[mpich-discuss] error in mpiexec

[Dave Goodell]

This topic/suggestion has come up before, but it's harder to do in practice than you might think.

The primary problem is that the "singleton init" case has no mpiexec, so the check can't be quite as simple as just "assert some environment variable contains the right magic cookie value".  At best we could try to detect some of the Open MPI environment variables and the OMPI guys could try to detect some of our environment variables.  It would never really be a robust solution.

-Dave

On Feb 3, 2012, at 12:16 PM CST, Gustavo Correa wrote:

> 
> On Feb 2, 2012, at 11:08 PM, Jed Brown wrote:
> 
>> On Thu, Feb 2, 2012 at 21:30, Jeff Hammond <jhammond at alcf.anl.gov> wrote:
>> (This question belongs on mpich-discuss at mcs.anl.gov)
>> 
>> It seems that you are using mpiexec from OpenMPI with a binary compiled with MPICH2.  This is not a good idea.
>> 
>> Please use the mpiexec from MPICH2.  If you do not know where it is, try 'find /home/mpich2-1.4.1p1/ -name mpiexec'.
>> 
>> You can see what mpiexec you are using by calling 'which mpiexec'.
>> 
>> This problem is way too common. Why doesn't mpiexec put some token in the environment that MPI_Init() can check and error if it doesn't match (including version, if necessary)? Same with Open MPI. The error message could even helpfully point out the location of the correct mpiexec.
> 
> 1+ on that.
> Excellent idea.  
> Thanks Jed.
> 
> For newbies, and also for those who need to keep various versions and flavors of MPI,
> this would be at least a minimal safety net.
> 
> However, some caution must be exerted to allow the deliberate use of compatible mpiexec
> releases.  For instance, we use OSC mpiexec with MPICH2, it works well,  and we don't
> want to preclude its use.
> 
> May thanks,
> Gus Correa
> 
>> _______________________________________________
>> mpich-discuss mailing list     mpich-discuss at mcs.anl.gov
>> To manage subscription options or unsubscribe:
>> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss
> 
> _______________________________________________
> mpich-discuss mailing list     mpich-discuss at mcs.anl.gov
> To manage subscription options or unsubscribe:
> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss