Skip to content

[v4.1.x] RDMA osc: perform CAS in shared memory if possible #10513

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
bwbarrett merged 1 commit into from
Jul 13, 2022
Merged

[v4.1.x] RDMA osc: perform CAS in shared memory if possible #10513

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
26 changes: 18 additions & 8 deletions ompi/mca/osc/rdma/osc_rdma_accumulate.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -814,14 +814,24 @@ int ompi_osc_rdma_compare_and_swap (const void *origin_addr, const void *compare
lock_acquired = true;
}

/* either we have and exclusive lock (via MPI_Win_lock() or the accumulate lock) or the
* user has indicated that they will only use the same op (or same op and no op) for
* operations on overlapping memory ranges. that indicates it is safe to go ahead and
* use network atomic operations. */
ret = ompi_osc_rdma_cas_atomic (sync, origin_addr, compare_addr, result_addr, dt,
peer, target_address, target_handle, lock_acquired);
if (OMPI_SUCCESS == ret) {
return OMPI_SUCCESS;
/* operate in (shared) memory if there is only a single node
* OR if we have an exclusive lock
* OR if other processes won't try to use the network either */
bool use_shared_mem = module->single_node ||
(ompi_osc_rdma_peer_local_base (peer) &&
(ompi_osc_rdma_peer_is_exclusive (peer) ||
!module->acc_single_intrinsic));

if (!use_shared_mem) {
/* either we have an exclusive lock (via MPI_Win_lock() or the accumulate lock) or the
* user has indicated that they will only use the same op (or same op and no op) for
* operations on overlapping memory ranges. that indicates it is safe to go ahead and
* use network atomic operations. */
ret = ompi_osc_rdma_cas_atomic (sync, origin_addr, compare_addr, result_addr, dt,
peer, target_address, target_handle, lock_acquired);
if (OMPI_SUCCESS == ret) {
return OMPI_SUCCESS;
}
}

if (!(lock_acquired || ompi_osc_rdma_peer_is_exclusive (peer))) {
Expand Down