Skip to content

RMA WG 03 14 2019

Jump to bottom
Naveen Namashivayam Ravichandrasekaran edited this page Mar 16, 2019 · 3 revisions

Agenda

  1. Discuss on the interaction between put-with-signal and other OpenSHMEM RMA and AMOs (Naveen)

Attendees

  • Pasha, Jim, Nick, Naveen, Bob, Anshuman, Min, Akil, Dave, Wazi, Khalid, and others

Open Action Items

  • Continue discussion on signal reset operation
  • Clarify on shmem_wait/test changes if anything necessary

New Action Items

  • None

Notes

put-with-signal interaction with other OpenSHMEM RMA and AMOs

  1. The overall acceptance of the following interactions with different operations are achieved in this RMA WG meeting
shmem_wait shmem_test put_signal(SET) put_signal(ADD) shmem_RMW_atomics
put_signal(SET) data race free data race free data race free no guarantees no guarantees
put_signal(ADD) data race free data race free no guarantees data race free no guarantees

  1. Pasha and Khalid suggested to include a separate API/semantic to include ordered/bundled operation. This can be orthogonal to put-with-signal. Though put-with-signal can be considered as a subset of this bundled operation, it is useful to have put-with-signal for performance reasons. Still a good proposal to consider for future.

  2. If target PEs is sure that no further update on the signal buffer is pending through any incoming put-with-signal on the wire, then it is free to reset the signal buffer using simple store operation. When it is hard to determine this condition, we might need some form of network reset operation - possibly the light-weight shmem_signal. There is no clear consensus on this, but still everyone understands that it is not for put-with-signal to decide this semantics. In case, if shmem_signal is introduced, then this operation should be data race free.

  3. Min asked about shmem_wait / shmem_test changes - do we need any new API for this? Discussion was pending.

  4. Overall the following semantic was decided to be acceptable for put-with-signal ** atomic with respect to itself ** atomic with respect to other put-with-signals of the same operation ** no guarantees with respect to shmem_RMW_atomics ** no guarantees with respect to other put-with-signals of different operations

shmem_signal

  1. Jim suggested Anshuman to hold the shmem_signal proposal for the time being
  2. It looks like shmem_signal is just a put-with-signal with 0-size put operation
  3. shmem_signal can provide programmers with some code quality - hence, we could wait on this proposal until all the issues with put-with-signal are sorted out
Clone this wiki locally