterial
geometries. Australasian Physical and Engineering Sciences in Medicine 27, 207212.
Bientinesi, P., Gunnels, J. A., M. E. M. E. S. Q.-O., and van de Geijn, R. A. (2005). The
science of deriving dense linear algebra algorithms. ACM Transactions on Mathematical
Software 31, 126.
Burgers, J. M. (1948). A mathematical model illustrating the theory of turbulence. Adv. Appl.
Mech. (1), 2527.
Canuto, C., M. Y. Hussaini, A. Quarteroni, and T. A. Zang (2006). Spectral Methods:
Fundamentals in Single Domains. Springer-Verlag, Berlin/Heidelberg.
Canuto, C. M. Y. Hussaini, A. Quarteroni, and T. Zang (2007). Spectral Methods. Evolution to Complex Geometries and Applications to Fluid Dynamics. Springer-Verlag,
Berlin/Heidelberg.
Chaitin, G. J. (1996). How to run algorithmic complexity theory on a computer: Studying the
limits of mathematical reasoning. Complexity 2(1), 1521.
Chapman, B., G. Jost, R. van der Pas, and D. J. Kuck (2007). Using OpenMP: Portable Shared
Memory Parallel Programming. The MIT Press, Cambridge, MA.
Chivers, I. D. and J. Sleightholme (2010). Compiler support for the Fortran 2003 and 2008
standards. ACM Fortran Forum 29(1), 2932.
Clarke, L. A. and D. S. Rosenblum (2006). A historical perspective on runtime checking in
software development. ACM SIGSOFT Software Engineering Notes 31(3), 2537.
Cole, J. D. (1951). On a quasilinear parabolic equation occuring in aerodynamics. Q. Appl.
Math. 9, 225236.
Collins, J. B. (2004). Standardizing an ontology of physics for modeling and simulation. In
Proceedings of the 2004 Fall Simulation Interoperability Workshop, Orlando, FL. Paper
04FSIW096.
Cosgrove, J., J. Buick, S. Tonge, C. Munro, C. Greated, and D. Campbell (2003). Application of the lattice boltzmann method to transition in oscillatory channel ow. Journal of
Physics 36, 26092630.
Dalhgren, T. L. (2007) Performance-Driven Interface Contract Enforcement for Scientic
Components. Ph.D. Dissertation, University of California, Davis (also Technical Report
UCRL-TH-235341, Lawrence Livermore National Laboratory).
Dahlgren, T. L., T. Epperly, G. Kumfert, and J. Leek (2009). Babel users guide. Technical
Report 230026, Lawrence Livermore National Laboratory, Livermore, CA.
Davidson, P. (2001). An Introduction to Magnetohydrodynamics. Cambridge University Press,
New York.
de Bruyn Kops, S. M. and J. J. Riley (1998). Direct numerical simulation of laboratory
experiments in isotropic turbulence. Phys. Fluids 10(9), 2527.

Bibliography
Decyk, V. K. and H. J. Gardner (2006). A factory pattern in Fortran 95, Volume 4487,
pp. 583590. Springer, Berlin/Heidelberg.
Decyk, V. K. and H. J. Gardner (2007). Object-oriented design patterns in Fortran 90/95.
Computer Physics Communications 178(8), 611620.
Decyk, V. K., C. D. Norton, and B. K. Szymanski (1997a). Expressing object-oriented concepts
in Fortran 90. ACM Fortran Forum 16(1), 1318.
Decyk, V. K., C. D. Norton, and B. K. Szymanski (1997b). How to express C++ concepts in
Fortran 90. Scientic Programming 6(4), 363390.
Decyk, V. K., C. D. Norton, and B. K. Szymanski (1998). How to support inheritance
and run-time polymorphism in Fortran 90. Computer Physics Communications 115,
917.
Deitel, P. J. (2008). C++: How to Program. Prentice Hall, Upper Saddle River, NJ.
Dijkstra, E. W. (1968). Go-to statement considered harmful. Communications of the
ACM 11(3), 147148.
Doviak, R. and D. Zrnic (1984). Doppler radar and weather observations. Academic Press,
New York.
Eichenberger, A. E., P. Wu, and K. OBrien (2004). Vectorization for simd architectures with
alignment constraints. ACM SIGPLAN Notices 39(6), 8293.
Fenton, N. E. and N. Ohlsson (2000). Quantitative analysis of faults and failures in a complex
software system. IEEE Transactions on Software Engineering 26(8), 797814.
Frisch, U., B. Hasslacher, and Y. Pomeau (1986). Lattice gas cellular automata for the navierstokes equations. Phys. Rev. Lett. 56, 15051508.
Gamma, E., R. Helm, R. Johnson, and J. Vlissides (1995). Design Patterns: Elements of
Reusable Object-Oriented Software. Addison-Wesley, New York.
Gardner, H. and G. Manduchi (2007). Design Patterns for e-Science. Springer-Verlag,
Berlin/Heidelberg.
Grant, P. W., M. Haveraaen, and M. F. Webster (2000). Coordinate free programming of
computational uid dynamics problems. Scientic Programming 8, 211230.
Gray, M. G., R. M. Roberts, and T. M. Evans (1999). Shadow-object interface between Fortran
95 and C++. Computing in Science and Engineering 1(2), 6370.
Hatton, L. (1997). The T Experiments: Errors in scientic software. IEEE Computational
Science and Engineering 4(2), 2738.
He, X. and L.-S. Luo (1997). Theory of the lattice boltzmann: from the boltzmann equation
to the lattice boltzmann equation. Phys. Rev. E 56, 68116817.
Henshaw, W. D. (2002). Overture: An object-oriented framework for overlapping grid applications. Technical Report 147889, Lawrence Livermore National Laboratory (also 32nd
AIAA Conference on Applied Aerodynamics, St. Louis, Missouri, June 2427, 2002).
Heroux, M. A., R. A. Bartlett, V. E. Howle, R. J. Hoekstra, J. J. Hu, T. G. Kolda,
R. B. Lehoucq, K. R. Long, R. P. Pawlowski, E. T. Phipps, A. G. Salinger, H. K. Thornquist,
R. S. Tuminaro, J. M. Willenbring, A. Williams, and K. S. Stanley (2005). An overview of
the trilinos project. ACM Transactions on Mathematical Software 31(3), 397423.
Hopf, E. (1950). The partial differential equation ut +uux = uxx . Commun. Pure Appl. Math.
3, 201230.
Huang, K. (1963). Statistical Mechanics, Wiley, Hoboken.
J3 Fortran Standards Technical Committee (1998). Technical report. Technical Report
ISO/IEC 15581:1998(E), International Organization for Standards/International Electrotechnical Committee, Geneva, Switzerland.
Karniadakis G. E. and Sherwin S. J. (2005). Spectral/hp Element Methods for Computational
Fluid Dynamics (Numerical Mathematics and Scientic Computation). Oxford University
Press, New York.
Kirk, S. R. and S. Jenkins (2004). Information theory-based software metrics and obfuscation.
J. Sys. Soft. 72(2), 179186.
Knuth, D. (1974). Structured programming with go to statements. Computing Surveys 6(4),
261301.

375

376

Bibliography
Koplik, J. and H. Levine (1993). Vortex reconnection in superuid helium. Physical Review
Letters 71, 13751378.
Krawezik, G., G. Allon, and F. Cappello (2002). SPMD OpenMP versus MPI on a IBM SMP
for 3 Kernels of the NAS Benchmarks, Volume 2327/2006 of Lecture Notes in Computer
Science, 515518.
Lewis, J. P. (2001). Limits to software estimation. ACM SIGSOFT Soft. Eng. Notes 26(4),
5459.
Lieber, B., A. Stancampiano, and A. Wakhloo (1997). Alteration of hemodynamics in
aneurism models by stenting: Inuence on stent porosity. Annals of Biomedical Engineering 25(3), 460469.
Linde, G. J. and M. T. Ngo (2008). WARLOC: A high-power coherent 94 GHz radar. In
Aerospace and Electronic Systems, IEEE Transactions, Milan, Italy, pp. 11021117.
Liu, Z., B. Chapman, T. Weng, and O. Hernandez (2003). Improving the Performance of
OpenMP by Array Privatization, Volume 2716/2003 of Lecture Notes in Computer Science,
244259.
Long, K. (2004). Sundance 2.0 tutorial. Technical Report 2004-4793, Sandia National
Laboratories.
Lorenz, E. N. (1963). Deterministic Nonperiodic Flow. Journal of the Atmospheric Sciences,
130141.
Lu, J., S. Wang, and S. Norville (2002). Method and apparatus for magnetically stirring a
thixotropic metal slurry. U.S. Patent No. 6,402,367 B1.
Maario, B., S. Viriato, and C. Francisco (2007). Dynamics in spectral solutions of burgers
equation. Journal of computational and applied mathematics 205(1), 296304.
Machiels, L. and M. O. Deville (1997). Fortran 90: An entry into object-oriented programming for the solution of partial differential equations. ACM Transactions on Mathematical
Software 23, 3242.
Malawski, M., K. Dawid, and V. Sunderam (2005). Mocca towards a distributed cca
framework for metacomputing. In Proceedings of the 19th IEEE International Parallel and
Distributed Processing Symposium (IPDPS05) Workshop 4 Volume 05. IEEE Computer
Society.
Manz, B., L. Gladden, and P. Warren (1999). Flow and dispersion in porous media:
Lattice-boltzmann and nmr studies. AIchE Journal of Fluid Mechanics and Transport
Phenomena 45(9), 18451854.
Markus, A. (2003). Avoiding memory leaks with derived types. ACM Fortran Forum 22(2),
16.
Markus, A. (2006). Design patterns and Fortran 90/95. ACM Fortran Forum 26(1), 1329.
Markus, A. (2008). Design patterns in Fortran 2003. ACM Fortran Forum 27(3), 215.
Martin, K. and B. Hoffman (2008). Mastering CMake 4th edition. Kitware, Inc.
Martin, R. C. (2002). Agile software development, principles, patterns and practices. Prentice
Hall, Upper Saddle River, NJ.
Mattson, T. G., B. A. Sanders, and B. L. Massingill (2005). Patterns in Parallel Programming.
Addison-Wesley, New York.
Metcalf, M., J. K. Reid, and M. Cohen (2004). fortran 95/2003 explained. Oxford University
Press, New York.
Mittal, R. (2005). Immersed boundary methods. Annual Reviews of Fluid Mechanics 37,
239261.
Moin, P. (2001). Fundamentals of Engineering Numerical Analysis. Cambridge University
Press, New York.
Morris, K. (2008). A direct numerical simulation of superuid turbulence. Ph. D. thesis, The
Graduate Center of The City University of New York.
Morris, K., R. A. Handler, and D. W. I. Rouson (2010, [in review]). Intermittency in the
turbulent ekman layer. Journal of Turbulence.
Morris, K., J. Koplik, and D. W. I. Rouson (2008). Vortex locking in direct numerical
simulations of quantum turbulence. Physical Review Letters 101, 015301.
Mulcahy, J. (1991). Pumping liquid metals. U.S. Patent No. 5011528.

Bibliography
Norton, C. D. and V. K. Decyk (2003). Modernizing Fortran 77 legacy codes. NASA Tech
Briefs 27(9), 72.
Numrich, R. W. (2005). Parallel numerical algorithms based on tensor notation and Co-Array
Frotran syntax. Parallel Computing, 31(6), 588607.
Numrich, R. W. and J. K. Reid (1998). Co-array fortran for parallel programming. ACM
Fortran Forum 17(2), 131.
Oliveira, J. N. (1997). http://www3.di.uminho.pt/ jno/html/camwfm.html.
Oliveira, S. and D. E. Stewart (2006). Writing Scientic Software: A Guide to Good Style.
Cambridge University Press, New York.
Pace, D. K. (2004). Modeling and simulation verication challenges. Technical Report 2, Johns
Hopkins APL Technical Digest.
Press, W. H., S. A. Teukolsky, W. T. Vetterling, and B. P. Flannery (1996). Numerical recipes
in fortran 90: The art of parallel scientic computing, volume 2. Cambridge University Press,
New York.
Press, W. H., W. T. Vetterling, B. P. Flannery, and S. A. Teukolsky (2002). Numerical Recipes
in C++: The Art of Scientic Computing, 2nd edition. Cambridge University Press, New
York.
Pressman, R. (2001). Software Engineering: A Practitioners Approach, 5th Ed. McGraw-Hill,
New York.
R. van Engelen, L. W. and G. Cats (1997). Tomorrows weather forecast: Automatic code
generation for atmospheric modeling. IEEE Computational Science and Engineering.
Rasmussen, C. E., M. J. Sottile, S. S. Shende, and A. D. Maloney (2006). Bridging the language
gap in scientic computing: the chasm approach. Concurrency and Computation: Practice
and Experience 18(2), 151162.
Reid, J. K. (2005). Co-arrays in the next fortran standard. ACM Fortran Forum 24(2), 417.
Rouson, D., K. Morris, and X. Xu (2005). Dynamic memory de-allocation in fortran 95/2003
derived type calculus. Scientic Programming 13(3), 189203.
Rouson, D., X. Xu, and K. Morris (2006). Formal constraints on memory management for
composite overloaded operations. Scientic Programming 14(1), 2740.
Rouson, D. W. I. (2008). Towards analysis-driven scientic software architecture: The case
for abstract data type calculus. Scientic Programming 16(4), 329339.
Rouson, D. W. M., H. Adalsteinsson, and J. Xia (2010). Design patterns for multiphysics
modeling in Fortran 2003 and C++. ACM Transactions on Mathematical Software 37(1),
Article 3.
Rouson, D. W. I., S. C. Kassinos, I. Moulitsas, I. Sarris, and X. Xu (2008a). Dispersed-phase
structural anisotropy in homogeneous magnetohydrodynamic turbulence at low magnetic
reynolds number. Physics of Fluids 20(025101).
Rouson, D. W. I., R. Rosenberg, X. Xu, I. Moulitsas, and S. C. Kassinos (2008b). A gridfree abstraction of the navier-stokes equation in fortran 95/2003. ACM Transactions on
Mathematical Software 34(1), Article 2.
Rouson, D. W. I. and Y. Xiong (2004). Design metrics in quantum turbulence simuluations:
How physics inuences software architecure. Scientic Programming 12(3), 185196.
Rubin, F. (1987). GOTO considered harmful considered harmful. Communications of the
ACM 30(3), 195196.
Satir, G. and D. Brown (1995). C++: The Core Language. OReilly & Associates, Inc.,
Sebastopol, CA.
Schach, S. (2002). Object-Oriented and Classical Software Engineering. Addison-Wesley, New
York.
Schwarz, K. W. (1985). Three-dimensional vortex dynamics in superuid 4 He: line-line and
line-boundary interactions. Physical Review B 31, 57825804.
Schwarz, K. W. (1988). Three dimensional vortex dynamics in superuid 4 He: homogeneous
superuid turbulence. Physical Review B 38, 23982417.
Shalloway, A. and J. R. Trott (2002). Design Patterns Explained. Addison-Wesley, New York.
Shannon, C. (1948). A mathematical theory of communication. The Bell System Technical
Journal.

377

378

Bibliography
Shepperd, M. and D. C. Ince (1994). A critique of three metrics. Journal of Systems and
Software 26, 197210.
Spalart, P. R., R. D. Moser, and M. M. Rogers (1991). Spectral methods for the navier-stokes
equations with one innite and two periodic directions. J. Comp. Phys. 96(297), 297324.
Stevens, W. P., G. J. Myers, and L. L. Constantine (1974). Structured design. IBM Systems
Journal 13(2), 115.
Stevenson, D. (1997). How goes cse? Thoughts on the IEEE cs workshop at Purdue.
Stewart, G. (2003). Memory leaks in derived types revisited. ACM Fortran Forum 22(3),
2527.
Stokes, G. (1851). On the effect of the internal friction of uids on the motion of pendulums.
Trans. Camb. Phil. Soc. 9(Part II), pp. 8106.
Strang, G. (2003). Introduction to Linear Algebra. Wellesley Cambridge Press, Wellesley,
MA.
Stucki, L. G. and G. L. Foshee (1971). New assertion concepts for self-metric software validation. In Proceedings of the International Conference on Reliable Software, pp. 5971. ACM
SIGPLAN and SIGMETRICS.
Tatarskii, V. (1961). Wave propagation in a turbulent medium. McGraw-Hill, New York.
Thomas, H. (1949). Elliptic problems in linear difference equations over a network. Technical
report, Watson Sci. Comput. Lab. Rept., Columbia University, New York.
Vandevoorde, D. and N. M. Josuttis (2003). C++ Templates, the complete guide. Addison
Wesley, New York.
Vinen, W. F. and J. J. Niemela (2002). Quantum turbulence. Journal of Low Temperature
Physics 128(56), 167231.
Wallcraft, A. J. (2000). Spmd openmp versus mpi for ocean models. Concurrency: Practice
and Experience 12, 11551164.
Wallcraft, A. J. (2002). A comparison of co-array fortran and openmp fortran for spmd
programming. The Journal of Supercomputing 22(3), 231250.
Warmer, J. and A. Kleppe (2003). The Object Constraint Language: Getting Your Models
Ready for MDA, 2nd Ed. Addison-Wesley, New York.
Won, J., S. Suh, H. Ko, K. Lee, W. Shim, B. Chang, D. Choi, S. Park, and D. Lee (2006).
Problems encountered during and after stentgraft treatment of aortic dissection. Journal of
Vascular and Interventional Radiology 17, 271281.
Xu, X. and J. Lee (2008). Application of the lattice boltzmann method to ow in aneurysm
with ring-shaped stent obstacles. Int. J. Numer. Meth. Fluids 59(6), 691710.
Zhang, K., K. Damevski, V. Venkatachalapathy, and S. G. Parker (2004). Scirun2: A cca
framework for high performance computing. In Ninth International Workshop on HighLevel Parallel Programming Models and Supportive Environments (HIPS04), pp. 7279.

Index

abstract, 34
calculus, 129, 130, 132, 137, 140, 141
class, 94
data type, 34
data type calculus, 58, 59, 62, 63, 66, 69, 70, 72,
73, 7578, 80, 99, 130, 132, 141
interface, 79
abstract calculus, 202, 236, 293, 310
abstract factory, 228, 234
abstract factory pattern, 99
abstract syntax tree, 253
actual argument, 54
adapter pattern, 99
ADT
calculus, 111
ADT calculus, 141
aggregation, 48, 57, 62, 78, 365367
allocatable, 35, 38, 41, 62
allocation
sourced, 155
ANSI C, 254
application, 96
application programming interface
(API), 295
architecture, 7
argument
actual, 41, 54, 79, 145, 202, 268, 269
dummy, 41, 50, 79, 111, 135, 268
passed-object dummy, 38, 51, 54, 135
assertion, 232, 233, 241, 242, 247
assertions, 232
assignment
dened, 55
intrinsic, 55, 58, 70, 78
association, 362
attribute, 34
auto-parallelization, 293
auto-vectorization, 293
Babel, 254
back substitution, 337, 338, 341, 343
backward difference, 349

backward Euler, 59
beginners mind, 87
binding
deferred, 95
bisection method, 71, 72
BLACS, 108
BLAS, 141
body count, 262
Boolean, 35
BOOST, 116
Builder, 227
Burgers equation, 202, 204, 310
buttery effect, 101
bzip2, 69
C++
0x standard, 99
class, 36
data members, 36
exception handling, 137
function prototype, 39
namespace, 36
Standard Template Library, 23, 36, 129
virtual member function, 36
call tree, 234
central difference, 349
chain of responsibility pattern, 99
chaos, 101
character string
deferred-length, 38
xed-width, 38
Chasm, 252254
child, 57, 78
class
interface, 95
coarray, 309
cohesion, 6971, 109
Common Component Architecture (CCA), 5, 23
compiler
Cray Fortran, 31
Gnu Fortran, 31
Intel, 31

379

380

Index
compiler (cont.)
Numerical Algorithms Group, 31
XL Fortran, 31
complexity, 67, 69, 79
algorithmic, 67, 69
computational, 69
Kolmogorov, 67, 76
theory, 71
component, 23
-based, 24
-based software engineering, 23
derived type, 29
composition, 48, 57, 365367
concrete class, 94
concurrency, 55
constructor, 32, 34, 38, 4143, 49, 50, 53, 55
generic name, 37
context, 370
copy
deep, 70
shallow, 70
coupling, 6971, 81
Cray, 31
dead-code removal, 28, 246
decorator pattern, 99
deferred binding, 79, 95
dened assignment, 55
dependency, 362
depth count, 111
derived type
components, 29
design by contract, 24
design patterns, 5
creational, 203
diagonally dominant, 339
distributed programming, 99
dynamic dispatching, 54
encapsulation, 21, 31, 36, 39, 50, 54
exception handling, 136
expression templates, 77
extended types, 22
extensibility, 45, 54, 55
Extensible Markup Language (XML), 252
facade pattern, 108
factory method, 203, 204, 234
fault, 231
nal subroutine, 109111
nite differences
central, 105
Pade, 105
Runge-Kutta, 105
attening, 255, 257
fork-join, 295
formal
methods, 231, 232
specication, 231

verication, 231
formal constraints, 24
formal methods, 24
Fortran
2008 standard, 39, 99, 155
forward difference, 348
forward elimination, 337
forward Euler, 58
forward reference, 145
framework, 96, 331
function
intrinsic, 22, 41, 53
function evaluation, 131
Gang of Four (GoF), 9396, 99, 127, 141144, 164,
202204, 226228, 240
garbage collection, 117
Gaussian elimination, 337342
generalization, 358, 362, 363
generic
constructor name, 37
interface, 41
name, 41
programming, 35
type, 32
graphical user interface, 98
GUI, 98
heat equation, 345
hierarchy
class, 54, 57
inheritance, 54
high-performance computing (HPC), 4, 107
IBM, 31
IEEE, 256
implicit Euler, 351
information
entropy, 74, 75
theory, 69, 73
information hiding, 21, 31, 36, 38, 39, 54, 57
inheritance, 22, 31, 48, 57, 62, 75, 78, 363, 364
initialization, 255, 257
inline, 253
instantiate, 34
interface, 19, 22, 32, 39, 4143, 50, 53, 57, 58, 61,
67, 73, 75, 76, 78, 79, 361, 363, 370, 371
abstract, 79, 95
blocks, 39
class, 94, 95
explicit, 39
graphical user, 98
implicit, 39
logical, 255
physical, 255
specications, 39
interface block, 262
interface body, 262
intermediate object representation (IOR), 254

Index
intrinsic assignment, 55, 58, 70, 78
intrinsic function, 22, 41, 53
intrinsic operation, 62
intrinsic type, 35
invariant, 240
iterator pattern, 127
Java, 38, 98, 99, 108, 254
Virtual Machine, 98
kernel, 245
koan, 87
LAPACK, 141
legacy
interfaces, 57
legacy Fortran, 40, 41, 45
Lorenz equations, 101
magnetohydrodynamics (MHD), 100
Mathematical Acceleration Subsystem (MASS),
293
memory leak, 41
memory management, 231
message, 34, 36, 38, 53, 62, 74
Message Passing Interface (MPI), 297
method, 21, 23, 32, 34, 35, 41, 50, 53, 54
module, 62
molecular diffusion, 100
Morfeus, 330
Morpheus, 53
multicore, 294
multiphysics, 3, 4, 6, 100, 285
multiplicity, 49, 243, 367
multithreading, 99, 293295
multphysics, 7
name mangling, 255
Newtons method, 343
object, 34
object pattern, 108, 109, 127, 237
object-based, 251, 255
object-oriented
analysis (OOA), 33, 34, 38, 40, 45
design (OOD), 5, 33, 34, 38, 40, 42, 43, 45
programming (OOP), 6, 12, 19, 2124, 26,
3134, 3638, 40, 45, 48, 50, 54, 57, 62
object-oriented analysis (OOA), 285
object-oriented design (OOD), 285
object-oriented programming (OOP), 285
observer pattern, 99
OCL (see UML), 232
opaque pointer, 257
Open Multi-Processing (OpenMP), 20, 295
operating system, 257
operation, 32, 3436, 43, 51, 53
intrinsic, 62

overloading, 256
operator, 22
operator , 22
package, 130
Pade, 226
parallel programming, 99
parent, 57, 75, 78
partial pivoting, 338, 339, 342, 343
passed-object dummy argument, 38, 51, 54, 135
pattern language, 86, 88, 141
PETSc, 141
phase space, 102
pointer
assignment, 112
polymorphism, 22, 31, 32, 45, 53, 54
dynamic, 38, 53, 54, 57
static, 53
postcondition, 240
precondition, 240
primitive type, 35
procedure
external, 39, 41
nal, 42
module, 21, 39, 53
signature, 32, 79
type-bound, 36, 38, 39, 51, 53, 55, 75
procedure pointer, 215
programming
function, 26
functional, 20, 21
generic, 23
object-based, 21, 22, 24
object-oriented, 4, 6, 22
purely functional, 21
structured, 20
unstructured, 20, 21
protocol, 54
puppeteer, 169, 202, 227
Python, 254
quantum vortices, 100
realization, 95
refactor, 40, 59
reference counting, 117
reference-counted pointer (RCP), 117, 273
reuse, 22, 38, 48, 54, 57, 58
reverse engineering, 61, 62, 69
Runge-Kutta, 131
2nd order (RK2), 144
Runge-Kutta methods, 216
scalable abstract calculus, 285
ScaLAPACK, 108, 141
scaling
strong, 292
weak, 292

381

382

Index
Scientic Interface Denition Language (SIDL),
254
scientic OOP (SOOP), 57, 62, 79, 80
scope, 62
private, 34, 39
public, 34
semidiscrete, 101, 347
sequence association, 268
shadow object, 255
shell, 245, 248, 250, 269, 270
shoshin, 87
side effects, 13, 21, 29, 238
signature, 32, 79
Simplied Wrapper and Interface Generator
(SWIG), 252
single instruction multiple data, 293
skeleton, 253
software engineering
component-based (CBSE), 23
speedup, 291
state pattern, 99
stereotype, 34
STL, 23, 36
strategy, 202
strong scaling, 292
struct ID, 269, 270, 273
structure constructor, 37, 55
structured
programming, 20, 24
stub, 253
superuidity, 100
surrogate, 202
symmetric multiprocessing (SMP), 294
Taoism, 87
template, 23, 32
class, 35, 36, 117, 121
emulation, 23
expression, 77
metaprogramming, 116
template method, 228
tensor, 331
calculus, 309
test-driven development, 43

toolkit, 96
Trilinos, 108, 141, 169
truncation error, 348
type
child, 78
derived, 21
extended, 50, 53, 54, 79
extension, 49
nalization, 111
generic, 32, 35
guard, 136
intrinsic, 35, 59, 78
parameter (UML), 35
parameterized, 32
parameterized (UML), 35
parent, 57, 75, 78
primitive, 35
selection, 135
type-bound procedure, 21
typed allocation, 215
types
extended, 22
UML, 33, 232
adorned relationships, 205
attribute, 3436
class diagram, 3436, 49, 56, 357, 359, 366
generalization, 50, 363, 364
object diagram, 49, 357, 366
OCL, 232
operation, 32, 3436, 43, 51, 53
package, 130
package diagram, 357
realization, 95
sequence diagram, 62, 66, 80, 357
use case, 33, 34, 36, 42, 357, 358
Unied Modeling Language (UML), 32
annotation, 240
weak scaling, 292
wrapping legacy Fortran, 40, 41, 45
Zen Buddhism, 87