This is a biggy and likely important to customers using Open MPI with GPU accelerators. See https://github.com/mpi-forum/mpi-issues/issues/580 and entry 17 of section B.1.2 of the MPI 4.1 standard