We have encountered some latency issues with MCPL phase-space files being broacast in MPI to the different nodes at the beginning of execution (this is on AWS. on a local multi-node cluster speed is great).
I could be wrong, but I believe the base full phase-space file is sent to all nodes.
In mcnp-NAGSS (the photon transport stage of Advantg’s ORCS R2S), the NAGSS photon phase-space files are split into per OMP per node. Such that each node receives its own.
(It isn’t compiled with MPI, so it does this with multiple individual execution and then a final merge_mctal at end).
It could be an idea to develop some mechanisms into OpenMC and the phase-space writing/reading/etc. to allow the files to be split and individually distributed.