RMA WG 08 30 2018

Agenda

1. Test/wait update on proposed changes
2. Continue memory model discussion

Attendees

• David Ozog, Wasi Rahman, Jim Dinan (Intel)
• Anshuman Goswami (Nvidia)
• Ferrol, Swaroop Pophale (ORNL)
• Min Si (ANL)
• Naveen (Cray)

Open Action Items

• None

Action Items

Notes

• Wait/test update:

  • Straw poll from f2f: Yes to adding *_any interfaces
  • Straw poll from f2f: Yes to adding *_all interfaces
  • Is mask a necessary feature or something optional that should be part of the function name?
    • Mask takes the place of MPI_REQUEST_NULL
    • Without the mask wait_any doesn't work for common usage models
  • Should wait_some check all elements?
    • Min: Make semantics consistent between wait and test some routines
    • Min: Both should check all ivars
    • Reviewed MPI text; text indicating that each request is visited is in advice (non-normative), but probably is the universal behavior across implementations
    • Waitsome should visit every ivar at least once
    • Testsome should visit every ivar exactly once
    • Min: MPICH test/wait-some both check all elements
    • Will look into adding restrict to bindings

• Memory models update:

  • Review of discussion at the f2f meeting
  • First, clarify existing operational semantics as best we can
  • In parallel, develop more rigorous axiomatic memory model
  • Anshuman: Prefer wait_until definition where only used with SHMEM atomics
  • Nick measured 15% performance hit when using atomics versus scalar put to satisfy wait_until
  • Reviewed possible definitions for wait_until, will resume discussion here next time