mpiexec and PBSPro's new syntax

[Pete Wyckoff]

bfp at purdue.edu wrote on Tue, 06 May 2008 11:03 -0400:
> If I apply Matt Ford's patch to "get_hosts.c" in mpiexec-0.83, I'm able to 
> succesfully use mpiexec with PBSPro 9, as long as I request nodes using 
> the PBSPro "old style" syntax for requesting nodes, as in
> 
> qsub -I -l nodes=2:ppn=2 ...
> cat $PBS_NODEFILE
> tempest-000
> tempest-000
> tempest-001
> tempest-001
> 
> mpiexec -np 4 ./a.out
> 
> 
> However, if I use PBSPro's latest syntax for rquesting nodes, I get an 
> error from mpiexec.
> 
> qsub -I -l select=2:mpiprocs=2,place=scatter
> cat $PBS_NODEFILE
> tempest-000
> tempest-000
> tempest-001
> tempest-001
> 
> mpiexec -np 4 ./a.out
> mpiexec: Error: get_hosts: PBS reports fewer tasks 2 than TM 4.
> 
> 
> Would there by any chance be another patch available for PBSPro 
> version 9 ?

There's nothing newer than that partial patch.  The hope was we
could figure something out that would work for all versions of PBS.
If you dig around the archives a bit, you might find a request from
me to a PBSPro user to send the ouput of qstat -f in a few different
job configurations.  Mpiexec uses the Resource_List.* variables (and
could use others) to figure out what the hosts are.

A bit of explanation on what that select=... syntax means would be
useful too.  Like what mpiexec should do depending on what it finds
in the qstat -f output.

If you come up with a snazzy patch, it will be easier for me to
integrate and make sure open PBS still works.

		-- Pete