Cluster computing, getting started

A few notes about using the research cluster to run lots of jobs. We have machines that are situated in both CS and in ECE, therefore you need an account in both CS and ECE. Please contact one of the faculty (Lebeck or Sorin) to get these accounts created and added to the group dukearch.

• First read the man page on qsub "man qsub". If it says the manpage is not found then you need to specify the path to the manpages. On the CS research machine nicl.cs.duke.edu you can type "man -M /usr/research/admin/sge/current/man qsub" and that should find the man page for qsub. For those in the know, you can add this to your $MANPATH environment variable...
  
  
• The directory that you plan to write output to from the batch system must be in either /usr/research/arch or in /usr/research/arch1 and it must be group dukearch with group write permissions. To accomplish this use the following commands:
  chgrp dukearch <directory>
  chmod g+w <directory>
  
  
• To submit jobs you must be logged into either nicl.cs.duke.edu with your CS research account or into platypus.ee.duke.edu with your ECE account.
  
  
• The command for submitting jobs is qsub. You must have a script file as an input argument to qsub, type qsub -help to see the list of available arguments. The -cwd argument is very useful, it tells qsub to use the "current working directory" as the place to search for binaries, input files and to write output files.
  
  
• Most people find it useful to use scripts to submit a large number of jobs. You can write this script in your favorite scripting languate (csh, perl, etc.) Below is a csh script that submits 10 instances of a job to the cluster.
  
  
  #!/bin/csh

	# Example to submit 10 jobs of hello world

	foreach i (0 1 2 3 4 5 6 7 8 9)
   		qsub -cwd -N hello_$i run_hello
	end

• The job submitted is specified in the script file "run_hello" which resides in the current working directory. The contents of "run_hello" are below:

	#
 	hello

• hello is a linux/x86 binary in the current working directory. For a good test, simply change hello to "date" which will simply print the date and time to standard out.