RMA WG 07 05 2018

Agenda

1. Memory Model Proposal (Anshuman Goswami)

Attendees

• Note taker: Jim Dinan (Intel)
• Attendees: Anshuman Goswami and Sreeram Potluri (NVIDIA), Nick Park (DOD), Khaled Hamidouche (AMD), Megan Grodowitz (ARM), Naveen and Bob (Cray), Swaroop Pophale and Manju (ORNL), Min Si (ANL), Michael Raymond (HPE)

Open Action Items

• None

Action Items

• None

Notes

• Applied updates suggested in previous meeting, including updates to ordering table
  • Q: Why include barrier in the table? Is the inter-PE synchronization relevant here?
  • A: Focus is on memory ordering and barrier is another way to achieve the quiet
• Discussion of ordering in OpenSHMEM model
  • Discussion of read-after-read ordering
  • Motivation to consider additional semantics clarifications: compiler optimizations involving OpenSHMEM operations -- What is allowed by compiler in terms of reordering?
  • Q: Is collective ordering a basic property? Current ordering semantics may be more relevant to pSync versus user's data buffers.
• Discussion of reads-from and happens-before
  • Are shmem_p and shmem_g atomic compatible operations?
  • Is wait_until compatible with put/get operations? Yes, it is.
  • Polling using shmem_g instead of wait_until -- what are the implications to the shmem_g routine and SHMEM memory model?