Rajeev Thakur <[email protected]> writes:

It doesn't have to be a matching Wait. The implementation is required to make progress on other pending communication while it is blocked in a Wait.

Or blocked in any of a number of other places, if non-blocking collective requests have the same completion semantics as point-to-point requests. E.g., this shouldn't deadlock: MPI_Iallreduce(...,COMM_WORLD,&req); if (!rank) MPI_Wait(&req); MPI_Allgather(...,COMM_WORLD); if (rank) MPI_Wait(&req);