[mvapich-discuss] [hydra] one MPI process per node (-bynode)

[Christian Krause]

Thank you, this is very helpful to know in advance.

On Wed, Dec 19, 2012 at 8:03 PM, Jeff Hammond <jhammond at alcf.anl.gov> wrote:
> Make sure you set the environment variable MV2_ENABLE_AFFINITY=0 so
> that the OpenMP threads can be spawned on different cores.
>
> I've observed catastrophic performance with MPI+Pthreads codes because
> the default settings (enabling thread affinity) of MVAPICH2 do not
> permit the threads to be spawned on different cores.
>
> Jeff
>
> On Wed, Dec 19, 2012 at 12:08 PM, Yauheni Zelenko <zelenko at cadence.com> wrote:
>> Hi, Christian!
>>
>> I think -ppn 1 in Hydra command line will solve ypur problem.
>>
>> Eugene.
>> ________________________________________
>> From: mvapich-discuss-bounces at cse.ohio-state.edu [mvapich-discuss-bounces at cse.ohio-state.edu] On Behalf Of Christian Krause [kizkizzbangbang at googlemail.com]
>> Sent: Wednesday, December 19, 2012 5:41 AM
>> To: mvapich-discuss at cse.ohio-state.edu
>> Subject: [mvapich-discuss] [hydra] one MPI process per node (-bynode)
>>
>> Hello,
>>
>> Is it possible with the hydra launcher to start one MPI process per node?
>>
>> We currently want to support hybrid MPI / OpenMP jobs at our cluster
>> for MVAPICH. We are using hydra to tightly integrate MVAPICH with the
>> Sun Grid Engine.
>>
>> We also support OpenMPI which has a "-bynode" parameter to start one
>> MPI process per node, which is really useful for starting hybrid MPI /
>> OpenMP jobs.
>>
>> Is there something like "-bynode" for the hydra (or another) launcher?
>> I've browsed through the man page but unfortunately found nothing.
>>
>> Best Regards
>> _______________________________________________
>> mvapich-discuss mailing list
>> mvapich-discuss at cse.ohio-state.edu
>> http://mail.cse.ohio-state.edu/mailman/listinfo/mvapich-discuss
>>
>> _______________________________________________
>> mvapich-discuss mailing list
>> mvapich-discuss at cse.ohio-state.edu
>> http://mail.cse.ohio-state.edu/mailman/listinfo/mvapich-discuss
>
>
>
> --
> Jeff Hammond
> Argonne Leadership Computing Facility
> University of Chicago Computation Institute
> jhammond at alcf.anl.gov / (630) 252-5381
> http://www.linkedin.com/in/jeffhammond
> https://wiki.alcf.anl.gov/parts/index.php/User:Jhammond
> _______________________________________________
> mvapich-discuss mailing list
> mvapich-discuss at cse.ohio-state.edu
> http://mail.cse.ohio-state.edu/mailman/listinfo/mvapich-discuss