RMA WG 01 17 2019

Agenda

1. Wait/test changes to put-with-signal proposal (Naveen)

Attendees

• Not recorded

Open Action Items

• None

New Action Items

• None

Notes

• Latest update on the pt-to-pt synchronization issue under the memory model umbrella : 1) introduce shmem_signal which is atomic wrt itself and shmem_wait_until but not wrt other shmem_atomic_ operations; 2) specify shmem_wait_until to be atomic wrt both shmem_signal and shmem_atomic_ operations; 3) shmem_signal and shmem_atomic_ can be used for pt-to-pt synchronization
• There is concern that shmem_wait_until performance will be badly affected because it has to go through network atomics for some platforms
• Khaled thinks that there is no OpenSHMEM implementation that does network atomics in shmem_wait_until but applications still work so adding this restriction on shmem_wait_until is not needed
• The note to implementors in shmem_wait_until prohibits partial update so current implementations not using network atomics suggests that the implementors note is ignored : should the implementors note be removed?
• Khaled suggested that making the signal 1 byte wide will ensure that partial updates are not seen because no known implementation splits a message at less than 1 byte : the problem there is 1 byte signal is restrictive for something like continuously increasing counter used as signal
• Khaled and Manju thinks there is value in allowing partial updates to be seen by shmem_wait_until for an usecase where the signal is just used to communicate arrival of the initiator PE : proposal for two variants of shmem_wait came up; 1) shmem_wait can observe partial updates 2) shmem_wait_signal cannot observe partial updates
• The change in shmem_wait_until to make it atomic wrt to shmem_signal and shmem_atomic_ affects the shmem_waitsome/testsome proposal by Dave : what happens to the scheduled reading in next specification meeting?
• Naveen and Bob are fine with the proposed changes in shmem_wait_until and introduction of shmem_signal : shmem_put_with_signal to rely on shmem_signal in that case
• Manju brought up the similarity that the modified shmem_wait_until has to shmem_atomic_fetch
• Khaled mentioned that user feedback is needed before changing shmem_wait_until