RMA WG 02 22 2021

Memory model (updated by Akhil)

• Status: Proposal is still in preparation
• Plan: Akhil will add issue links to wiki; Min will upload June slides

Performance variable (updated by Dave)

• Status: willing to get feedback from group to simplify the proposed API
• Comments:
  • Manju: worth to get experiences from tools group (e.g., TAO, DDT)
  • Dave: feedback from VTune: MPI_T-like API will be beneficial
  • Min: will initiate email to MPI tools WG to get feedback about MPI_T
• Plan: Was presented at Austine F2F. Plan to get feedback from MPI and tools groups, and present an updated version in March spec meeting

Bundled communication (led by Dave)

• Status: slides (motivation/possible API) presented, will be updated to issue page
• Additional notes: use case example for motivation can be found at [slides] (https://chapel-lang.org/releaseNotes/1.19/05-benchmark-opts.pdf) and [github repo] (https://github.com/jdevinney/bale)
• Comments:
  • Manju: convert from a latency-bound problem to bw-bound
  • Min: Atomicity guarantee for bundled AMO? Maybe easier to go with only RMA first.
  • Akhil: better to allow different threads to issue bundled op; bundled collectives like API exists in NCCL
  • Manju: need key use cases for bundled collectives
  • Manju/Naveen/Min: Explicit v.s. Implicit API
  • Dave: Bundle across contexts might not make sense to have, for performance reason
• Plan: continue presentation & discussion at next call