[MPICH] debug flag

[Anthony Chan]

It is possible that the Cray has its own process manager that yod is only
way to interact with the process manager and probably the only one knows
the nodes allocated.  If that is the case, you could try this in your PBS
script.

> cat <your_pbs_script>
#!/bin/sh
...
#PBS -l nodes=<num_of_nodes>
...
YOD_NODEFILE=<my_node_file>
rm -f ${YOD_NODEFILE}
yod -size <num_of_nodes> hostname > ${YOD_NODEFILE}
...
mpdboot -n <num_of_nodes> -f ${YOD_NODEFILE}
...
mpiexec -n <num_of_nodes> a.out



Not sure if this works, may worth a try.

A.Chan

On Sat, 26 May 2007, Wei-keng Liao wrote:

>
> The machine is a Cray XT3, 4. The batch is PBS and the command to launch
> an MPI job is "yod" as described in the user guide.
>
> I tried to replace yod with mpich mpdboot and mpiexec, etc. I also used
> the PBS environment variable PBS_NODEFILE in mpdboot, but the node file
> does not contain the hostnames of nodes allocated by the PBS. It actually
> always contains just one yod node, no matter how many nodes I request in
> PBS script.
>
> Any idea of how to do it?
>
> Wei-keng
>
>
> On Sat, 26 May 2007, Rajeev Thakur wrote:
>
> > Are you sure you can't run your own MPICH2 on the machine? It is just a
> > user-level library. Once the scheduler assigns you a set of nodes, you can
> > run mpd and your own mpiexec.
> >
> > Rajeev
> >
>
>