Ralph Castain <rhc_at_[hidden]> writes:
> On Apr 25, 2013, at 5:33 PM, Vladimir Yamshchikov <yaximik_at_[hidden]> wrote:
>> $NSLOTS is what requested by -pe openmpi <ARG> in the script, my understanding that by default it is threads.
Is there something in the documentation
<http://arc.liv.ac.uk/SGE/htmlman/manuals.html> that suggests that? [It
currently incorrectly says processes, rather than slots, in at least one
place I'll fix.]
> What you want to do is:
> 1. request a number of slots = the number of application processes * the number of threads each process will run
[If really necessary, maybe use a job submission verifier to fiddle what
the user supplies.]
> 2. execute mpirun with the --cpus-per-proc N option, where N = the number of threads each process will run.
> This will ensure you have one core for each thread. Note, however,
> that we don't actually bind a thread to the core - so having more
> threads than there are cores on a socket can cause a thread to bounce
> across sockets and (therefore) potentially across NUMA regions.
Does that mean that binding is suppressed in that case, as opposed to
binding N cores per process, which is what I thought it did? (I can't
immediately test it.)
I don't understand the problem in this specific case which causes
over-subscription. However, if the program's runtime needs instruction,
you can do things like setting OMP_NUM_THREADS with an SGE JSV; see
archives of the gridengine list. (The SGE_BINDING variable that recent
SGE provides to the job can be converted to GOMP_CPU_AFFINITY etc., but
that's probably only useful for single-process jobs.)
There may be a case for OMPI to support this sort of thing for DRMs like
SGE which don't start the MPI processes themselves; you potentially need
to export the binding information per-process.
Community Grid Engine: http://arc.liv.ac.uk/SGE/