Skip to content

RMA WG 06 21 2018

Jump to bottom
spotluri edited this page Jun 21, 2018 · 3 revisions

Agenda

  1. Memory Model Proposal (Anshuman Goswami)

Attendees

  • TBD

Open Action Items

  • None

Action Items

  • TBD

Notes

  • slide 3: fence orders non-blocking puts
  • slide 4:
    • collective operations issued by different threads (using different psync/pwrk arrays) are unordered collective operations from the same thread using different psync/pwrk arrays are unordered correction: collectives issued in program order and use same active set, psync/pwrk arrays are ordered
    • non-fetch AMOs or non-blocking AMOs are unordered correction: blocking fetch AMOs issued in program order are ordered
    • correction: change the prefix statement that the two accesses are issued by the same PE in program order
  • slide 5:
    • Spec already mentions "OpenSHMEM programs are expected to exhibit progression of communication both with and without OpenSHMEM calls" Need to agree that interpretation of this is the stronger progress guarantee being proposed
    • Clarificaition: SHMEM library having a progress thread does not mean guaranteed "thread-safety" at shmem API level. The SHMEM library can still only provide "thread-single".
Clone this wiki locally