RMA WG 07 19 2018

Agenda

1. Memory Model Proposal (Anshuman Goswami)

Attendees

• Note taker: Wasi (Intel)
• Attendees: Anshuman Goswami and Sreeram Potluri (NVIDIA); Khaled Hamidouche (AMD); Naveen (Cray); Ferrol and Manju (ORNL); Min Si (ANL); Jim, Dave, and Wasi (Intel)

Open Action Items

• None

Action Items

• None

Notes

• Using shmem_put/iput/p to trigger shmem_wait_until is platform defined

  • Manju: backward compatibility issue
  • Khaled: Why a separate API?
  • Manju: Most implementation will grab a lock before doing an update (putmem)
  • Sreeram: will revisit this; to make int_p/g work, it needs a vector, so a new API
  • Jim: architecture supporting different storage instructions for atomic read/write v/s. read/modify/write
  • Sreeram: GPUs, the strong and weak operations, atomic set and atomic cswap; draw it pictorially to see if this makes sense

• Next steps?

  • New subsection in the spec, making sure the consistency and sync is defined in the spec is consistent with this proposal
  • Create all the open issues in the github before start making changes
  • Two separate tickets to make sure everything is coherent
  • Sreeram: Leave the non-atomic operations when ordering is concerned since that is undefined
  • Jim: Need a to-do list, and go over each changes individually
  • Started on each of the topics and preparing the texts related, no explicit implication on put-with-signal proposal (question asked by Naveen); propose updates in the spec through email