RMA WG 09 20 2021

Today we review active WG topics to gather status and next steps.

1. Memory model clarifications (Akhil + Jim) [summary of issues presented on June 08, 2020: slides] #477
2. Performance variables (Wasi) #295
   • Follow up on Wasi's OpenSHMEM workshop talk
3. Bundled communication (Dave) #189
   • AI: Manju to prepare a proposal on a work queue extension to OpenSHMEM
   • AI: Dave to draft bundled communication proposal
   • Jim proposed that we explore both implicit and explicit operation queueing models at the same time
4. Block interleaved communication (Jim) #448
   • We discussed opportunities top optimize the communication using features like request chaining, InfiniBand UMR, etc.
   • Dave: should we add nonblocking variants?
5. Noncontiguous communication (more general than ibput) (Jim) #365
   • Yanfei will review notes from Min and current proposals
   • Decisions:
     1. Keep or deprecate interleaved APIs
     2. If keep, should we add block size?
     3. If not keep, block interleaved proposal can be dropped and replaced with something better (not strictly either/or)
     4. What’s the right noncontiguous API to support HPC workloads and enable optimization?
6. Signal operation (Jim + Naveen) #382
   • Need to revisit, this is weaker than an atomic and can provide better performance if the platform has a path for it (e.g. single copy atomic store).
7. Extending OpenSHMEM for GPUs (Jim, Khaled, Min, et al.)
   • Will review topics that Jim mentioned in OpenSHMEM workshop talk.