Pavan Balaji <[email protected]> writes:

SEPARATE is too expensive for such cases.

What is the lightest weight way for the source to create an ordering at the target (which will be using MPI_Win_sync)? How does the overhead of that compare to MPI_Win_sync?