RMA WG 06 18 2020

OpenSHMEM 1.6 Topics Discussion

Even Weeks Team (Chair: Jim)

• Memory model (Akhil + Jim)
• Performance variables (Wasi)
• Bundled communication (Jim)
• SHMEM_TEAM_PTR (Jim)
• Block interleaved communication (Jim)
• Noncontiguous communication (more general than ibput) (Jim)
• Signal operation (Jim + Naveen)
• Extending OpenSHMEM for GPUs (Jim, Khaled, Min, et al.)
  • Host initiated transfers of GPU memory (may be addressed through memory spaces)
  • Device triggered transfers (host sets up the transfer, device releases it when data buffers are ready)
  • Device initiated transfers (single-threaded and collective on a thread group)
  • Integration of OpenSHMEM operations into device scheduled flows (streams, graphs, etc.)
• User-level threading integration/support (Wasi)

Odd Squad (Chair: Manju)

• Memory spaces (Naveen)
• Per-PE quiet and fence (Manju)
• Interoperability (Min Si)
• Glossary
• Nonblocking collectives (Dave)
• Alltoallv (Manju)