RMA WG 11 08 2018

Agenda

1. Memory Model Discussion, Anshuman (NVidia)

Attendees

• Anshuman, Akhil (NVIDIA)
• Jim Dinan, David Ozog, Wasiur Rahman (Intel)
• Ferrol (MTSU)
• Naveen (Cray)
• Manjunath Gorentla
• Nick

Notes

• (Anshuman) Not to break existing applications
• Leave shmem_p as non-atomic operation
  • Jim - Do not let things open for interpretation
• (Jim)Two implementations of shmem_p
  • Those that can be used in Wait operation
  • Any types that cannot be used in wait operation
• (Manju) We cannot guarantee partial update or complete update
• (Jim) Partial updates are disallowed if wait would see and react to them in that implementation
  • Should it be implementation defined?
• (Jim) Put with signal operation - did not want to specify that the signal update would be atomic
  • shmem_p and signal update are simple
• (Manju) Even with serialization, not 100% sure if it is enough
  • It gives atomicity but not single copy atomicity
• (Nick) Atomicity because of cache alignment should not be an issue in any known architecture* Jim and Manju are OK with non-atomic shmem_p compatible with shmem_wait_until and shmem_test (so implementation ensures that partial update does not cause wakeup)* Only shmem_atomic_set and shmem_p using the same type are compatible with shmem_wait_until and shmem_test