Skip to content

RMA WG 04 26 2018

Jump to bottom
James Dinan edited this page Apr 26, 2018 · 3 revisions

Agenda

  1. Collect opens, assign note taker
  2. Check in with proposal leads
  3. Atomic memory model proposal

Attendees

  • Dave Ozog (Intel), Ferrol (ORNL), Min Si (ANL), Naveen (Cray), Swaroop (ORNL), Thomas Naugton (ORNL), Tony (SBU), Wasiur Rahman (Intel), Bob (Cray), Jim Dinan (Intel)

  • Put with signal

    • Need clarity on signal operation (i.e. atomic write/add or put)

  • SHMEM wait-then-set

    • no update this week

  • Wait/test-some

    • Index array of input/output values should be boolean
    • Discussion of return value -- all completed events or number completed by this call to shmem_wait_some
    • Consider whether entries in array of status should have more than two states (e.g. completed, incomplete, ignore)

  • Nonblocking AMOs

    • Proposed for fetching and non-fetching
    • Decided to drop nonblocking versions of non-fetching AMOs
    • Current implementations of non-fetching AMOs are already nonblocking (i.e. buffered/passed by value)
    • Gathering performance results

  • Atomic memory model #204

    • Jim read an updated draft for feedback
    • Dave: suggest including in the prose how to fix each anti-example
      • e.g. in Scenario 2, adding quiet isn't sufficient, a barrier is needed to ensure remote AMOs are not concurrent
    • Min: suggest including an example of how to use API properly (all PEs increment X on 0, 0 waits on X)
    • Most WG members thought the proposal is an improvement, although perhaps straying a bit close to a user tutorial
    • Jim will sync with other committee members who had voiced this concern ahead of the next meeting to get their feedback
Clone this wiki locally