SALSA

Twister: A Runtime for Iterative MapReduce

Jaliya EkanayakeCommunity Grids Laboratory,

Digital Science Center

Pervasive Technology Institute

Indiana University

HPDC – 2010 MAPREDUCE’10 Workshop, Chicago, 06/22/2010

SALSA

Acknowledgements to:

• Co authors:Hui Li, Binging Shang, Thilina GunarathneSeung-Hee Bae, Judy Qiu, Geoffrey FoxSchool of Informatics and Computing Indiana University Bloomington

• Team at IU

SALSA

MotivationData

DelugeMapReduce Classic Parallel

Runtimes (MPI)

Experiencing in many domains

Data Centered, QoS Efficient and Proven techniques

Input

Output

map

Inputmap

reduce

Inputmap

reduce

iterations

Pij

Expand the Applicability of MapReduce to more classes of Applications

Map-Only MapReduceIterative MapReduce

More Extensions

SALSA

Features of Existing Architectures(1)

• Programming Model– MapReduce (Optionally “map-only”)– Focus on Single Step MapReduce computations (DryadLINQ supports

more than one stage)

• Input and Output Handling– Distributed data access (HDFS in Hadoop, Sector in Sphere, and shared

directories in Dryad)– Outputs normally goes to the distributed file systems

• Intermediate data– Transferred via file systems (Local disk-> HTTP -> local disk in Hadoop)– Easy to support fault tolerance– Considerably high latencies

Google, Apache Hadoop, Sector/Sphere, Dryad/DryadLINQ (DAG based)

SALSA

Features of Existing Architectures(2)

• Scheduling– A master schedules tasks to slaves depending on the availability – Dynamic Scheduling in Hadoop, static scheduling in Dryad/DryadLINQ– Naturally load balancing

• Fault Tolerance– Data flows through disks->channels->disks– A master keeps track of the data products– Re-execution of failed or slow tasks– Overheads are justifiable for large single step MapReduce computations– Iterative MapReduce

SALSA

A Programming Model for Iterative MapReduce

• Distributed data access

• In-memory MapReduce

• Distinction on static data and variable data (data flow vs. δ flow)

• Cacheable map/reduce tasks (long running tasks)

• Combine operation

• Support fast intermediate data transfers

Reduce (Key, List<Value>)

Iterate

Map(Key, Value)

Combine (Map<Key,Value>)

User Program

Close()

Configure()Staticdata

δ flow

Twister Constraints for Side Effect Free map/reduce tasks

Computation Complexity >> Complexity of Size of the Mutant Data (State)

SALSA

Twister Programming ModelconfigureMaps(..)

Two configuration options :1. Using local disks (only for maps)2. Using pub-sub bus

configureReduce(..)

runMapReduce(..)

while(condition){

} //end while

updateCondition()

close()

User program’s process space

Combine() operation

Reduce()

Map()

Worker Nodes

Communications/data transfers via the pub-sub broker network

Iterations

May send <Key,Value> pairs directly

Local Disk

Cacheable map/reduce tasks

SALSA

Twister API1.configureMaps(PartitionFile partitionFile)

2.configureMaps(Value[] values)

3.configureReduce(Value[] values)

4.runMapReduce()

5.runMapReduce(KeyValue[] keyValues)

6.runMapReduceBCast(Value value)

7.map(MapOutputCollector collector, Key key, Value val)

8.reduce(ReduceOutputCollector collector, Key

key,List<Value> values)

9.combine(Map<Key, Value> keyValues)

SALSA

Twister Architecture

Worker Node

Local Disk

Worker Pool

Twister Daemon

Master Node

Twister Driver

Main Program

B

BB

B

Pub/sub Broker Network

Worker Node

Local Disk

Worker Pool

Twister Daemon

Scripts perform:Data distribution, data collection, and partition file creation

map

reduce Cacheable tasks

One broker serves several Twister daemons

SALSA

Input/Output Handling

• Data Manipulation Tool:– Provides basic functionality to manipulate data across the local disks

of the compute nodes

– Data partitions are assumed to be files (Contrast to fixed sized blocks in Hadoop)

– Supported commands:• mkdir, rmdir, put,putall,get,ls,

• Copy resources

• Create Partition File

Node 0 Node 1 Node n

A common directory in local disks of individual nodese.g. /tmp/twister_data

Data Manipulation Tool

Partition File

SALSA

Partition File

• Partition file allows duplicates• One data partition may reside in multiple nodes• In an event of failure, the duplicates are used to re-schedule

the tasks

File No Node IP Daemon No File partition path

4 156.56.104.96 2 /home/jaliya/data/mds/GD-4D-23.bin

5 156.56.104.96 2 /home/jaliya/data/mds/GD-4D-0.bin

6 156.56.104.96 2 /home/jaliya/data/mds/GD-4D-27.bin

7 156.56.104.96 2 /home/jaliya/data/mds/GD-4D-20.bin

8 156.56.104.97 4 /home/jaliya/data/mds/GD-4D-23.bin

9 156.56.104.97 4 /home/jaliya/data/mds/GD-4D-25.bin

10 156.56.104.97 4 /home/jaliya/data/mds/GD-4D-18.bin

11 156.56.104.97 4 /home/jaliya/data/mds/GD-4D-15.bin

SALSA

The use of pub/sub messaging• Intermediate data transferred via the broker network• Network of brokers used for load balancing

– Different broker topologies• Interspersed computation and data transfer minimizes

large message load at the brokers• Currently supports

– NaradaBrokering– ActiveMQ

100 map tasks, 10 workers in 10 nodes

Reduce()

map task queues

Map workers

Broker networkE.g.

~ 10 tasks are producing outputs at once

SALSA

Scheduling• Twister supports long running tasks• Avoids unnecessary initializations in each

iteration• Tasks are scheduled statically– Supports task reuse– May lead to inefficient resources utilization

• Expect user to randomize data distributions to minimize the processing skews due to any skewness in data

SALSA

Fault Tolerance• Recover at iteration boundaries• Does not handle individual task failures• Assumptions:– Broker network is reliable– Main program & Twister Driver has no failures

• Any failures (hardware/daemons) result the following fault handling sequence– Terminate currently running tasks (remove from memory)– Poll for currently available worker nodes (& daemons)– Configure map/reduce using static data (re-assign data

partitions to tasks depending on the data locality)– Re-execute the failed iteration

SALSA

Performance Evaluation• Hardware Configurations

• We use the academic release of DryadLINQ, Apache Hadoop version 0.20.2, and Twister for our performance comparisons.

• Both Twister and Hadoop use JDK (64 bit) version 1.6.0_18, while DryadLINQ and MPI uses Microsoft .NET version 3.5.

Cluster ID Cluster-I Cluster-II# nodes 32 230# CPUs in each node 6 2

# Cores in each CPU 8 4

Total CPU cores 768 1840Supported OSs Linux (Red Hat Enterprise Linux

Server release 5.4 -64 bit) Windows (Windows Server 2008 -64 bit)

Red Hat Enterprise Linux Server release 5.4 -64 bit

SALSA

Pair wise Sequence Comparison using Smith Waterman Gotoh

• Typical MapReduce computation• Comparable efficiencies• Twister performs the best

SALSA

Pagerank – An Iterative MapReduce Algorithm

• Well-known pagerank algorithm [1]• Used ClueWeb09 [2] (1TB in size) from CMU• Reuse of map tasks and faster communication pays off[1] Pagerank Algorithm, http://en.wikipedia.org/wiki/PageRank[2] ClueWeb09 Data Set, http://boston.lti.cs.cmu.edu/Data/clueweb09/

M

R

Current Page ranks (Compressed)

Partial Adjacency Matrix

Partial Updates

CPartially merged Updates

Iterations

SALSA

Multi-dimensional Scaling

• Maps high dimensional data to lower dimensions (typically 2D or 3D)• SMACOF (Scaling by Majorizing of COmplicated Function)[1]

[1] J. de Leeuw, "Applications of convex analysis to multidimensional scaling," Recent Developments in Statistics, pp. 133-145, 1977.

While(condition){ <X> = [A] [B] <C> C = CalcStress(<X>)}

While(condition){ <T> = MapReduce1([B],<C>) <X> = MapReduce2([A],<T>) C = MapReduce3(<X>)}

SALSA

Conclusions & Future Work

• Twister extends the MapReduce to iterative algorithms• Several iterative algorithms we have implemented

– K-Means Clustering– Pagerank– Matrix Multiplication– Multi dimensional scaling (MDS)– Breadth First Search

• Integrating a distributed file system• Programming with side effects yet support fault

tolerance

SALSA

Related Work• General MapReduce References:– Google MapReduce– Apache Hadoop– Microsoft DryadLINQ– Pregel : Large-scale graph computing at Google– Sector/Sphere– All-Pairs– SAGA: MapReduce– Disco

SALSASALSA

Questions?

Thank you!

SALSA

Extra Slides

SALSA

Hadoop (Google) Architecture

• HDFS stores blocks, manages replications, handle failures• Map/reduce are Java processes, not long running• Failed maps are re-executed, failed reducers collect data from maps again

HDFS

M

Local

R

Task Tracker

Job Tracker

Map output goes to local disk first

Task Tracker

Local

Map task readsInput data from HDFS

Task tracker notifies job tracker

Job tracker assigns some map outputs to a reducer

Reducer downloads map outputs using HTTP

Reduce output goes to HDFS

1

2

3

4

56

SALSA

Twister Architecture

• Scripts for file manipulations• Twister daemon is a process, but Map/Reduce tasks are Java Threads

(Hybrid approach)

M

Local

R

TwisterDaemon

Map output goes directly to reducer

TwisterDaemon

Local

Reduce output goes to local disk OR to Combiner1

3

4

Read static data from local disk

1

B

B B B

Broker Connection

Receive static data (1)ORVariable data (key,value)via the brokers (2)

4

2

Broker Network

TwisterDriver

Main program1. configureMaps(PartitionFile

partitionFile)

2. configureMaps(Value[] values)

3. configureReduce(Value[] values)

4. String key=addToMemCache(Value

value)

5. removeFromMemCache(String key)

6. runMapReduce()

7. runMapReduce(KeyValue[]

keyValues)

8. runMapReduceBCast(Value value)

Twister API

SALSA

Twister• In-memory MapReduce• Distinction on static data and

variable data (data flow vs. δ flow)

• Cacheable map/reduce tasks (long running tasks)

• Combine operation• Support fast intermediate data

transfers

Different synchronization and intercommunication mechanisms used by the parallel runtimes

Reduce (Key, List<Value>)

Iterate

Map(Key, Value)

Combine (Key, List<Value>)

User Program

Close()

Configure()Staticdata

δ flow

SALSA

Publications1. Jaliya Ekanayake, (Advisor: Geoffrey Fox) Architecture and Performance of Runtime Environments

for Data Intensive Scalable Computing, Accepted for the Doctoral Showcase, SuperComputing2009.

2. Xiaohong Qiu, Jaliya Ekanayake, Scott Beason, Thilina Gunarathne, Geoffrey Fox, Roger Barga, Dennis Gannon, Cloud Technologies for Bioinformatics Applications, Accepted for publication in 2nd ACM Workshop on Many-Task Computing on Grids and Supercomputers, SuperComputing2009.

3. Jaliya Ekanayake, Atilla Soner Balkir, Thilina Gunarathne, Geoffrey Fox, Christophe Poulain, Nelson Araujo, Roger Barga, DryadLINQ for Scientific Analyses, Accepted for publication in Fifth IEEE International Conference on e-Science (eScience2009), Oxford, UK.

4. Jaliya Ekanayake and Geoffrey Fox, High Performance Parallel Computing with Clouds and Cloud Technologies, First International Conference on Cloud Computing (CloudComp2009), Munich, Germany. – An extended version of this paper goes to a book chapter.

5. Geoffrey Fox, Seung-Hee Bae, Jaliya Ekanayake, Xiaohong Qiu, and Huapeng Yuan, Parallel Data Mining from Multicore to Cloudy Grids, High Performance Computing and Grids workshop, 2008. – An extended version of this paper goes to a book chapter.

6. Jaliya Ekanayake, Shrideep Pallickara, Geoffrey Fox, MapReduce for Data Intensive Scientific Analyses, Fourth IEEE International Conference on eScience, 2008, pp.277-284.

7. Jaliya Ekanayake, Shrideep Pallickara, and Geoffrey Fox, A collaborative framework for scientific data analysis and visualization, Collaborative Technologies and Systems(CTS08), 2008, pp. 339-346.

8. Shrideep Pallickara, Jaliya Ekanayake and Geoffrey Fox, A Scalable Approach for the Secure and Authorized Tracking of the Availability of Entities in Distributed Systems, 21st IEEE International Parallel & Distributed Processing Symposium (IPDPS 2007).