Description
MPI 4.0 has introduced new info keys that allow applications to signal the non-use of some MPI P2P semantics that otherwise might impose unwanted overhead, namely "mpi_assert_allow_overtaking", "mpi_assert_no_any_tag","mpi_assert_no_any_source", and "mpi_assert_exact_length". We can make use of the first two info keys in our application but I have not found any trace of them in the pml/ucx (there seems to be support in pml/ob1). I am wondering what it would take to add support to pml/ucx but I couldn't find any way to signal such assertions to UCX. I guess "mpi_assert_allow_overtaking" could allow the use of thread-private endpoints (similar to osc/ucx) to mitigate the overheads of multi-threaded message ordering? Are there extensions planned in UCX?