RMA WG 01 11 2021

Attendees: Naveen, David, Nick, Akhil, Manju, Wasi, Jim, Min

Agenda

• Go over all active topics and discuss & update plan

Notes

1. Performance variable and APIs

   • Wasi: discussed offline with Dave. Currently low priority at Intel. Will have a plan at the next meeting.

2. Bundled communication

   • Dave and Manju will work on this ticket. Dave plans to create slides to capture the available options so that vendor/user can give feedback
   • Using context to support? E.g., user enables bundled comm for a specific context.
   • Naveen: will provide some inputs on this idea after proposal out
   • Jim: the proposal aims to reduce cost to push commands to network. Would benefit if target to specific middleware and app, not generic optimization; might benefit to implement in a library built on top of SHMEM implementation as the bundling implementation might be complex.
   • Min: like a hint to runtime. I.e., the user defines the opportunity but runtime does not have to bundle if it is not good for performance on the specific hardware or runtime implementation.
   • Jim: the bundling API is to hint runtime certain operation is not latency critical and can be delayed

3. Memory model clarifications

   • Jim: trying to map C++ memory model
   • Akhil has presented slides at previous F2F meeting and forum meeting.
   • Trying to identify specific incomplete semantics and missing API after 1.5 spec, e.g., memory ordering of AMO, missing memory fence.
   • Fresh ticket for 1.6 spec, the original memory model (driven by Anshuman) will be for reference.
   • Separate ticket for each topic? better to do a single ticket as the memory model is the basis.
   • Jim and Akhil: will create slides to kick off next steps.
   • Previous discussion can be found at #339

4. SHMEM_TEAM_PTR

   • Dave will lead the ticket

5. Noncontiguous APIs

6. Github discussion feature enabled

Plan for next meeting

• Go over the remaining active topics and discuss any specific item