[torqueusers] forcing use of one processor per node?

Glen Beane glen.beane at gmail.com
Wed Apr 16 10:37:01 MDT 2008


On Wed, Apr 16, 2008 at 11:52 AM, Jim Coyle <jjc at iastate.edu> wrote:

> John,
>
> This is a function of the launcher.


no, it's really a function of the job scheduler


>
>
> I use mpirun and create a different machinefile.
> sort -u $PBSNODEFILE > ./tmp1
> cat tmp1 tmp1  > tmp2
> mpirun -np 4 -machinefile tmp2 ...
>

using a machine file with torque is a bad idea.  You should be using a TM
based job launcher (OpenMPI and LAM have TM support,  OSC's mpiexec provides
it for mpich), and you should be submitting your job in such a way that the
scheduler knows to schedule one task per node

as for the original poster:  you can configure moab or maui to honor your
explicit node request rather than pack multiple ppn=1 requests onto one
node.  I would try adding pmem to your resource request though, like brock
requested
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://www.supercluster.org/pipermail/torqueusers/attachments/20080416/9912c0a2/attachment.html


More information about the torqueusers mailing list