Flush GPU recv buffer after non-pof2 unfold receives#4
Open
mjwilkins18 wants to merge 1 commit into
Open
Conversation
The Phase-3 unfold PMPI_Recv on even folded ranks lands the final result in recvbuf via GPU-aware MPI (posted PCIe writes), but unlike every other receive site no cail_gpu_flush_recv_buf follows. The application may launch a kernel reading recvbuf immediately after MPI_Allreduce returns and observe stale device memory. Add the missing flush after the unfold receive in recursive doubling and Rabenseifner.
mjwilkins18
force-pushed
the
mjwilkins18/unfold-recv-flush
branch
from
a25ffcc to
53c264b
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
For non-power-of-two process counts, the Phase-3 unfold
PMPI_Recvoneven folded ranks lands the final allreduce result in
recvbufviaGPU-aware MPI — posted PCIe writes with no ordering guarantee against
subsequent kernel reads. Every other receive site in the algorithms is
followed by
cail_gpu_flush_recv_buf(), but these two are not:src/coll/allreduce/cail_allreduce_recursive_doubling.c(unfold receive)src/coll/allreduce/cail_allreduce_rabenseifner.c(unfold receive)CAIL launches no kernel after the unfold, but the application may launch
one reading
recvbufimmediately afterMPI_Allreducereturns andobserve stale device memory — affecting exactly the folded ranks at
non-pof2 scale.
This adds the missing flush after each unfold receive (no-op on the
host path).