10 27 2021

HACC WG 10-27-2021

Logically concurrent:

• Dan working on runnable litmus tests
• AI: Track down OMP version of my magic thread barrier example

Continuations proposal:

• Issue and PR posted
• Joseph is working on Open MPI implementation of proposed API
• Open question: Is session a good point of attachment?
  • Cannot share MPI objects across sessions

Digression into sessions progress semantics:

Two libraries (Sessions A and B):

Process 0:

MPI_Isend(to Process 1, Session A Comm)
MPI_Recv(from Process 1, Session B Comm) (If no progress on MPI_Isend, can deadlock)

Process 1:

MPI_Recv(from Process 0, Session A Comm)
MPI_Isend(to Process 0, Session B Comm)

• Conclusion that sessions should not cause this example to deadlock

Continuations proposal:

• MPI_Continue could allow requests from different sessions to be attached to the same continuation request
• The text would need to call this out as not allowed
• Otherwise, waiting on the continuation request is like waiting on an array of requests from different sessions (which MPI 4.0 forbids)
• Discussed binding of callbacks to MPI objects
• Which objects does it make sense to tie to?