10 13 2021

Jump to bottom Edit New page

James Dinan edited this page Oct 13, 2021 · 2 revisions

October 13, 2021

Agenda:

Message ordering in the case of multiple threads (#117)

Action items:

Jim to update Hybrid WG wiki, create missing tickets, etc.
WG members to reflect on three options below and bring preferences to next meeting.

Logically current isn't issue:

Reviewed notes posted on 09-29-2021
- Discussed two options for resolving ambiguity in message ordering with MPI_THREAD_MULTIPLE
  1. MPI must order operations in the order that it sees them.
  2. MPI can always treat operations from distinct threads as unordered.
- Consensus is trending toward option (a). If we choose this option,
  - Need a clear statement that it is a change (possibly breaking change)
  - Need to be careful that we don't introduce new serialization in semantics (e.g. must constrain to same communicator)
- Third "legacy preserving" option proposed by Martin Schulz:
  - Leave MPI_THREAD_MULTIPLE ambiguous and introduce two new threading levels to capture options 1-2 above.