Parallel Computing using Condor on Windows PCs

1
Parallel Computing using Condor on Windows PCs
Peng Wang and Corey Shields Research and Academic
Computing Division University Information
Technology Services Indiana University
2
Problem Description

• Turn Windows desktop systems in STC labs (around
  2000) into a parallel scientific computer

3
Discussion

• Parallel applications need coordination through
  message passing
• MPI does not handle ephemeral processes well
• Multiplexing communication among processes
• Ports brokered among multiple parallel sessions

4
What do we have ?

• Condor NT, vanilla universe
• match-making, file transfer, fair sharing, job
  submission, suspension, preemption, restart,
  security
• Test application fastDNAml-p
• Parallel application, master-worker model,
  small granularity of work

5
How we did it

• Simple Message Brokering Library (SMBL)
• Process and Port Manager (PPM)
• A mechanism for users to submit jobs (web portal)

6
SMBL

• An IO multiplexing server in charge of message
  delivery for each parallel session (serialize
  communication)
• SMBL client library implements selected MPI-like
  calls
• Both the server and the client library are based
  on a TCP socket abstraction library

7
Process and Port Manager

• Assigns port to each of the SMBL server process
• start the SMBL server and application processes
  on demand
• direct workers to their servers

8
The Portal

• Apache based
• PHP web interface
• Creates and submits the condor submit files

9
The Big Picture
The shaded box indicates components hosted on
multiple desktop computers
10
Statistics
Red total owner Blue total idle Green total
Condor
11
Scalability Issues

• Needed big server
• Adjusted condor_config
• MAX_JOBS_RUNNING 1000
• SHADOW_SIZE_ESTIMATE 900KB
• MAX_STARTD_LOG 640KB
• Lost workers because of
• per-process file descriptor limit (1024)

12
Summary

• Built a large parallel scientific computing
  facility using Condor
• Built parallel message passing library to deal
  with ephemeral resources
• Built port broker to handle multiple parallel
  sessions
• Built web portal
• It is open source, visit http//smbl.sourceforge.
  net