Autodyn Parallel Processing Tutorial

ANSYS, Inc. Release 15.0

Southpointe November 2013
275 Technology Drive
Canonsburg, PA 15317 ANSYS, Inc. is
[email protected] certified to ISO
9001:2008.
http://www.ansys.com
(T) 724-746-3304
(F) 724-514-9494
Copyright and Trademark Information

© 2013 SAS IP, Inc. All rights reserved. Unauthorized use, distribution or duplication is prohibited.

ANSYS, ANSYS Workbench, Ansoft, AUTODYN, EKM, Engineering Knowledge Manager, CFX, FLUENT, HFSS and any
and all ANSYS, Inc. brand, product, service and feature names, logos and slogans are registered trademarks or
trademarks of ANSYS, Inc. or its subsidiaries in the United States or other countries. ICEM CFD is a trademark used
by ANSYS, Inc. under license. CFX is a trademark of Sony Corporation in Japan. All other brand, product, service
and feature names or trademarks are the property of their respective owners.

Disclaimer Notice

THIS ANSYS SOFTWARE PRODUCT AND PROGRAM DOCUMENTATION INCLUDE TRADE SECRETS AND ARE CONFID-
ENTIAL AND PROPRIETARY PRODUCTS OF ANSYS, INC., ITS SUBSIDIARIES, OR LICENSORS. The software products
and documentation are furnished by ANSYS, Inc., its subsidiaries, or affiliates under a software license agreement
that contains provisions concerning non-disclosure, copying, length and nature of use, compliance with exporting
laws, warranties, disclaimers, limitations of liability, and remedies, and other provisions. The software products
and documentation may be used, disclosed, transferred, or copied only in accordance with the terms and conditions
of that software license agreement.

ANSYS, Inc. is certified to ISO 9001:2008.

U.S. Government Rights

For U.S. Government users, except as specifically granted by the ANSYS, Inc. software license agreement, the use,
duplication, or disclosure by the United States Government is subject to restrictions stated in the ANSYS, Inc.
software license agreement and FAR 12.212 (for non-DOD licenses).

Third-Party Software

See the legal information in the product help files for the complete Legal Notice for ANSYS proprietary software
and third-party software. If you are unable to access the Legal Notice, please contact ANSYS, Inc.

Published in the U.S.A.

Table of Contents
1. Introduction ............................................................................................................................................ 1
2. Parallelization Methodology .................................................................................................................. 3
2.1. Parallelization of Structured Part Calculations .................................................................................... 3
2.2. Parallelization of SPH Calculations ..................................................................................................... 3
2.3. Parallelization of Unstructured Calculations ....................................................................................... 4
2.4. Parallelization of Contact Interactions ............................................................................................... 4
2.5. Parallelization of Euler-Lagrange Coupling Interactions ..................................................................... 5
3. Establishing a Parallel Processing Environment ..................................................................................... 7
3.1. Windows Systems using Platform MPI ............................................................................................... 9
3.1.1. Running Autodyn with MPIRUN ................................................................................................ 9
3.2. Linux Systems using Platform MPI ................................................................................................... 10
3.2.1. Host Configuration Set to Run Autodyn in Parallel ................................................................... 10
4. Creating Domain Decomposition Sets .................................................................................................. 13
4.1. Decomposing Structured Calculations ............................................................................................. 13
4.2. Decomposing SPH Calculations ....................................................................................................... 27
4.3. Decomposing Unstructured Calculations ........................................................................................ 30
4.4. Decomposing Euler-Lagrange Coupled Problems ............................................................................ 35
4.5. Fully Automatic Decomposition on Windows ................................................................................... 38
A. Overview of Parallelized Features and Known Limitations ........................................................................ 41

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
of ANSYS, Inc. and its subsidiaries and affiliates. iii
 Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
iv of ANSYS, Inc. and its subsidiaries and affiliates.
Chapter 1: Parallel Processing Introduction
Autodyn has been developed specifically for analyzing non-linear, dynamic events such as impacts and
blast loading of structures and components. The program offers users a variety of numerical techniques
with which to solve their problems. These include Lagrange, Shell, Euler, ALE (Arbitrary Lagrange Euler)
and SPH (Smooth Particle Hydrodynamics) solvers. As reliance on computational simulations becomes
accepted, the complexity of the problems to be solved increases in size and resolution. However, the
practical computation of these very large simulations has been restrained by the lack of performance
of available computers. Problems requiring millions of elements and run-times that can run many weeks
are not uncommon. Even the fastest single CPUs cannot easily cope with these larger problems.

One approach to overcoming these limitations is to utilize parallel systems. Parallel algorithms have
been implemented in Autodyn to take advantage of parallel systems that allow simultaneous use of
multiple CPUs either on a single multi-processor machine or over a distributed network of computers.
This tutorial describes the method used by Autodyn to process problems in parallel and explains how
the user sets up and runs a calculation using parallel processing. Currently Autodyn supports the use
of a maximum of 127 tasks (slave processes) in a single parallel analysis.

Users set up problems for parallel processing in exactly the same way as they do for serial processing,
and the processing of results (for example, plotting, saving) are also performed in the usual way.

At the current time, the structured as well as unstructured Part calculations of all 3D solvers have been
parallelized:

• Lagrange (with joins)

• ALE (with joins)

• Shell (with joins)

• SPH (with joins)

• Euler Ideal Gas

• Euler Multi-Material

• Beams (with joins)

Note

Although the joins between unstructured parts have been parallelized it is strongly advised
to avoid the use of unstructured parts in combination with joins in parallel analyses.

In Autodyn a new option is available that will merge joined unstructured nodes that reside
at the same physical location in the model into one single unstructured node. The option is
available under the Join menu and will increase robustness in many applications involving
joins.

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
of ANSYS, Inc. and its subsidiaries and affiliates. 1
Introduction

Coupled calculations have been parallelized between the Euler Ideal Gas and Euler multi-material solver
and structured and unstructured Lagrange and Shell solvers.

Interaction calculations have been parallelized between Lagrange, ALE, Shell and SPH solvers, using the
Gap and Trajectory contact algorithm. Parallel simulations using Trajectory contact are subject to the
following restrictions:

• The contact method must be set to penalty; the Decomposition Response method is not available in
parallel.

• Trajectory contact cannot be run in parallel when an SPH part is present in the model.

Bonded connections are available as a beta option in parallel.

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
2 of ANSYS, Inc. and its subsidiaries and affiliates.
Chapter 2: Parallelization Methodology
This chapter discusses the various methodologies used to parallelize the different types of solutions
available in Autodyn.

2.1. Parallelization of Structured Part Calculations

Domain Decomposition is used to parallelize the structured Part computations performed by all solvers
which utilize an IJK mesh. Using this method, each part is divided along index planes in the I, J and K
directions (J and K directions for Shell parts) to form smaller parts called sub-domains. At the present
time, users must define the domain decomposition manually to be used for each problem they wish
to run in parallel.

These sub-domains are distributed among the CPUs of the parallel machine using an algorithm that
attempts to minimize inter-CPU communications and balance the computational load on each machine.
Each sub-domain is processed in parallel as if it were a standard part in serial processing. This importantly
allows most of the source code for serial processing to be used without modification.

Because the part structure does not normally change during simulations, a static decomposition of the
entire index space is usually sufficient to achieve good parallel performance.

Efficient parallel processing of the Part calculations requires not only good load-balancing of the com-
putation, but also the efficient exchange of data at sub-domain boundaries. This is best achieved by
users clearly understanding the process involved and choosing their domain decompositions to suit
the specific host configuration they intend to use.

Automatic decomposition can be used to decompose Euler-FCT Parts using the recursive bi-section
method described below.

2.2. Parallelization of SPH Calculations

One of the main benefits of the SPH solver is its ability to evaluate the governing variables of a nonlinear
dynamic analysis without utilizing a mesh or grid structure, thus allowing for very high levels of deform-
ation. This benefit means that the domain cannot be decomposed by choosing grid lines and intersections
as occurs with grid calculations.

In order to decompose the SPH calculation we utilize the virtual work units which are set up to facilitate
the searching algorithm which is integral to the SPH solver. Each work unit is a cube and together they
encompass the entire computational domain with a grid structure. The number of SPH nodes within a
work unit is evaluated and a similar algorithm to that used to decompose the grid calculations is used
to decompose the work units and therefore SPH nodes in order to minimize inter-processor communic-
ation. At present this is a static decomposition; each SPH node stays on the same processor for the
length of the calculation.

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
of ANSYS, Inc. and its subsidiaries and affiliates. 3
Parallelization Methodology

2.3. Parallelization of Unstructured Calculations

As the numbering of nodes/elements of the unstructured solvers does not adhere to the same restrictions
as the structured solvers, the parts cannot be simply decomposed by I, J, K index lines. The decompos-
ition is done in a similar manner to SPH calculations. Virtual work units are applied to the model and
the work done in each work unit is evaluated. A recursive bisection method is then applied in order to
effectively load balance the model. The amount of communication between processors is minimized
but this is a secondary consideration compared to the load balancing. This method is completely
automatic and requires a minimal amount of information from the user.

2.4. Parallelization of Contact Interactions

To understand how contact interactions are parallelized, we must first understand how the algorithm
works in serial calculations.

Contact logic is used when one surface element of a part attempts to penetrate another surface element
of the same or a different part. This requires a global search of Cartesian space to find possible contacts.
Using the same sub-domain decomposition used for the grid calculation is not an efficient way to par-
allelize contact interactions. This is because the contact surfaces that need to be processed come from
elements that lie on the surface of a part and thus comprise only a subset of the total elements in the
part. It is easy to see that some sub-domains might contain many surface elements, while others none
at all. Moreover, if a calculation allows the erosion or removal of elements as penetration occurs, the
actual surface elements of a part will change as the calculation proceeds. We are therefore forced to
use a second, dynamic domain decomposition of Cartesian space for the contact calculations.

Generally, any two surface elements anywhere in the simulation can come in contact with each other
during some time step, even those that belong to the same object (self-interaction). Checking for all
such contacts requires a global search in Cartesian space that in practice can take up to 50% of the
overall CPU time. For efficiency, the contact nodes and faces are spatially sorted to speed this compu-
tation and to avoid unnecessary tests of distant elements. Thus, the contact algorithm used in Autodyn
can be considered in two parts. Firstly, a calculation is performed to identify neighboring nodes/faces
that require to be checked for interaction. Secondly, a detailed interaction calculation is performed for
all these identified nodes/faces.

Determining which nodes/faces require to be checked for interactions is achieved with a bucket-sort.
A grid of virtual work units is defined in Cartesian space. Each work unit is a cube, with sides twice the
smallest face dimension of all interacting faces. In tests, this cube size was found to not only yield the
most efficient computing times (due to the fine sort), but also to generate sufficient work units to allow
efficient load-balancing for the parallelization. These work units are virtual because storage for a partic-
ular work unit is only allocated when it is determined that the work unit contains nodes/faces that are
to be tested for interaction.

The bucket-sort loops over all the surface nodes and faces of a problem, and constructs a list of the
actual work units required for a particular analysis. The sort is performed in two passes, in which all the
nodes are sorted, and then the faces are sorted. First, each surface node is added to the work unit,
which contains it. Next, each surface face is added to all work units, which contain nodes that might
interact with the face. This is determined by checking each node of the face to see if it is contained
within, or is in close proximity to, a work unit’s domain. At this stage, only work units that already
contain surface nodes are considered. The proximity test is based on the size of the contact detection
zone used for the interaction logic and the amount of “slack” allowed enabling the calculations described
here to be performed less frequently than every cycle.

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
4 of ANSYS, Inc. and its subsidiaries and affiliates.
 Parallelization of Euler-Lagrange Coupling Interactions

Finally, the node and face tables built for each work unit in the list are examined to determine the total
number of node/face interactions that will be required to be computed for the work unit (this number
is used to facilitate load-balancing in the parallel version). In general, this will equal the total number
of nodes in the work unit times the total number of faces. However, this can be reduced if, for example,
self-interaction of a part with itself is not permitted, or two parts have been specified not to interact
with each other. If the total number of interactions required to be computed for a work unit is found
to be zero, then the work unit is removed from the list.

At the end of this procedure a compact group of work units has been generated, each containing a list
of surface nodes and faces that require testing for interaction. Each node has been uniquely assigned
to a particular work unit. Faces have been assigned to multiple work units, as required. These lists may
be valid for a number of computational cycles, depending on the proximity test used to determine
potential node-face interactions and on changes in surface definitions (if an element is eroded or re-
moved, surfaces need to be redefined).

Within each work unit, detailed interaction calculations are performed between the nodes and faces in
each work unit list. The calculation is very robust in that every impact is detected and dealt with correctly
regardless of the deformation and relative movement of bodies or changes in surface definitions

Parallelization of the contact algorithm described above is fairly straightforward. Once the work units
containing the node and face lists to be tested for impact have been generated, a load-balancing al-
gorithm efficiently distributes them among the available CPUs, assuming each CPU has either the same
speed or a pre-determined relative speed provided by the user. The data decomposition used for the
contact interaction calculation is different from the one used for the grid calculation, so the load-balancing
algorithm attempts to minimize the inter-CPU communications required between these two decompos-
itions. Although a static decomposition is used for the grid calculation, a dynamic decomposition has
to be used for the contact calculation. Consequently, load balancing of the newly formed work units is
performed for each cycle on which a sort is carried out. This allows contact calculations to remain well
load-balanced even when surfaces are reconfigured as the simulation progresses, or during the erosion
(removal) of elements.

Results have shown that the contact algorithm generates sufficient work units during the sort phase
to allow efficient load balancing for parallel processing. Furthermore, the scheme uses simpler commu-
nication patterns than those that rely on recursive coordinate bisection (RCB) to assign equal amounts
of nodes to all CPUs, and adapts well to heterogeneous systems where CPUs may have different speeds
and workloads that may vary with time.

2.5. Parallelization of Euler-Lagrange Coupling Interactions

Aside from handling the decomposition of unstructured parts, the automatic parallel decomposition
algorithm can also handle the decomposition of Euler parts, Euler Ideal Gas as well as Euler Multi-ma-
terial. This facilitates the decomposition of complicated Euler/Lagrange coupled models; the user only
needs to define the number of tasks over which the model should be assigned and Autodyn will auto-
matically produce a decomposition configuration with good load balancing qualities and minimal inter-
processor communication.

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
of ANSYS, Inc. and its subsidiaries and affiliates. 5
Parallelization Methodology

To further enhance the efficiency of coupled calculations the sub-domains of the FE structure will be
placed on the same processor as the Euler sub-domains located in the same geometric space. This de-
creases the necessary inter-processor communication for the coupling calculations.

Table 2.1: 8 CPU Automatic Decomposition of a Casing Fragmentation Analysis

Note

The multi-material Euler part should not be joined to other Euler Parts.

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
6 of ANSYS, Inc. and its subsidiaries and affiliates.
Chapter 3: Establishing a Parallel Processing Environment
Autodyn has been designed for parallel processing on a variety of systems ranging from a Massively
Parallel Processor (MPP) using shared memory to heterogeneous distributed networks of computers.

In this chapter, we outline the procedure for establishing a parallel processing environment for Autodyn.
The procedure should be similar for most platforms. Platform specific procedures for configuring the
parallel processing environment are given in referenced appendices.

When using Autodyn for parallel processing, data must be exchanged between cooperating tasks, and
some message passing protocol has to be used to achieve this. We currently use MPI (Message Passing
Interface) to allow a heterogeneous collection of computers networked together to be viewed by
Autodyn as a single parallel computer. The component computers can be single or multiple processor
machines, including MPPs (Massively Parallel Processors).

If you wish to use the parallel processing options described in this tutorial, you must install the Platform
MPI message passing protocol on all the machines you intend to use.

Platform MPI is a part of the unified installation packages on both Linux and Windows platforms.

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
of ANSYS, Inc. and its subsidiaries and affiliates. 7
Establishing a Parallel Processing Environment

Figure 3.1: Schematic of the Parallel Processing Environment

Whether you intend to run a calculation in serial mode or parallel mode, Autodyn is always started by
activating a single process on one machine. This is done by starting the autodyn process.

If a serial calculation is being performed, this is the only process that is run. It performs all aspects of
the problem (for example, setup, execution, I/O, post processing).

If a parallel calculation is being performed, this process is considered to be the master process and
additional slave processes are started on each CPU used by the parallel system (including the CPU
running the master process). These slave processes are spawned automatically by the master process,
which issues the required system commands to start the adslave process once on each CPU. This
parallel processing environment is shown schematically in the figure above.

In what follows, we assume that the configuration we are implementing comprises of n hosts, each
having m CPU’s, and that the hostname of the nth host is hostn. Host1 is special, in that it runs the
master process (autodyn) and is usually the system on which the program files are stored.

When running Autodyn in parallel on a system of networked computers, the program executables and
libraries must be accessible on each of the computers. They can either be installed on all computers in
the system or installed on one computer (usually the one running the master process) and shared with
the others using NFS mounting or Windows Sharing. We recommend sharing as it simplifies the handling
of the files. Setup is further simplified if the shared mount points/folders are the same on each host
(including the master).

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
8 of ANSYS, Inc. and its subsidiaries and affiliates.
 Windows Systems using Platform MPI

3.1. Windows Systems using Platform MPI

Autodyn is made available on Windows operating systems using Platform MPI as the parallel message
passing protocol. Dynamic spawning of slave processes from the Autodyn component system is not
possible. Therefore the number of slave tasks is specified before starting the Autodyn program using
mpirun. The process for this is outlined below. Once the Autodyn executable is started the user would
have to restart their Autodyn session should they require a different number of slave tasks.

To establish the parallel environment Platform MPI must be installed on each machine in the cluster,
following the installation instructions for Platform MPI as provided in ANSYS online documentation.
Users may find it necessary to set the MPI password on each machine in their cluster. To do this they
can call mpirun with the -cache argument the first time they run mpirun, entering their user password
when prompted. This argument should be removed for subsequent parallel runs.

Note

• Please refer to the MPI documentation found in your MPI installation (C:\Program Files
(x86)\IBM\Platform-MPI\doc\pcmpi.win.09.01.00.00.rn.pdf for a default in-
stallation) for the description of command options such as: -cache, -pwcheck, -pass, and so on.

• Please be aware that for the Windows 7 (and higher) OS, you must disable User Account Control
(UAC) on all of the machines being used for the parallel solves. To do so:

1. In the Windows 7 Search box on the Start menu, type UAC and press Enter.

2. Move the UAC settings slider to Never notify, click OK, and then reboot the machine.

• The mixture of 32-bit and 64-bit Windows platforms in distributed runs is currently not suppor-
ted. Both master and slave processes must be on two or more 32-bit machines, or on two or
more 64-bit machines.

When Platform MPI is installed using the ANSYS installer, it will install Platform MPI version 9.1. The in-
staller will set MPI_ROOT to point to this version and will add MPI_ROOT\bin to the path environment
variable.

3.1.1. Running Autodyn with MPIRUN

Create a file called applfile. The location of the file for Windows 7 systems is C:\Users\<user-
name>\AppData\Roaming\Ansys\v150\AUTODYN. For Windows XP systems, the location of the
file is C:\Documents and Settings\<username>\Application Data\Ansys\v150\AUTO-
DYN. In this file enter the following text, making the relevant changes for your parallel model, machine
names, and installation path should the installation not be in the default location:
-e MPI_FLAGS=y0 -e MPI_ANSYS_EXD=ON -h machine1
-np 1 "C:\Program Files\ANSYS Inc\v150\AISOL\AUTODYN\winx64\autodyn.exe"
-h machine1 -np 2 "C:\Program Files\ANSYS Inc\v150\AISOL\AUTODYN\winx64\adslave.exe"
-h machine2 -np 2 "C:\Program Files\ANSYS Inc\v150\AISOL\AUTODYN\winx64\adslave.exe"

The above applfile will launch 1 master process, 2 slaves on machine1, and 2 additional slaves on
machine2. If using Windows XP 32-bit, winx64 should be replaced by intel.

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
of ANSYS, Inc. and its subsidiaries and affiliates. 9
Establishing a Parallel Processing Environment

In order to signify to the solver that the solver is running in parallel and to cross-check this with the
user settings in the model file, it is required to pass the MPI_ANSYS_EXD environment flag to the
solver. The solver also cross-checks the MPI_FLAGS environment variable.

Autodyn is then started using mpirun. It is important that you reference the correct version of MPI, or
else the wrong mpirun command is picked up from the path environment variable. This can be done
globally as noted in Windows Systems using Platform MPI (p. 9) by permanently setting the MPI_ROOT
to point to version 9.1, or by temporarily changing it through a batch script.

To start Autodyn, you could call mpirun from a command prompt, but we suggest that you create a
batch file in the same application data directory as the applfile file and execute that. For example,
create a blank file called autodyn_mpi.bat. In this file enter:
set MPI_ROOT=%AWP_ROOT150%\commonfiles\MPI\Platform\9.1\winx64
"%MPI_ROOT%\bin\mpirun.exe" -prot -e MPI_WORKDIR="C:\Program Files\ANSYS Inc\v150\AISOL\AUTODYN\winx64" -f applfile

The MPI_WORKDIR text should be modified if the user’s installation of Autodyn is not in the default
location.

Once the batch file is executed, the Autodyn program starts along with the number of slaves specified
in applfile. The user can now create their model or load an existing one as usual.

3.2. Linux Systems using Platform MPI

In what follows, we assume that Autodyn and Platform MPI files are installed on host1, in a directory
named /usr/ansys_inc/, and that this directory is NFS mounted on all other hosts (with the same
mount point). To NFS mount this directory, add the following line to the file /etc/fstab on each
host (except host1):
host1: /usr/ansys_inc/ /usr/ansys_inc/ nfs defaults 0 0

/etc/fstab may have a different name for other Linux systems

Following the normal installation procedures for Autodyn products, all Autodyn files will be installed
in /usr/ansys_inc/v150/autodyn/. In particular, the Autodyn master and slave executables (autodyn
and adslave) will reside in the directory /usr/ansys_inc/v150/autodyn/bin/linx64.

After installing Autodyn and Platform MPI on all the cluster machines, the environment variable MPI_ROOT
can be set to be the directory where Platform MPI is installed and the path to the MPI_ROOT/bin directory
added to the path environment variable. These changes are not required to run Autodyn in parallel,
but when set they will override the defaults used by Autodyn.

3.2.1. Host Configuration Set to Run Autodyn in Parallel

To run Autodyn in parallel, you must define the host machine configuration you wish to use. Autodyn
allows you to define up to ten host configuration sets and select one of these sets as the active config-
uration.

Host configuration set data is saved in an external file called parallel.cfg, located in your running
directory. This is a text file that must be edited manually.

Here is an example of the typical contents of this file.

#@EPDEF=\\host1\autodyn (1)
#@PPDEF office (2)
#@PPCFG office (3)
host1 sp=1000 (4)

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
10 of ANSYS, Inc. and its subsidiaries and affiliates.
 Linux Systems using Platform MPI

#@ mem=128 cpu=1 task=1 (5)

host2 sp=500 ep=\\host2\autodyn (6)
#@ mem=256 cpu=2 task=2 (7)

(1) Defines the default path for the Autodyn master and slave executables (autodyn.e150 and ad-
slave.e150) to be \\host1\autodyn , where autodyn is a shared directory on host1. When
the master and slave executables reside in the directory /usr/ansys_inc/v150/auto-
dyn/bin/linx64, the share \\host1\autodyn should point to that directory.

(2) Defines the configuration set name that is to be used (office).

(3) Indicates that the following lines, until the next #@PPCFG statement, defines configuration set office.
(There is only one configuration set defined in this example.)

(4) Adds host1 to configuration set office. It has a relative speed of 1000 and uses the default path
for the Autodyn executable (since no ep parameter is defined).

(5) Specifies host1 has 128 Mb of memory, 1 CPU, and is to have 1 task (slave process) started on it.

(6) Adds host2 to configuration set office. It has a relative speed of 500 and /autodyn as the path
for the Autodyn master and slave executables.

(7) Specifies host2 has 256 Mb of memory, 2 CPUs, and is to have 2 tasks (slave processes) started on
it. The operating system will automatically allocate 1 task to each CPU.

The parallel.cfg file is an external file that is read whenever a project is executed in Autodyn. Host
configuration sets are therefore not project dependent (unlike the domain decomposition sets described
in the next chapter). When you create or modify a host configuration set while working on a particular
project, that set becomes available for all projects.

When an Autodyn analysis is started using the command script autodyn150, the host configuration
set defined in the parallel.cfg file is read, applfile is generated automatically and mpirun is
started (similar as is done on Windows operating systems).

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
of ANSYS, Inc. and its subsidiaries and affiliates. 11
 Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
12 of ANSYS, Inc. and its subsidiaries and affiliates.
Chapter 4: Creating Domain Decomposition Sets
To enable parallel processing the problem must be decomposed and the various sub-domains assigned
to the available processors. Autodyn offers two ways of decomposition generation:

Manual
The user manually creates domain decomposition sets that are associated with a particular model, and
these are stored along with all other project data whenever the model is saved. Only Structured solvers
can be decomposed manually.

Fully automatic
This option is only available for models using unstructured solvers or a combination of unstructured
solvers and Euler. Once this option is chosen, no further input is required by the user in terms of model
setup.

The following sections give examples of how to setup different types of decompositions. Files containing
the initial project setup can be found in the samples directory of the Workbench installation which is
by default C:\Program Files\ANSYS Inc\v150\aisol\Samples\AUTODYN.
4.1. Decomposing Structured Calculations
4.2. Decomposing SPH Calculations
4.3. Decomposing Unstructured Calculations
4.4. Decomposing Euler-Lagrange Coupled Problems
4.5. Fully Automatic Decomposition on Windows

4.1. Decomposing Structured Calculations

In the following example, you will create a domain decomposition set for a Lagrange impact benchmark
calculation. Load in the model B3D2P2. Though this example specifically applies to a Lagrange process,
the method shown is appropriate to decompose any grid calculation, including Euler processes.

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
of ANSYS, Inc. and its subsidiaries and affiliates. 13
Creating Domain Decomposition Sets

The problem consists of a single Lagrange part (49x100x10 cells) filled with two materials. The steel has
an initial z-velocity of 0.06 cm/microsecond, while the aluminum is initially at rest.

If you run this problem for 100 cycles and then view a velocity vector plot, you will see the screen
above, showing the impact of the steel on the aluminum.

This model already contains Decomposition Set information. The current decomposition is shown in
the Decomposition Set text box.

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
14 of ANSYS, Inc. and its subsidiaries and affiliates.
 Decomposing Structured Calculations

Presently the current decomposition set is set to None. The current decomposition set can be changed
by using the drop-down list activated by pressing the list button on the right of the text box. The image
below shows a close up of the Parallel panel.

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
of ANSYS, Inc. and its subsidiaries and affiliates. 15
Creating Domain Decomposition Sets

Choose the decomposition set denoted by TWO. Further buttons allowing the modification of the de-
composition set will appear.

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
16 of ANSYS, Inc. and its subsidiaries and affiliates.
 Decomposing Structured Calculations

Select the check box Plot parallel decomposition. The decomposition of the model will now be displayed
(you may need to manually refresh the screen if automatic refresh is not activated):

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
of ANSYS, Inc. and its subsidiaries and affiliates. 17
Creating Domain Decomposition Sets

This image shows the decomposition of the model into its constitutive sub-domains. The scale illustrates
which sub-domain is assigned to which task. Here the bottom sub-domain is assigned to task 1 and
the top sub-domain assigned to task 2. This is the optimum decomposition for efficient parallel com-
puting over two tasks. The model has been decomposed in the J direction. As the J direction contains
the most elements, this decomposition will result in the smallest amount of information being passed
between tasks.

There are 490 elements on each sub-domain boundary; 49 in the I direction and 10 in the K direction.
If we had decomposed the part in the I direction, there would have been 1000 elements on each sub-
domain boundary; 100 in the J direction and 10 in the K direction. Obviously there is a computational
overhead associated with the amount of information passed between tasks and therefore each element
on the sub-domain boundaries increases the amount of communicated information and therefore
computational expense.

Before proceeding click Delete to discard the TWO decomposition. The instructions to enable you to
replace it are detailed below.

Click Add in the Decomposition Set section of the Parallel panel.

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
18 of ANSYS, Inc. and its subsidiaries and affiliates.
 Decomposing Structured Calculations

Additional buttons and information will then appear in the Parallel panel:

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
of ANSYS, Inc. and its subsidiaries and affiliates. 19
Creating Domain Decomposition Sets

Type TWO in the Selected decomposition text box. This will be the name of the decomposition set
that will be defined. This name is also displayed in the Current text box:

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
20 of ANSYS, Inc. and its subsidiaries and affiliates.
 Decomposing Structured Calculations

Click Subdomains to activate a popup window in which the decomposition information can be entered.
The following window will appear:

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
of ANSYS, Inc. and its subsidiaries and affiliates. 21
Creating Domain Decomposition Sets

As we wish to decompose the part into two sub-domains in the J direction, enter 2 in the Number of
J-sections text box. Autodyn will automatically evaluate the optimum dividing plane for the decompos-
ition and enter it in the Dividing plane #1 text box:

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
22 of ANSYS, Inc. and its subsidiaries and affiliates.
 Decomposing Structured Calculations

You may alter the J value of the dividing plane by changing the number within the text box. There may
be situations where this would be advantageous, but this would involve more complicated models.

Click OK to store this information.

In this case the divisions can be chosen such that the resulting two sub-domains are identical in size.
This is the optimum way to set up a parallel processing calculation for excellent load-balancing and
most efficient execution.

This may not always be possible. You may have multiple parts, each of different dimensions, making it
more difficult to decompose each part so that one sub-domain of equal size is created for each processor.

• For the current host configuration, if you have two parts of differing size, it may be possible to subdivide
each part into two equal size sub-domains, so that you have a total of four sub-domains. These sub-domains
could then be allocated one to each processor (one from each part) to allow perfect load balancing. There
would be a small amount of overhead, due to processing two sub-domains instead of one on each pro-
cessor, but this is likely to be a small price to pay for perfect load-balancing.

• Taking this principal a step further, for multiple parts that do not offer very uniform decomposition, you
can decompose the parts into many more sub-domains than there are processors and let Autodyn auto-
matically load-balance the sub-domains over all processors (a discussion of automatic load-balancing fol-
lows). The more sub-domains Autodyn has to distribute, the easier it will be to effectively load-balance a
problem. However, the more sub-domains created, the more overhead will be involved to exchange data
at sub-domain boundaries. You may have to experiment to find a good balance, but our experience so
far indicates that parallel calculation are generally computation intensive, so it is likely that good load-
balancing is more important than minimizing data communication between sub-domains. Most of our
benchmarking has been on well load-balanced problems with minimal communication overhead, so we
cannot offer more detailed guidelines at present.

The part has now been decomposed into sub-domains. Though, if you view the parallel decomposition
you will see that both of the sub-domains have been automatically allocated to one task. Therefore,
we need to assign these sub-domains to the available tasks in order to produce an efficient calculation.
This process is initiated be clicking Task Assignments. The following popup window will appear:

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
of ANSYS, Inc. and its subsidiaries and affiliates. 23
Creating Domain Decomposition Sets

The sub-domains can be assigned manually or automatically. If the Automatic option button is selected
then the sub-domains will be assigned at run time. If the model is a very simple one then this course
of action could be taken. Here we will manually assign the sub-domains to the available tasks in order
to become familiar with the process.

We wish to assign the two J-sections of the model to two different tasks. To do this, enter 1 into the
to J-Section text box and click Assign Now:

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
24 of ANSYS, Inc. and its subsidiaries and affiliates.
 Decomposing Structured Calculations

These actions have assigned the first J-section sub-domain to the first task.

To assign the second J-section sub-domain to the second task enter 2 into the From J-section, to J-
section, and Assign to task text boxes and click Assign Now.

If you have automatic refresh activated the image of the model on screen should have been updated
to show the two sub-domains on two different tasks:

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
of ANSYS, Inc. and its subsidiaries and affiliates. 25
Creating Domain Decomposition Sets

We could have balanced the task assignments automatically and viewed the results instantly by selecting
the Balance automatically option button in the Task Assignment window. This results in a change
to the configuration of the window:

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
26 of ANSYS, Inc. and its subsidiaries and affiliates.
 Decomposing SPH Calculations

By clicking Balance Now, the task assignments will be immediately processed automatically. If you view
the domain decomposition image then you will see that the decomposition is identical to that which
we defined manually.

The model is now ready to be run in parallel. Click Run to execute the calculation.

4.2. Decomposing SPH Calculations

Decomposing SPH parts adopts a similar approach to that applied for grid calculations, though there
is no manual option to decompose the problem.

Load in the SPH problem SPHTAY from the samples directory within the Autodyn distribution.

This is an iron Taylor Test problem where the deformation of a cylinder is studied after impacting a wall
at 221 ms-1. The model consists of a single SPH object made up of 18540 SPH nodes. Running the
model to 0.06 ms will produce the following results:

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
of ANSYS, Inc. and its subsidiaries and affiliates. 27
Creating Domain Decomposition Sets

To decompose the model, click the Parallel button on the Setup toolbar and activate parallel processing
by selecting the appropriate check box.

Following the same method as when decomposing the Lagrange calculation; click Add in the Decom-
position Sets section of the Parallel Processing panel. Again enter TWO as the name of the decompos-
ition and click Sub-domains, which will result in the following window being displayed.

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
28 of ANSYS, Inc. and its subsidiaries and affiliates.
 Decomposing SPH Calculations

Enter 2 in the Number of Task for SPH text box and click the check button.

Now, click Task Assignments button on the main Parallel Processing panel. The following window
will be displayed.

By default, the Automatic option button will be selected. By leaving the default setting, the sub-domains
would be assigned to tasks at runtime. As we wish to view the decomposition, select the Manual option
button, which will result in further options being displayed.

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
of ANSYS, Inc. and its subsidiaries and affiliates. 29
Creating Domain Decomposition Sets

If we wished to assign the entire SPH object to one task, which may be a viable option if the model
contained multiple parts, we would enter the task number in the Assign to task text box and click
Assign Now button. As the model consists of a single SPH object we need to decompose the object
onto the two available processors in order to efficiently compute the results.

Select the Balance automatically option button and the available options will be modified.

The number of tasks will automatically be placed in the Number of tasks text box and therefore we
only have to click Balance Now in order to assign the SPH nodes to the different tasks. Autodyn will
attempt to assign the nodes in the most efficient configuration possible for parallel processing. After
doing this, click the check button to change the focus back to the main Autodyn window. Select the
Plot parallel decomposition check box and the decomposition will be displayed.

The model can be now efficiently run as a parallel computation.

4.3. Decomposing Unstructured Calculations

The decomposition of unstructured parts is applied automatically in a similar method to the decompos-
ition of SPH grids.

Load in CYRAIL from the samples directory within the Workbench distribution.

Convert this structured model into unstructured parts by selecting the Import > Convert IJK Parts to
Unstructured menu item and selecting all of the available options:

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
30 of ANSYS, Inc. and its subsidiaries and affiliates.
 Decomposing Unstructured Calculations

This makes two unstructured parts called VOLUME 1 and VOLUME 2. Rename these parts to TARG and
CYLINDER in the parts menu.

Go to the Parallel panel and tick the box to activate parallel processing. Add a decomposition set and
call it TWO:

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
of ANSYS, Inc. and its subsidiaries and affiliates. 31
Creating Domain Decomposition Sets

You will see that there are two sections within the Decomposition Sets area of the menu (Manual and
Automatic). For unstructured parts, the second section should be used. Clicking Set Decompositions
activates a window in which the information regarding the decomposition can be entered:

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
32 of ANSYS, Inc. and its subsidiaries and affiliates.
 Decomposing Unstructured Calculations

All the unstructured parts within the model will be displayed in the Parts to be decomposed window
and all are highlighted by default. To decompose the model over two tasks (task 1 and task 2) enter 2
into the Number of Tasks box and click Decompose Now:

Click the check button.

Select the Plot parallel decomposition option in the Parallel Processing window if you want to see
the parallel decomposition.

If you want to decompose the part TARG in Task 2, click Set Decomposition again and highlight only
TARG, enter 2 into the Start Task box, and click Decompose Now:

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
of ANSYS, Inc. and its subsidiaries and affiliates. 33
Creating Domain Decomposition Sets

In a similar manner the model can be decomposed over 4 slaves:

and over 8 slaves:

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
34 of ANSYS, Inc. and its subsidiaries and affiliates.
 Decomposing Euler-Lagrange Coupled Problems

4.4. Decomposing Euler-Lagrange Coupled Problems

In the following mine blast example, you will create a domain decomposition set for an analysis where
an Euler Ideal Gas parts interacts with a vehicle structure and soil through Euler-Lagrange coupling.

The decomposition of models that include Euler parts interacting with unstructured parts can be applied
automatically.

Load in the model MINPAR from the samples directory within the Workbench distribution.

The model consists of an Euler Ideal Gas part (50x20x40) filled with air and two structural parts that
model the soil (29.000 Tetra elements) and vehicle (1650 shell elements). The Euler part is filled with
high pressurized gas in the region of the soil part where the mine is located and the vehicle will be hit
by the blast wave and eroded nodes of the ejected soil. The blast wave interaction is calculated through
Euler-Lagrange Interaction and the interaction of the ejected soil and vehicle is modeled through Lag-
range-Lagrange interaction and erosion.

If this problem is run for 2 ms and a combined material/velocity vector plot is selected, you will see the
following view:

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
of ANSYS, Inc. and its subsidiaries and affiliates. 35
Creating Domain Decomposition Sets

Go to the Parallel Processing panel and select the box to activate parallel processing. Add a decom-
position set and call it TWO:

Euler in combination with unstructured parts can automatically be decomposed. Click Set Decomposition.
This activates a window in which the information regarding the decomposition can be entered:

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
36 of ANSYS, Inc. and its subsidiaries and affiliates.
 Decomposing Euler-Lagrange Coupled Problems

The unstructured parts within the model and structured Euler Ideal Gas part will be shown in the Parts
to be decomposed window and all are highlighted by default. To decompose the model over two tasks
(task 1 and task 2) simply enter 2 into the Number of Tasks box and click Decompose Now:

Click the check button. Select Plot parallel decomposition in the Parallel Processing window to see
the following automatic decomposition:

The decomposition algorithm automatically produces a decomposition configuration with good load
balancing qualities and minimal inter-processor communication.

To further enhance the efficiency of the coupled calculation the sub-domains of the FE structure are
placed on the same processor as the Euler Ideal Gas sub-domains located in the same geometric space.
This decreases the necessary inter-processor communication for the coupling calculations, as can be
seen below.

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
of ANSYS, Inc. and its subsidiaries and affiliates. 37
Creating Domain Decomposition Sets

In a similar manner the model can be decomposed over 4 slaves:

Again the sub-domains of the FE structure are placed on the same processor as the Euler Ideal Gas sub-
domains located in the same geometric space to decrease the necessary inter-processor communication
for the coupling calculations, as can be seen above.

4.5. Fully Automatic Decomposition on Windows

In a Windows environment, as an even simpler (and sometimes more efficient) alternative to the
method outlined in Decomposing Euler-Lagrange Coupled Problems (p. 35), Autodyn offers a fully
automatic decomposition option for models containing any combination of Euler and unstructured
solvers.

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
38 of ANSYS, Inc. and its subsidiaries and affiliates.
 Fully Automatic Decomposition on Windows

When using this fully automatic option there is no further input required by the user in terms of model
set-up. On execution the model is automatically decomposed over the number of tasks specified in the
Platform MPI appl file. At any point after the model has been initialized the resultant decomposition
can be viewed using the Plot parallel decomposition option also found in the Parallel Processing
panel.

Load in the model MINPAR from the samples directory within the Workbench distribution.

Select the Fully automatic Decomposition option and run the analysis for 1 cycle with 2 CPU’s defined
in the Platform MPI applfile file. The automatic domain decomposition that has been used can be
seen in the picture below.

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
of ANSYS, Inc. and its subsidiaries and affiliates. 39
 Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
40 of ANSYS, Inc. and its subsidiaries and affiliates.
Appendix A. Overview of Parallelized Features and Known
Limitations
Feature Available Limitation
in Paral-
lel?
Solvers Structured Yes
Unstructured Yes
SPH Yes
NBS Tet Yes
Rigid Bodies Yes
Euler Ideal Gas Yes
Euler Multi-Mater- Yes
ial
Contact Gap Yes Edge on edge impact is not supported.
Trajectory Yes Analyses using trajectory contact cannot
be run in parallel if there is an SPH part
or if the Decomposition Response
method is selected.
Connections Bonded/Break- Yes
able
Spotwelds Yes
Reinforcement Yes
Beams
Joins Structured/Un- Yes • Although the joins between unstruc-
structured tured parts have been parallelized it
is strongly advised to avoid the use
of unstructured parts in combination
with joins in parallel analyses.

• An option is available that will merge

joined unstructured nodes that reside
at the same physical location in the
model into one single unstructured
node. The option is available under
the Join menu and will increase ro-
bustness in many applications in-
volving joins.

SPH-Structured Yes
Euler Ideal Gas Yes

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
of ANSYS, Inc. and its subsidiaries and affiliates. 41
Overview of Parallelized Features and Known Limitations

Feature Available Limitation

in Paral-
lel?
Euler Multi-Mater- Yes Joins cannot be used in combination
ial with Euler/Lagrange coupling.
Euler/Lag- Euler Ideal Gas Yes
rangeInterac- Euler Multi-Mater- Yes Joined Euler cannot be used in combin-
tion ial ation with Euler/Lagrange coupling.
Boundary Stress/Velo- Yes
Conditions city/Bend-
ing

Flow/Trans-
mit
Analytical Blast No
Remote Displace- Yes
ment
Materials Autodyn Materi- Yes
als
Sesame Library No

Release 15.0 - © SAS IP, Inc. All rights reserved. - Contains proprietary and confidential information
42 of ANSYS, Inc. and its subsidiaries and affiliates.