It seems to work well with various codes and numbers of nodes and processes.
Again, the patch I sent is kind of nasty. It will likely not work with meshes with very badly-shaped elements.
So, what was the problem exactly?
Your MPI implementation is having some difficulties with MPI_Waitany
and/or MPI_Waitall
+ MPI_Status
queries. I’d advise trying to start with a fresh installation using Intel compilers and IntelMPI, which do not have such issues (that’s what I’m using daily). If it still fails with those, there may be something wrong with your interconnect, but it is not possible for me to solve such issues remotely.