[mpich-devel] detecting mpirun mismatch

[Dave Goodell (dgoodell)]

On Apr 22, 2014, at 4:58 PM, Jeff Hammond <jeff.science at gmail.com> wrote:

> PMI is specific to MPICH-family, right?  The problem that was observed
> had to do with MPICH/OpenMPI mixing.  'mpirun -np 2' led to two copies
> of the application running in serial...

I'm pretty sure that Open MPI can speak some flavors of PMI.  I don't have any personal experience with using that functionality though, so it may not be very automagical.  It might require some special build flags or run-time config.  So keep that in mind when considering solutions.

Also, don't forget that the "singleton init" case is supported by both OMPI and MPICH, so the absence of PMI_* vars isn't sufficient for you to decide that you were launched by some other MPI implementation's process manager.  You actually have to check for other common variables from other launchers / MPI implementations and warn that you know you cannot handle them.


I feel like we've had this conversation before on the MPICH lists and decided there was no good solution for this problem... hah, here it is:

http://lists.mcs.anl.gov/pipermail/mpich-discuss/2012-February/011756.html

I completely concede that Jed's correct here in his response (modulo the "Open MPI is system-installed on Macs", which is no longer true for modern Macs):

http://lists.mcs.anl.gov/pipermail/mpich-discuss/2012-February/011757.html

-Dave