18
TORQUE Kerry Chang CCLS December 13, 2010

TORQUE Kerry Chang CCLS December 13, 2010. O UTLINE Torque How does it work? Architecture MADA Demo Results Problems Future Improvements

Embed Size (px)

Citation preview

TORQUEKerry Chang

CCLS

December 13, 2010

OUTLINE

Torque How does it work? Architecture

MADA Demo Results Problems Future Improvements

TORQUE – WHAT IS IT?

Open source project by Cluster Resources Inc.

Cluster resource manager Manages batch jobs

A series of programs to be executed without manual intervention

Manages distributed compute nodes Distributed servers on which to execute batch

jobs

TORQUE ARCHITECTURE

TORQUE SCHEDULER

Currently using standard built-in schedule (FIFO)

MOAB – more advanced scheduler

WHAT HAVE I DONE?

Used MADA as an application of TORQUE Treated the application as a blackbox Text parallelization on input

Created a series of scripts for text manipulation and job submission to Torque queue

Linear improvement in processing time by using Torque

MADA

System for Morphological Analysis and Disambiguation for Arabic

Input file is separated line by line

MADA ARCHITECTURE

HOW DO THE SCRIPTS WORK?

1) First split the text file evenly across the number of specified jobs to be

submitted 2) Create a script for each newly split text file

e.g. If you wanted to run 5 jobs, split the text into 5 files and create a script to run each of the 5 files.

3) Submit each script to Torque 4) Concatenate the output of each script

DEMO

Demonstration of Torque and MADA 3 Output Files

file.bw file.bw.mada file.bw.mada.tok

RESULTS

30 lines

0 2 4 6 8 10 12 14 16 18 200

50

100

150

200

250

300

350

400

With TorqueWithout Torque

RESULTS

300 Lines

0 1 2 3 4 5 6 7 8 9 100

200

400

600

800

1000

1200

1400

1600

With TorqueWithout Torque

RESULTS

3,000 Lines

0 5 10 15 20 25 30 35 400

200

400

600

800

1000

1200

1400

1600

1800

With TorqueWithout Torque

RESULTS

30,000 Lines

0 1 2 3 4 5 6 7 80

2000

4000

6000

8000

10000

12000

14000

16000

With TorqueWithout Torque

RESULTS

Network – Local Temp comparison (seconds)

Network Local Temp Improvement

114 108 6

201 193 8

1,477 1,400 77

14,544 13,973 571

145,105 131,649 13,456

PROBLEMS

How do we know when MADA has finished and we can concatenate the results?

Where do we run MADA and have the results output to?

Submission to compute node hangs Use smarter scheduler Supply machines dedicated to running Torque

jobs

FUTURE IMPROVEMENTS

Pipeline many jobs to Torque Work from local temp folders instead of on

the network Split and rebuild certain output files by

looking at provided testing.madaconfig file MADA TOKAN Preprocessor

QUESTIONS