Cloud Programming Models: eScience, Big Data, etc.

IN4392 Cloud Computing
Cloud Programming Models

Cloud Computing (IN4392)

D.H.J. Epema and A. Iosup. With contributions from Claudio Martella and Bogdan Ghit.

2012-2013

1

Parallel and Distributed Group/TU Delft

Terms for Today’s Discussion
Programming model
= language + libraries + runtime system that create
a model of computation (an abstract machine)
= ―an abstraction of a computer system‖ Wikipedia
Examples: message-passing vs shared memory, data- vs
task-parallelism, …

Abstraction level Q: What is the best abstraction level?
= distance from physical machine
Examples: Assembly low-level vs Java is high level
Many design trade-offs: performance, ease-of-use,
common-task optimization, programming paradigm, …
2012-2013 2

Characteristics of a Cloud Programming Model

1. Cost model (Efficiency) = cost/performance, overheads, …
2. Scalability
3. Fault-tolerance
4. Support for specific services
5. Control model, e.g., fine-grained many-task scheduling
6. Data model, including partitioning and placement, out-of-
memory data access, etc.
7. Synchronization model

2012-2013 3

Agenda

1. Introduction
2. Cloud Programming in Practice (The Problem)
3. Programming Models for Compute-Intensive Workloads
4. Programming Models for Big Data
5. Summary

2012-2013 4

Today’s Challenges

• eScience
• The Fourth Paradigm
• The Data Deluge and Big Data
• Possibly others

2012-2013 5

eScience: The Why
• Science experiments already cost 25—50% budget
• … and perhaps incur 75% of the delays
• Millions of lines of code with similar functionality
• Little code reuse across projects and application domains
• … but last two decades‘ science is very similar in structure

• Most results difficult to share and reuse
• Case-in-point: Sloan Digital Sky Survey
digital map of 25% of the sky x spectra
40TB+ sky survey data
200M+ astro-objects (images)
1M+ objects with spectrum (spectra)
How to make it work for this and
the next generation of scientists?
2012-2013 6

Source: Jim Gray and Alex Szalay,―eScience -- A Transformed Scientific Method‖,
http://research.microsoft.com/en-us/um/people/gray/talks/NRC-CSTB_eScience.ppt

eScience (John Taylor, UK Sci.Tech., 1999)
• A new scientific method
• Combine science with IT
• Full scientific process: control scientific instrument or produce data
from simulations, gather and reduce data, analyze and model
results, visualize results
• Mostly compute-intensive, e.g., simulation of complex phenomena

• IT support
• Infrastructure: LHC Grid, Open Science Grid, DAS, NorduGrid, …
• From programming models to infrastructure management tools

• Examples Q: Why is CompSci an example here?
• * physics, Bioinformatics, Material science, Engineering, CompSci
2012-2013 7

The Fourth Paradigm: The Why (An Anecdotal Example)
The Overwhelming Growth of Knowledge
―When 12 men founded the Number of 1993 1997
Royal Society in 1660, it was Publications 1997 2001
possible for an educated
person to encompass all of
scientific knowledge. […] In
the last 50 years, such has
been the pace of scientific
advance that even the best
scientists cannot keep up
with discoveries at frontiers
outside their own field.‖
Tony Blair,
PM Speech, May 2002 Data: King,The scientific impact of nations,Nature‘04.

The Fourth Paradigm: The What
From Hypothesis to Data
• Thousand years ago:
science was empirical describing natural phenomena
2

• Last few hundred years:
.
a 4 G c2

theoretical branch using models, generalizations a 3 a2

• Last few decades:
a computational branch simulating complex phenomena
• Today (the Fourth Paradigm): Q1: What is the Fourth
data exploration Paradigm?
unify theory, experiment, and simulation
• Data captured by instruments or generated by simulator
• Processed by software
Q2: What are the dangers
• Information/Knowledge stored the Fourth Paradigm?
of in computer
• Scientist analyzes results using data management and statistics
2012-2013 9

Source: Jim Gray and ―The Fourth Paradigm‖,
http://research.microsoft.com/en-us/collaboration/fourthparadigm/

The “Data Deluge”:
The Why

"Everywhere you look, the quantity
of information in the world is
soaring. According to one
estimate, mankind created 150
exabytes (billion gigabytes) of
data in 2005. This year, it will
create 1,200 exabytes. Merely
keeping up with this flood, and
storing the bits that might be
useful, is difficult enough.
Analysing it, to spot patterns and
extract useful information, is
harder still.―
The Data Deluge, The Economist,
25 February 2010
10

“Data Deluge”: The Personal Memex Example

• Vannevar Bush in the 1940s: record your life
• MIT Media Laboratory: The Human Speechome
Project/TotalRecall, data mining/analysis/visio
• Deb Roy and Rupal Patel ―record practically every
waking moment of their son‘s first three years‖
(20% privacy time…Is this even legal?! Should it be?!)
• 11x1MP/14fps cameras, 14x16b-48KHz mics, 4.4TB
RAID + tapes, 10 computers; 200k hours audio-video
• Data size: 200GB/day, 1.5PB total

11

“Data Deluge”: The Gaming Analytics Example

• EQ II: 20TB/year all logs
• Halo3: 1.4PB served statistics
on player logs

12

“Data Deluge”: Datasets in Comp.Sci.

Dataset
Size
http://gwa.ewi.tudelft.nl
1TB/yr
The Failure
Trace 1TB GamTA
Archive
100GB P2PTA
http://fta.inria.fr
10GB

Peer-to-Peer Trace Archive 1GB
… PWA, ITA, CRAWDAD, …
‗06 ‗09 ‗10 ‗11 Year

• 1,000s of scientists: From theory to practice
13

13

“Data Deluge”:
The Professional World Gets Connected

Feb 2012
2012-2013 14

Source: Vincenzo Cosenza, The State of LinkedIn,
http://vincos.it/the-state-of-linkedin/

What is “Big Data”?

• Very large, distributed aggregations of loosely structured
data, often incomplete and inaccessible

• Easily exceeds the processing capacity of conventional
database systems

• Principle of Big Data: ―When you can, keep everything!‖

• Too big, too fast, and doesn‘t comply with the traditional
database architectures
2011-2012 15

The Three “V”s of Big Data

• Volume
• More data vs. better models
• Data grows exponentially
• Analysis in near-real time to extract value
• Scalable storage and distributed queries
• Velocity
• Speed of the feedback loop
• Gain competitive advantage: fast recommendations
• Identify fraud, predict customer churn faster
• Variety
• The data can become messy: text, video, audio, etc.
• Difficult to integrate into applications
2011-2012 16

Adapted from: Doug Laney, ―3D data management‖, META Group/Gartner report,
Feb 2001. http://blogs.gartner.com/doug-laney/files/2012/01/ad949-3D-Data-
Management-Controlling-Data-Volume-Velocity-and-Variety.pdf

Agenda

1. Introduction
3. Programming Models for Compute-Intensive
Workloads
1. Bags of Tasks
2. Workflows
3. Parallel Programming Models
5. Summary

2012-2013 18

What is a Bag of Tasks (BoT)? A System View

BoT = set of jobs sent by a user…

…that start at most Δs after
the first job Q: What is the user‘s view?

Time [units]
• Why Bag of Tasks? From the perspective
of the user, jobs in set are just tasks of a larger job
• A single useful result from the complete BoT
• Result can be combination of all tasks, or a selection
of the results of most or even a single task
2012-2013 19
Iosup et al., The Characteristics and
Performance of Groups of Jobs in Grids,
Euro-Par, LNCS, vol.4641, pp. 382-393, 2007.

Applications of the BoT Programming Model

• Parameter sweeps
• Comprehensive, possibly exhaustive investigation of a model
• Very useful in engineering and simulation-based science

• Monte Carlo simulations
• Simulation with random elements: fixed time yet limited inaccuracy
• Very useful in engineering and simulation-based science

• Many other types of batch processing
• Periodic computation, Cycle scavenging
• Very useful to automate operations and reduce waste
2012-2013 20

BoTs Became the Dominant Programming
Model for Grid Computing
(US) TeraGrid-2 NCSA
(US) Condor U.Wisc.
(EU) EGEE
(CA) SHARCNET
(US) Grid3
(US) GLOW
(UK) RAL
(NO,SE) NorduGrid
(FR) Grid'5000
(NL) DAS-2

From jobs [%] 0 20 40 60 80 100

(US) TeraGrid-2 NCSA
(US) Condor U.Wisc.
(EU) EGEE
(CA) SHARCNET
(US) Grid3
(US) GLOW
(UK) RAL
(NO,SE) NorduGrid
(FR) Grid'5000
(NL) DAS-2

From CPUTime [%]
0 20 40 60 80 100

2012-2013 21

Iosup and Epema: Grid Computing Workloads.
IEEE Internet Computing 15(2): 19-26 (2011)

Practical Applications of the BoT Programming Model
Parameter Sweeps in Condor [1/4]

• Sue the scientist wants to
―Find the value of F(x,y,z) for
10 values for x and y, and 6 values for z‖

• Solution: Run a parameter sweep, with
10 x 10 x 6 = 600 parameter values
• Problem of the solution:
• Sue runs one job (a combination of x, y, and z) on her low-end
machine. It takes 6 hours.
• That‘s 150 days uninterrupted computation on Sue‘s machine!

2012-2013 22

Source: Condor Team, Condor User‘s Tutorial.
http://cs.uwisc.edu/condor

Universe = vanilla
Executable = sim.exe
Input = input.txt
Output = output.txt
Error = error.txt
Log = sim.log

Requirements = OpSys == “WINNT61” && Complex SLAs can
Arch == “INTEL” && be specified easily
(Disk >= DiskUsage) && ((Memory * 1024)>=ImageSize)

InitialDir = run_$(Process) Also passed as
Queue 600 parameter to sim.exe
2012-2013 23


% condor_submit sim.submit
Submitting job(s)
................................................
................................................
................................................
................................................
................................................
...............
Logging submit event(s)
................................................
................................................
................................................
................................................
................................................
...............
600 job(s) submitted to cluster 3.

2012-2013 24


% condor_q
-- Submitter: x.cs.wisc.edu : <128.105.121.53:510> :
x.cs.wisc.edu
ID OWNER SUBMITTED RUN_TIME ST PRI SIZE CMD
3.0 frieda 4/20 12:08 0+00:00:05 R 0 9.8 sim.exe
3.1 frieda 4/20 12:08 0+00:00:03 I 0 9.8 sim.exe
3.2 frieda 4/20 12:08 0+00:00:01 I 0 9.8 sim.exe
3.3 frieda 4/20 12:08 0+00:00:00 I 0 9.8 sim.exe
...
3.598 frieda 4/20 12:08 0+00:00:00 I 0 9.8 sim.exe
3.599 frieda 4/20 12:08 0+00:00:00 I 0 9.8 sim.exe

600 jobs; 599 idle, 1 running, 0 held

2012-2013 25


Agenda

1. Introduction
Workloads
1. Bags of Tasks
2. Workflows
5. Summary

2012-2013 26

What is a Wokflow?

WF = set of jobs with
precedence
(think Direct Acyclic Graph)

2012-2013 27

Applications of the Workflow Programming
Model
• Complex applications
• Complex filtering of data
• Complex analysis of instrument measurements

• Applications created by non-CS scientists*
• Workflows have a natural correspondence in the real-world,
as descriptions of a scientific procedure
• Visual model of a graph sometimes easier to program

• Precursor of the MapReduce Programming Model
(next slides)
2012-2013 28

*Adapted from: Carole Goble and David de Roure, Chapter in ―The Fourth
Paradigm‖, http://research.microsoft.com/en-us/collaboration/fourthparadigm/

Workflows Existed in Grids, but Did Not
Become a Dominant Programming Model
• Traces

• Selected Findings

• Loose coupling
• Graph with 3-4 levels
• Average WF size is 30/44 jobs
• 75%+ WFs are sized 40 jobs or less, 95% are sized 200 jobs or less
2012-2013 29
Ostermann et al., On the Characteristics of Grid
Workflows, CoreGRID Integrated Research in Grid
Computing (CGIW), 2008.

Practical Applications of the WF Programming Model
Bioinformatics in Taverna

2012-2013 30

Source: Carole Goble and David de Roure, Chapter in ―The Fourth Paradigm‖,
http://research.microsoft.com/en-us/collaboration/fourthparadigm/

Agenda

1. Introduction
Workloads
1. Bags of Tasks
2. Workflows
5. Summary

2012-2013 31

Parallel Programming Models TUD course in4049
Introduction to HPC

• Abstract machines
Task (groups of 5, 5 minutes):
• (Distributed) shared memory
discuss parallel programming in clouds
• Distributed memory: MPI
Task (inter-group discussion):
discuss.
• Conceptual programming models
• Master/worker
• Divide and conquer
• Data / Task parallelism
• BSP
• System-level programming models
• Threads on GPUs and other multi-cores
2012-2013 32

Varbanescu et al.: Towards an Effective Unified
Programming Model for Many-Cores. IPDPS WS 2012

Agenda

1. Introduction
5. Summary

2012-2013 33

Ecosystems of Big-Data Programming Models
Q: Where does MR-on-demand fit?
Q: Where does Pregel-on-GPUs fit?
High-Level Language
Flume BigQuery SQL Meteor JAQL Hive Pig Sawzall Scope DryadLINQ AQL

Programming Model
PACT MapReduce Model Pregel Dataflow Algebrix

Execution Engine
Flume Dremel Tera Azure Nephele Haloop Hadoop/ Giraph MPI/ Dryad Hyracks
Engine Service Data Engine YARN Erlang
Tree Engine

Storage Engine
S3 GFS Tera Azure HDFS Voldemort L CosmosFS Asterix
Data Data F B-tree
Store Store * Plus Zookeeper, CDN, etc. S
2012-2013 34

Adapted from: Dagstuhl Seminar on Information Management in the Cloud,
http://www.dagstuhl.de/program/calendar/partlist/?semnr=11321&SUOG

Agenda

1. Introduction
1. MapReduce
2. Graph Processing
3. Other Big Data Programming Models
5. Summary

2012-2013 35

MapReduce
• Model for processing and generating large data sets

• Enables a functional-like programming model

• Splits computations into independent parallel tasks

• Makes efficient use of large commodity clusters

• Hides the details of parallelization, fault-tolerance,
data distribution, monitoring and load balancing

2011-2012 36

MapReduce: The Programming Model

A programmig model, not a programming language!
1. Input/Output:
• Set of key/value pairs
2. Map Phase:
• Processes input key/value pair
• Produces set of intermediate pairs
map (in_key, in_value) -> list(out_key, interm_value)
3. Reduce Phase:
• Combines all intermediate values for a given key
• Produces a set of merged output values
reduce(out_key, list(interm_value)) -> list(out_value)

2011-2012 37

MapReduce in the Cloud

• Facebook use case:
• Social-networking services
• Analyze connections in the graph of friendships to recommend new
connections
• Google use case:
• Web-base email services, GoogleDocs
• Analyze messages and user behavior to optimize ad selection and
placement
• Youtube use case:
• Video-sharing sites
• Analyze user preferences to give better stream suggestions

2011-2012 38

Wordcount Example
• File 1: ―The big data is big.‖
• File 2: ―MapReduce tames big data.‖

• Map Output:
• Mapper-1: (The, 1), (big, 1), (data, 1), (is, 1), (big, 1)
• Mapper-2: (MapReduce, 1), (tames, 1), (big, 1), (data, 1)

• Reduce Input • Reduce Output
• Reducer-1: (The, 1) • Reducer-1: (The, 1)
• Reducer-2: (big, 1), (big, 1), (big, 1) • Reducer-2: (big, 3)
• Reducer-3: (data, 1), (data, 1) • Reducer-3: (data, 2)
• Reducer-4: (is, 1) • Reducer-4: (is, 1)
• Reducer-5: (MapReduce, 1), (MapReduce, 1) • Reducer-5: (MapReduce, 2)
• Reducer-6: (tames, 1) • Reducer-6: (tames, 1)
2011-2012 39

Colored Square Counter

2011-2012 40

MapReduce @ Google: Example 1

• Generating language model statistics
• Count # of times every 5-word sequence occurs in large corpus
of documents (and keep all those where count >= 4)

• MapReduce solution:
• Map: extract 5-word sequences => count from document
• Reduce: combine counts, and keep if count large enough

2011-2012 41

• http://www.slideshare.net/jhammerb/mapreduce-pact06-keynote

MapReduce @ Google: Example 2

• Joining with other data
• Generate per-doc summary, but include per-host summary
(e.g., # of pages on host, important terms on host)

• MapReduce solution:
• Map: extract host name from URL, lookup per-host info,
combine with per-doc data and emit
• Reduce: identity function (emit key/value directly)

2011-2012 42

• http://www.slideshare.net/jhammerb/mapreduce-pact06-keynote

More Examples of MapReduce Applications

• Distributed Grep
• Count of URL Access Frequency
• Reverse Web-Link Graph
• Term-Vector per Host
• Distributed Sort
• Inverted indices
• Page ranking
• Machine learning algorithms
• …

2011-2012 43

Execution Overview (1/2)
Q: What is the performance problem raised by this step?

2011-2012 44

Execution Overview (2/2)
• One master, many workers example
• Input data split into M map tasks (64 / 128 MB): 200,000 maps
• Reduce phase partitioned into R reduce tasks: 4,000 reduces
• Dynamically assign tasks to workers: 2,000 workers

• Master assigns each map / reduce task to an idle worker
• Map worker:
• Data locality awareness
• Read input and generate R local files with key/value pairs
• Reduce worker:
• Read intermediate output from mappers
• Sort and reduce to produce the output
2011-2012 45

Failures and Back-up Tasks

2011-2012 46

What is Hadoop? A MapReduce Exec. Engine
• Inspired by Google, supported by Yahoo!

• Designed to perform fast and reliable analysis of the big data

• Large expansion in many domains such as:
• Finance, technology, telecom, media, entertainment, government,
research institutions

2011-2012 47

Hadoop @ Yahoo

• When you visit
yahoo, you are
interacting with
data processed
with Hadoop!

2011-2012 48

• https://opencirrus.org/system/files/OpenCirrusHadoop2009.ppt

Hadoop Distributed File System

• Assumptions and goals:
• Data distributed across hundreds or thousands of machines
• Detection of faults and automatic recovery
• Designed for batch processing vs. interactive use
• High throughput of data access vs. low latency of data access
• High aggregate bandwidth and high scalability
• Write-once-read-many file access model
• Moving computations is cheaper than moving data: minimize network
congestion and increase throughput of the system

2011-2012 50

HDFS Architecture
• Master/slave architecture

• NameNode (NN)
• Manages the file system namespace
• Regulates access to files by clients
• Open/close/rename files or directories
• Mapping of blocks to DataNodes
• DataNode (DN)
• One per node in the cluster
• Manages local storage of the node
• Block creation/deletion/replication initiated by NN
• Serve read/write requests from clients
2011-2012 51

HDFS Internals
• Replica Placement Policy
• First replica on one node in the local rack
• Second replica on a different node in the local rack
• Third replica on a different node in a different rack
improved write performance (2/3 are on the local rack)
preserves data reliability and read performance

• Communication Protocols
• Layered on top of TCP/IP
• Client Protocol: client – NameNode machine
• DataNode Protocol: DataNodes – NameNode
• NameNode responds to RPC requests issued by DataNodes / clients

2011-2012 52

Hadoop Scheduler
• Job divided into several independent tasks executed in parallel
• The input file is split into chunks of 64 / 128 MB
• Each chunk is assigned to a map task
• Reduce task aggregate the output of the map tasks

• The master assigns tasks to the workers in FIFO order
• JobTracker: maintains a queue of running jobs, the states of the
TaskTrackers, the tasks assignments
• TaskTrackers: report their state via a heartbeat mechanism

• Data Locality: execute tasks close to their data
• Speculative execution: re-launch slow tasks
2011-2012 53

MapReduce Evolution [1/5]
Hadoop is Maturing: Important Contributors
Number of patches

2012-2013 54

Source: http://www.theregister.co.uk/2012/08/17/community_hadoop/

Late Scheduler
• LATE Scheduler
• Longest Approximate Time to End
• Speculatively execute the task that will finish farthest into the future
• time_left = (1-progressScore) / progressRate
• Tasks make progress at a roughly constant rate
• Robust to node heterogeneity

• EC2 Sort running times
• LATE vs. Hadoop vs. No spec.

2011-2012 55

Zaharia et al.: Improving MapReduce performance in
heterogeneous environments. OSDI 2008.

FAIR Scheduling
• Isolation and statistical multiplexing
• Two-level architecture:
• First, allocates task slots across pools
• Second, each pool allocates its slots among multiple jobs
• Based on a max-min fairness policy

2011-2012 56

Zaharia et al.: Delay scheduling: a simple technique for
achieving locality and fairness in cluster
scheduling. EuroSys 2010. Also TR EECS-2009-55

Delay Scheduling
• Data locality issues:
• Head-of-line scheduling: FIFO, FAIR
• Low probability for small jobs to achieve data locality
• 58% of jobs @ FACEBOOK have < 25 maps
• Only 5% achieve node locality vs. 59% rack locality

2011-2012 57


Delay Scheduling
• Solution:
• Skip the head-of-line job until you find a job that can launch a local task
• Wait T1 seconds before launching tasks on the same rack
• Wait T2 seconds before launching tasks off-rack
• T1 = T2 = 15 seconds => 80% data locality

2011-2012 58


KOALA and MapReduce [1/9]
• KOALA • MR-Runner
• Placement & allocation • Configuration & deployment
• Central for all MR clusters • MR cluster monitoring
• Maintains MR cluster metadata • Grow/Shrink mechanism

• On-demand MR clusters
• Performance isolation
• Data isolation
• Failure isolation
• Version isolation

2011-2012 59

Koala and MapReduce [2/9]
Resizing Mechanism
• Two types of nodes:
• Core nodes: fully-functional nodes, with TaskTracker and DataNode (local disk access)
• Transient nodes: compute nodes, with TaskTracker

• Parameters
• F = Number of running tasks per number of available slots
• Predefined Fmin and Fmax thresholds
• Predefined constant step growStep / shrinkStep
• T = time elapsed between two successive resource offers

• Three policies:
• Grow-Shrink Policy (GSP): grow-shrink but maintain F between Fmin and Fmax
• Greedy-Grow Policy (GGP): grow, shrink when workload done
• Greedy-Grow-with-Data Policy (GGDP): GGP with core nodes (local disk access)
2011-2012 60

Workloads

• 98% of jobs @ Facebook process 6.9 MB and take less than a minute [1]

• Google reported in 2004 computations with TB of data on 1000s of machines [2]

[1] Y. Chen, S. Alspaugh, D. Borthakur, and R. Katz, ―Energy Efficiency for Large-Scale MapReduce Workloads with Significant Interactive Analysis,‖
pp. 43–56, 2012.
[2] J. Dean and S. Ghemawat, ―Mapreduce: Simplified Data Processing on Large Clusters,‖ Comm. of the ACM, Vol. 51, no. 1, pp. 107–113, 2008.

2011-2012 61

Wordcount (Type 0, full system)

CPU Disk

• 100 GB input data
• 10 core nodes with 8 map slots each
• 800 map tasks executed in 7 waves
• Wordcount is CPU-bound in the map phase

2011-2012 62

Sort (Type 0, full system)
CPU Disk

• Performed by the MR framework during the shuffling phase
• Intermediate key/value pairs are processed in increasing key order
• Short map phase with 40%-60% CPU utilization
• Long reduce phase which is highly disk intensive

2011-2012 63

Speedup (full system)

Optimum

a) Wordcount (Type 1) b) Sort (Type 2)

• Speedup relative to an MR cluster with 10 core nodes
• Close to linear speedup on core nodes

2011-2012 64

Execution Time vs. Transient Nodes

• Input data of 40 GB
• Wordcount output data = 20 KB
• Sort output data = 40 GB
• Wordcount scales better than Sort on transient nodes
2011-2012 65

Performance of the Resizing Mechanism

Fmin=0.25
Fmax=1.25
growStep=5
shrinkStep=2
TGSP =30 s
TGG(D)P =120 s

• Stream of 50 MR jobs
• MR cluster of 20 core nodes + 20 transient nodes
• GGP increases the size of the data transferred across the network
• GSP grows/shrinks based on the resource utilization of the cluster
• GGDP enables local writes on the disks of provisioned nodes
2011-2012 66

Use Case @ TUDelft
• Goal: Process 10s TB of P2P traces

• DAS-4 constraints:
• Time limit per job (15 min)
• Non-persistent storage

• Solution:
• Reserve the nodes for several days, import and process the data?

• Our approach:
• Split the data into multiple subsets
• Smaller data sets => faster import and processing
• Setup multiple MR clusters, one for each subset
2011-2012 67

Agenda

1. Introduction
1. MapReduce
2. Graph Processing
5. Summary

2012-2013 68

Graph Processing Example
Single-Source Shortest Path (SSSP)

• Dijkstra‘s algorithm
• Select node with minimal distance
• Update neighbors
• O(|E| + |V|·log|V|) with FiboHeap

• Initial dataset:
A: <0, (B, 5), (D, 3)>
B: <inf, (E, 1)>
C: <inf, (F, 5)>
D: <inf, (B, 1), (C, 3), (E, 4), (F, 4)>
...
2012-2013 69

Source: Claudio Martella, Presentation on Giraph at TU Delft, Apr 2012.

Graph Processing Example
SSSP in MapReduce
Q: What is the performance problem?
• Mapper output: distances
<B, 5>, <D, 3>, <C, inf>, ...
• But also graph structure
<A, <0, (B, 5), (D, 3)> ...

• Reducer input: distances
B: <inf, 5>, D: <inf, 3> ...
• But also graph structure
N jobs, where N is the B: <inf, (E, 1)> ...
graph diameter
2012-2013 70


The Pregel
Programming Model for Graph Processing

• Batch-oriented processing
• Runs in-memory
• Vertex-centric API
• Fault-tolerant
• Runs on Master-Slave architecture

• OK, the actual model follows in the next slides
2012-2013 71


The Pregel
Programming Model for Graph Processing
Processors
Local Computation
Based on Valiant‘s
Bulk Synchronous Parallel (BSP)
• N processing units with fast local memory
• Shared communication medium
• Series of Supersteps
• Global Synchronization Barrier Communication
• Ends when all voteToHalt

• Pregel executes initialization, one
or several supersteps, shutdown Barrier
Synchronization
2012-2013 72


Pregel
The Superstep

• Each Vertex (execution in parallel)
• Receive messages from other vertices
• Perform own computation (user-defined function)
• Modify own state or state of outgoing messages
• Mutate topology of the graph
• Send messages to other vertices
• Termination condition
• All vertices inactive
• All messages have been transmitted

2012-2013 73

Source: Pregel article.

Pregel: The Vertex-Based API

Input
message
Implements
processing
algorithm

Output
message

2012-2013 74


The Pregel Architecture
Master-Worker
• Master assigns vertices to Workers
• Graph partitioning Master
• Master coordinates Supersteps
• Master coordinates Checkpoints
Worker Worker
• Workers execute vertices compute()
1 n
• Workers exchange messages directly
5
1 2

4
3 7
k

2012-2013 75


Apache Giraph
An Open-Source Implementation of Pregel
Tasktracker Tasktracker Tasktracker Tasktracker
Map Map Map Map Map Map Map Map
Slot Slot Slot Slot Slot Slot Slot Slot

Persistent
Zookeeper Master
computation state NN & JT
• Loose implementation of Pregel
• Strong community (Facebook, Twitter, LinkedIn)
• Runs 100% on existing Hadoop clusters
• Single Map-only job
2012-2013 78

http://incubator.apache.org/giraph/

Page rank benchmarks
• Tiberium Tan
• Almost 4000 nodes, shared among numerous groups in Yahoo!
• Hadoop 0.20.204 (secure Hadoop)
• 2x Quad Core 2.4GHz, 24 GB RAM, 1x 6TB HD

• org.apache.giraph.benchmark.PageRankBenchmark
• Generates data, number of edges, number of vertices, # of
supersteps
• 1 master/ZooKeeper
• 20 supersteps
• No checkpoints
• 1 random edge per vertex

79

Source: Avery Ching presentation at HortonWorks.
http://www.slideshare.net/averyching/20111014hortonworks/

Worker scalability (250M vertices)
5000
4500
4000
3500
Total seconds

3000
2500
2000
1500
1000
500
0
0 50 100 150 200 250 300 350
# of workers
80


Vertex scalability (300 workers)
2500

2000
Total seconds

1500

1000

500

0
0 100 200 300 400 500 600
# of vertices (in 100's of millions)
81


Vertex/worker scalability
2500 600

# of vertices (100's of millions)
500
2000
Total seconds

400
1500
300
1000
200

500
100

0 0
0 50 100 150 200 250 300 350
# of workers
82


Agenda

1. Introduction
1. MapReduce
2. Graph Processing
5. Summary

2012-2013 83

Stratosphere

• Meteor query language, Supremo operator framework

• Programming Contracts (PACTs) programming model
• Extended set of 2nd order functions (vs MapReduce)
• Declarative definition of data parallelism

• Nephele execution engine
• Schedules multiple dataflows simultaneously
• Supports IaaS environments based on Amazon EC2, Eucalyptus

• HDFS storage engine

2012-2013 84

Stratosphere
Programming Contracts (PACTs) [1/2]
2nd-order function

Also in MapReduce
2012-2013 85

Source: PACT overview,
https://stratosphere.eu/wiki/doku.php/wiki:pactpm

Stratosphere
Programming Contracts (PACTs) [2/2]

Q: How can PACTs
optimize data
processing?

2012-2013 86

Source: PACT overview,
https://stratosphere.eu/wiki/doku.php/wiki:pactpm

Stratosphere
Nephele

2012-2013 87

Stratosphere vs MapReduce
• PACT extends MapReduce
• Both propose 2nd-order functions (5 PACTs vs Map & Reduce)
• Both require from user 1st-order functions (what‘s inside the Map)
• Both can benefit from higher-level languages
• PACT ecosystem has IaaS support

• Key-value
data model

2012-2013 88

Source: Fabian Hueske, Large Scale Data Analysis Beyond MapReduce, Hadoop Get
Together, Feb 2012.

Stratosphere vs MapReduce 3. Join edges to triads
Triangle Enumeration [1/3]
• Input: 2. Sort by
undirected graph smaller vertex ID
edge-based
• Output:
triangles
4. Triads to triangles
1. Read input graph

2012-2013 90

Together, Feb 2012 and Stratosphere example.


MapReduce
Formulation

2012-2013 91



Stratosphere
Formulation

2012-2013 92


Agenda

1. Introduction
5. Summary

2012-2013 93

Conclusion Take-Home Message
• Programming model = computer system abstraction

• Programming Models for Compute-Intensive Workloads
• Many trade-offs, few dominant programming models
• Models: bags of tasks, workflows, master/worker, BSP, …
• Programming Models for Big Data
• Big data programming models have ecosystems
• Many trade-offs, many programming models
• Models: MapReduce, Pregel, PACT, Dryad, …
• Execution engines: Hadoop, Koala+MR, Giraph, PACT/Nephele, Dryad, …

• Reality check: cloud programming is maturing
February 20, 2013 http://www.flickr.com/photos/dimitrisotiropoulos/4204766418/
94

Reading Material (Really Active Field)
• Workloads
• Alexandru Iosup, Dick H. J. Epema: Grid Computing Workloads. IEEE Internet Computing 15(2): 19-26 (2011)

• The Fourth Paradigm
• ―The Fourth Paradigm‖, http://research.microsoft.com/en-us/collaboration/fourthparadigm/

• Programming Models for Compute-Intensive Workloads
• Alexandru Iosup, Mathieu Jan, Omer Ozan Sonmez, Dick H. J. Epema: The Characteristics and Performance of Groups of Jobs in Grids.
Euro-Par 2007: 382-393
• Ostermann et al., On the Characteristics of Grid Workflows, CoreGRID Integrated Research in Grid Computing (CGIW), 2008.
http://www.pds.ewi.tudelft.nl/~iosup/wftraces07chars_camera.pdf
• Corina Stratan, Alexandru Iosup, Dick H. J. Epema: A performance study of grid workflow engines. GRID 2008: 25-32

• Programming Models for Big Data
• Jeffrey Dean, Sanjay Ghemawat: MapReduce: Simplified Data Processing on Large Clusters. OSDI 2004: 137-150
• Jeffrey Dean, Sanjay Ghemawat: MapReduce: a flexible data processing tool. Commun. ACM 53(1): 72-77 (2010)
• Matei Zaharia, Andy Konwinski, Anthony D. Joseph, Randy Katz, and Ion Stoica. 2008. Improving MapReduce performance in
heterogeneous environments. In Proceedings of the 8th USENIX conference on Operating systems design and implementation (OSDI'08).
USENIX Association, Berkeley, CA, USA, 29-42.
• Matei Zaharia, Dhruba Borthakur, Joydeep Sen Sarma, Khaled Elmeleegy, Scott Shenker, Ion Stoica: Delay scheduling: a simple technique
for achieving locality and fairness in cluster scheduling. EuroSys 2010: 265-278
• Tyson Condie, Neil Conway, Peter Alvaro, Joseph M. Hellerstein, Khaled Elmeleegy, Russell Sears: MapReduce Online. NSDI 2010: 313-328
• Grzegorz Malewicz, Matthew H. Austern, Aart J. C. Bik, James C. Dehnert, Ilan Horn, Naty Leiser, Grzegorz Czajkowski: Pregel: a system
for large-scale graph processing. SIGMOD Conference 2010: 135-146
• Dominic Battré, Stephan Ewen, Fabian Hueske, Odej Kao, Volker Markl, Daniel Warneke: Nephele/PACTs: a programming model and
execution framework for web-scale analytical processing. SoCC 2010: 119-130

2012-2013 95

Cloud Programming Models: eScience, Big Data, etc.

Recommended

More Related Content

What's hot (20)

Similar to Cloud Programming Models: eScience, Big Data, etc. (20)

Cloud Programming Models: eScience, Big Data, etc.

Editor's Notes