I was able to run a job with this.  It looks like the mpirun script takes
the same arguments as mpiexec.

The job I ran was the NetPipe benchmark - and the results I got make it
clear that this mpich is using Ethernet - the maximum throughput with
Ethernet is almost a factor of 10 slower than with InfiniBand.

Dave

-----Original Message-----
From: Research Computing Support [mailto:[log in to unmask]] On
Behalf Of Robin
Sent: Thursday, May 11, 2006 11:04 AM
To: [log in to unmask]
Subject: MPICH-1.2.6 (GNU-Intel)

All,

	I've added support for the MPICH-1.2.6 (Ethernet - GNU and Intel).
	module: mpich-gnu-1.2.6, mpich-intel-1.2.6. Since people ask that  
before.

	It will handle clean up, etc.

	The clean up support is given by PBSPro via PBSPro's wrapper script

(/usr/pbs/bin/pbs_mpirun).
Documented at PBSPro (Chapter 10 Admin Guide). PBSPro wrapper script  
ensures that all slave processes have PBS_JOBID. There are couple  
more things that they do e.g.
pbs_attach, etc.

	The backend are still the same. I needed to modify PBSPro wrapper  
script a bit.

	PBSPro doesn't supply the wrapper script for our mpirun_ssh. That's

why we needed one to have to ensure that all slave processes
can be identified. Master processes can easily be identified by  
PBSPro when jobs are qdel'ed. It's the slave processes need to be the
extra tag.

	Here is how to run:
#!/bin/bash -l
#PBS -N mpichjob
#PBS -j oe
#PBS -q parallel
#PBS -l nodes=4:ppn=1

module add mpich-intel-1.2.6
## or module add mpich-gnu-1.2.6

mpirun <yourexecutable>

#-- end file--

	Please note that you don't need to specify the number of CPUs.  
mpirun will find it out from the requested nodes and processor.
	
Thanks,
--------------
Robin
[log in to unmask]
513-529-1483

"Academia politics is the most vicious precisely because the stake is  
so small" - Kissinger