Here is where I am at with the split periodic gravity calculation over MPI.
I can survive for some steps but I can get stuck in reproducible ways. For instance, running
mpirun -np 4 swift_mpi -s -S -c -G -t 4 eagle_12.yml -v 1 always gets
stuck on step 43.
We end up with an un-balanced number of send-recv on that step. With one node having an extra un-matched recv blocking the calculation.
Note that this is not directly after a rebuild but that it may involve cells that have not had any action performed on them since a rebuild.
I have looked a the obvious things that would prevent the task activation mechanism from making a correct symmetric decision but unsuccessfully thus far. Will come back to this in a few days when other commitments have passed. As always, any comments or suggestions are welcome.