Repository navigation
UCT/ROCM: adjust rocm/ipc performance parameters - #9568
Merged
Merged
Conversation
edgargabriel
force-pushed
the
topic/rocm-ipc-v2-tuning
branch
2 times, most recently
from
December 28, 2023 20:56
360b138 to
8e1f267
Compare
Contributor
Author
|
setting the DoNotMerge flag since I might have a better solution than the min_zcopy (which did most of the heavy lifting in this fix) |
edgargabriel
force-pushed
the
topic/rocm-ipc-v2-tuning
branch
from
January 1, 2024 15:10
8e1f267 to
e1e7493
Compare
Contributor
Author
|
Lets stay for now with this solution here to deal with the intra-node collectives performance issue that I'm looking into. An alternative approach that I explored was to set the pre_send_overhead to a non-zero value in rocm_ipc, that gave me similar performance improvements as the solution in this pr. However, I started to hit issue #9372 with the pre_send_overhead in a number of tests. Until we find a solution to the problem in that issues, this solution here (which mostly relies on setting a minimum message length for the rocm/ipc/zcopy operations) seems to be more stable. This pr will also have to go to the 1.16 branch. |
yosefe
previously approved these changes
Jan 2, 2024
Member
edgargabriel
force-pushed
the
topic/rocm-ipc-v2-tuning
branch
from
January 2, 2024 14:36
e1e7493 to
6ef3cfa
Compare
introduce two new parameters for rocm/ipc that allow to fine tune the protocol selection. Specifically, they help to resolve the issue seen with v2 protocols for short message intra-node collective operations.
edgargabriel
force-pushed
the
topic/rocm-ipc-v2-tuning
branch
from
January 2, 2024 14:41
6ef3cfa to
1e91ddb
Compare
Contributor
Author
|
@yosefe thank you, I think I fixed the issues |
yosefe
approved these changes
Jan 3, 2024
yosefe
enabled auto-merge
January 3, 2024 07:59
ovidiusm
pushed a commit
to ai-dynamo/nixl
that referenced
this pull request
Oct 12, 2026
## What? Set `ROCM_IPC_MIN_ZCOPY=0` in the UCX context NIXL creates. ## Why? `rocm_ipc` defaults to a 128 byte minimum zero-copy size (openucx/ucx#9568, for small-message MPI collectives). UCP put/get zero-copy protocols reject any lane with a non-zero minimum, so NIXL VRAM to VRAM transfers on AMD never use `rocm_ipc` (XGMI) and are staged through host memory instead. `cuda_ipc` already uses 0. Setting it only in NIXL's UCX context leaves MPI and other UCX users unchanged. A user-set `UCX_ROCM_IPC_MIN_ZCOPY` still wins. On UCX builds without ROCm support the setting is unknown and is ignored with a debug log. With NIXL's default peer error handling, `rocm_ipc` is selected only once it advertises peer-failure support (openucx/ucx#11977, open). This PR does not depend on it, but the full benefit needs both. ## Results MI355X, NIXL WRITE GPU0 to GPU4, UCX master + openucx/ucx#11977: | size | without this PR | with this PR | |---|---|---| | 64 KiB | 1.68 GB/s | 4.54 GB/s | | 512 KiB | 11.89 GB/s | 23.74 GB/s | | 2 MiB | 30.05 GB/s | 43.69 GB/s | | 128 MiB | 60.22 GB/s | 60.85 GB/s | UCX protocol selection (`UCX_PROTO_INFO=y`) changes from a rendezvous read to direct `rocm_ipc` zero-copy for all sizes. <!-- This is an auto-generated comment: release notes by coderabbit.ai --> ## Summary by CodeRabbit * **Improvements** * Updated ROCm IPC zero-copy settings during UCX initialization. * Added a warning when AMD GPUs are detected but UCX lacks ROCm memory-type support, helping identify a configuration that may affect GPU communication. <!-- end of auto-generated comment: release notes by coderabbit.ai --> Co-authored-by: Colin Hirsch <chirsch@nvidia.com>
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
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.
What
introduce two new parameters in rocm/ipc that allow to fine tune the protocol selection.
Why ?
Help to resolve the issue seen with v2 protocols for short message intra-node collective operations.