MPI failures using mpipool.MPIPool about kombine

[adrn]

Some more information:

• The MPI implementation used on the cluster is MVAPICH2
• Here is a minimal bit of code that produces the error.
• Interestingly, using the MPI pool implementation from this thread does not produce this error. See this code for an example.

[adrn]

I just tested on a different cluster that uses openMPI v1.6.5 and I see the same behavior.

[juliohm]

@adrn , I don't remember what exactly I did, but you can see that I used this MPI pool class instead: https://github.com/juliohm/HUM/blob/master/pyhum/utils.py#L24

Good luck.

[bfarr]

I haven't been able to look into this too deeply yet, but I've managed to reproduce UnpicklingErrors locally.

If I run without load balancing it seems to go away, and it looks like the MPI pool class pointed to by @juliohm doesn't do any load balancing either.

@adrn, do you get UnpicklingError with load balancing off? If emcee works fine for you with the same configuration, it would be surprising if it were a problem with the load balancing...