[torqueusers] Newbie torque script questions

Glen Beane glen.beane+torque at gmail.com
Thu Dec 7 09:49:47 MST 2006


On 12/6/06, dave first <linux4dave at gmail.com> wrote:
> I am such a newbie that I squeek.  I hope this is the correct forum in which
> to ask this question.
>
>  I want to specify a nodelist other than that which would be $PBS_NODEFILE.
> I want to specify n10, n11, n12 and n13, each with 4 processors.  The node
> list looks something like this:
>
> n10:4
> n11:4
>  n12:4
> n13:4
>
> And it is called local_nodelist in the working directory.
>
> The script sets PBS_NODEFILE=`pwd`/local_nodelist


I am not sure why you would want to do this.  PBS_NODEFILE is created
when torque launches your job.  If you intend to override the nodes
assigned by the scheduler,  job run time would not be the time to do
it. What would be the point of using TORQUE?  If you really need
specific nodes, follow Jerry's advice:  -l
nodes=n10:ppn=4+n11:ppn=4+n12:ppn=4+n13:ppn=4 in your PBS script will
tell torque you need those nodes, and torque will assign them to your
job (when they are free).

Also, using the standard mpirun that comes with mpich is a bad idea.
As Jerry metioned, you should use Pete Wycoff's mpiexec. If you don't
follow this advice, you are doomed to a cluster full of rogue
processes of mpi jobs that are not properly killed and cleaned up by
torque.  Also, a nice feature of mpiexec is that it allows torque to
collect cputime, etc for the entire job - this is something torque
can't do with ssh spawned mpi jobs (which is what mpirun does).


More information about the torqueusers mailing list