Distributing a Java Workflow Across a Network of Workstations

1
Trellis Driver

• Distributing a Java Workflow Across a Network of
  Workstations

Nicholas Lamb, Paul Lu, and Alona
FysheDepartment of Computing ScienceUniversity
of Albertanlambpaullualona_at_cs.ualberta.ca
2
Motivation

• Some applications in science consist of main
  job that drives other jobs

J1
Driver Job

• Main or driver jobcontrols workflow

J2
J3
J4
3
Motivation

• Languages provide functions for running external
  jobs on local server (e.g., system() )
• Javas runtime.exec() has drawback that location
  of jobs must be statically specified
• RMI offloads jobs onto remote hosts, but
• RMI is a set of low-level mechanisms
• RMI does not include scheduling
• RMI does not provide job barrier function

4
Trellis Driver Introduction

• Trellis Driver Java package schedules jobs
  across metacomputers
• Trellis Driver integrates Java applications
  with Trellis metacomputing system
• TrellisDriver.exec() accepts command line and
  sends it to underlying metascheduler
• TrellisDriver.waitForAll() provides a job
  barrier

5
Trellis Driver Implementation
Java Code
Trellis Driver
Local OS
mqsub process
Producer
Consumer
next
Trellis
For i 1 to n td.exec( blast )
Meta-queue
6
Example ApplicationProteome Analyst (PA)

• Bioinformatics tool for annotating
  proteomeshttp//www.cs.ualberta.ca/bioinfo/PA
• Training process Machine-learn new classifier
  based on set of proteins with known annotations
• Prediction process Use existing classifier to
  predict annotations of unknown proteins
• PA is best at predicting subcellular
  localization (Bioinformatics, March 2004)

7
Workflow Pipeline Shape
Sequence
Sequence
BLAST
BLAST
Homologues
Homologues
Parsing
Parsing
Features
Features
PredictionProcess
TrainingProcess
Training
Prediction
Classifier
Class Label
8
Workflow Pipeline Shape
Sequence
Sequence
Phase 151927
BLAST
BLAST
Homologues
Homologues
Phase 200637
Parsing
Parsing
Features
Features
Phase 310817
Phase 401834
Training
Prediction
Classifier
Class Label
9
Homogenous Job Batching

• Homogenous Job Batching refers to grouping
  together of multiple calls to same program in
  common metacomputer job
• Job1 Job2 Job3 Job4
• Job1Job2 Job3Job4
• Grouping multiple BLAST invocations in single
  call to mqsub greatly amortizes overhead
• Concatenation of Blast commands transparent to
  PA application

10
Parallel BLAST Phase Data Flow
Sets batching factor
td.setGroup( Blast, 2)td.exec(Blast,
cmd1)td.exec(Blast, cmd2)td.exec(Blast,
cmdN-1)td.exec(Blast, cmdN)td.waitForAl
l()
mqsub cmd1 cmd2
mqsub cmdN-1 cmdN
11
Parallel BLAST Phase Data Flow
td.setGroup( Blast, 2)td.exec(Blast,
cmd1)td.exec(Blast, cmd2)td.exec(Blast,
cmdN-1)td.exec(Blast, cmdN)td.waitForAl
l()
Multiple concurrent mqsub processes
mqsub cmd1 cmd2
mqsub cmdN-1 cmdN
12
Parallel BLAST Phase Results
13
Parallel BLAST Phase Results
14
Parallel BLAST Phase Results
15
Future Work

• Parallelize Machine Learning phase (runtime was
  more than 1 hr)
• Provide Heterogeneous Job Batching for job
  pipelines to ensure data affinity
• Test with geographically distant servers
  connected via wide-area networks (WANs)
• Underlying Trellis system is already being used
  across administrative domains and WANs (e.g.,
  CISS-3)

16
Conclusions

• Scientific applications can benefit by
  scheduling workflows across metacomputers
• Contribution Trellis Driver module provides
  convenient replacement for Runtime.exec()
• PA and Trellis Driver can obtain reasonable
  speed-ups for large BLAST phase
• Trellis Drivers overheads can be amortized by
  homogeneous job batching

17
Trellis Driver Architecture

• Job submission results in producer-consumer
  data flow pattern
• Application threads (producers) generate jobs
• Trellis Driver threads (consumers) process jobs
• Consumer count decoupled from producer count by
  bounded buffer implementation

18
Trellis Metacomputing System

• Trellis metacomputing system provides
  user-level aggregation of multiple computing
  systems
• Trellis offers load-balancing of workloads
• Metaqueues provide single point of submission
  for jobs
• mqsub add jobs to metaqueue mqdel remove
  jobs from metaqueue mqstat list metaqueue
  contents

19
Trellis Driver API
setGroup(groupName, batchFactor) Register
new job group execSynch(commandLine) Run
command synchronously exec(commandLine,
groupName, prodId) Run command
asynchronously waitForOne(key, prodId)
Await completion of single job
waitForAll(prodId) Await completion of all
jobs
20
Java Support for Workflow Parallelism

• PA originally used Runtime.exec() to run BLAST
  jobs
• Runtime.exec() provides process-level
  concurrency within single server
• Not scalable to large inputs
• RMI works only between Java objects and methods

21
Parallel BLAST Phase Code Changes

• Code changes required to parallelize BLAST phase
  minimal
• add TrellisDriver.setGroup() at beginning
• replace Runtime.exec() with
  TrellisDriver.exec()
• add TrellisDriver.waitForAll() at end

22
Original PA Performance Results

• Trained and validated new classifier with
  training set of 3,916 sequences