It is important that all MPI ranks carry the same PYTHONHASHSEED.
python destroys member variables in a random order determined by the hashseed. If the seed is different it is possible some MPI enabled objects goes out of scope earlier than others. This is actually a MPI related issue and shall be noted in mpi4py.
I see a strange stalling with SGI's MPT and anaconda 3.5 due to this on Plaedias
@nickhand -- I suggest you mention this in the nbodykit documentation.
It is important that all MPI ranks carry the same PYTHONHASHSEED.
python destroys member variables in a random order determined by the hashseed. If the seed is different it is possible some MPI enabled objects goes out of scope earlier than others. This is actually a MPI related issue and shall be noted in mpi4py.
I see a strange stalling with SGI's MPT and anaconda 3.5 due to this on Plaedias
@nickhand -- I suggest you mention this in the nbodykit documentation.