[mvapich-discuss] MVAPICH2 with OpenMP

Martin Cuma martin.cuma at utah.edu
Wed Mar 2 14:41:24 EST 2011


Thanks Ben,

this did the trick. I was thinking that since OpenMPI did not seem to need 
this, MVAPICH2 would not need it too.

Sayantan, perhaps it would be good to stress this out in your otherwise 
very good user's guide? There's not much specifics about the Qlogic PSM, 
which is quite different from the Mellanox, at least from an user 
perspective.

Also, your OpenMP-MPI test code you sent works right even without the 
IPATH_NO_CPUAFFINITY - that is, it shows multiple threads per process, 
however, critical to me was to see 400% CPU load per MPI process in my 
program (ran as 1 process per socket on a quad-core CPU), which I only 
achieved with the IPATH_NO_CPUAFFINITY=1.

Thanks for all your help.

MC

On Wed, 2 Mar 2011, Ben Truscott wrote:

> Dear Martin
>
> I'm working with a configuration very much like yours. I wonder, did you 
> remember to set IPATH_NO_CPUAFFINITY=1? Without this, the PSM library makes 
> its own affinity settings that pre-empt those made by OpenMP.
>
> Regards
>
> Ben Truscott
> School of Chemistry
> University of Bristol
>
>>  Hi all,
>>
>>  I am having a strange problem with MVAPICH2 - the ones that I build don't
>>  seem to run multi-threaded programs with OpenMP. I suspect it's some kind
>>  of a configuration issue on my end as I tried to run MVAPICH2 from the
>>  stock OFED distribution and the multithreading there is fine.
>>
>>  Also, I have been building OpenMPI with fairly standard options and the
>>  OpenMP there works fine.
>>
>>  Does someone have similar experience, or, is there some trick with
>>  MVAPICH2 build or running to get multiple OpenMP threads going?
>>
>>  I configure as:
>>  configure --enable-romio --with-file-system=nfs+ufs  --with-device=ch3:psm
>>  with Intel or GNU compilers.
>>
>>  Then run as:
>>  mpiexec.hydra -genv OMP_NUM_THREADS 4 -genv MV2_ENABLE_AFFINITY 0 -n 16
>>  -machinefile nodefile ./program
>>
>>  This is all on RHEL 5.5 with the Qlogic or Mellanox IB adapters.
>>
>>  Thanks,
>>  MC
>>
>>  --
>>  Martin Cuma
>>  Center for High Performance Computing
>>  University of Utah
>
>

-- 
Martin Cuma
Center for High Performance Computing
University of Utah


More information about the mvapich-discuss mailing list