PR #21708: NUMA-pin host memory buffers for D2H/H2D transfers #22243
+177
−97
Conversation
This file contains 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
![copybara-service[bot]](https://avatars.githubusercontent.com/in/44061?s=60&v=4){kind=small}
copybara-service
bot
force-pushed
the
test_722688719
branch
from
1d31798 to
82fdb46
Compare
Imported from GitHub PR #21708 This ensures that the pinned host buffers used for transfers between host and device are pinned to the NUMA node closest to the device. It had a previous life as #15216. In a benchmark that triggers large, concurrent, copies from all devices to the host then achieved D2H throughput is around 33 GiB/s with NUMA pinning on a DGX H100 node (2xCPU, 8xH100). Without pinning, the achieved throughput is around 13.5 GiB/s from the same benchmark. While it is already possible to achieve the correct NUMA pinning in process-per-GPU and process-per-NUMA-node configurations using `numactl` or similar, achieving correct pinning in process-per-node configuration requires logic inside XLA. Copybara import of the project: -- 1a2d98b by Olli Lupton <[email protected]>: NUMA-pin host memory buffers for D2H/H2D transfers -- 60a4659 by Olli Lupton <[email protected]>: 256 byte alignment for host allocations when NUMA is not enabled -- 839da45 by Olli Lupton <[email protected]>: Address review comments -- b61ce94 by Olli Lupton <[email protected]>: Drop TENSORFLOW_USE_NUMA -- 793fde0 by Olli Lupton <[email protected]>: std::string_view -> absl::string_view Merging this change closes #21708 FUTURE_COPYBARA_INTEGRATE_REVIEW=#21708 from olupton:numa 793fde0 PiperOrigin-RevId: 722688719
copybara-service
bot
force-pushed
the
test_722688719
branch
from
82fdb46 to
051518c
Compare
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.
PR #21708: NUMA-pin host memory buffers for D2H/H2D transfers
Imported from GitHub PR #21708
This ensures that the pinned host buffers used for transfers between host and device are pinned to the NUMA node closest to the device. It had a previous life as #15216.
In a benchmark that triggers large, concurrent, copies from all devices to the host then achieved D2H throughput is around 33 GiB/s with NUMA pinning on a DGX H100 node (2xCPU, 8xH100). Without pinning, the achieved throughput is around 13.5 GiB/s from the same benchmark.
While it is already possible to achieve the correct NUMA pinning in process-per-GPU and process-per-NUMA-node configurations using
numactlor similar, achieving correct pinning in process-per-node configuration requires logic inside XLA.Copybara import of the project:
--
1a2d98b by Olli Lupton [email protected]:
NUMA-pin host memory buffers for D2H/H2D transfers
--
60a4659 by Olli Lupton [email protected]:
256 byte alignment for host allocations when NUMA is not enabled
--
839da45 by Olli Lupton [email protected]:
Address review comments
--
b61ce94 by Olli Lupton [email protected]:
Drop TENSORFLOW_USE_NUMA
--
793fde0 by Olli Lupton [email protected]:
std::string_view -> absl::string_view
Merging this change closes #21708
FUTURE_COPYBARA_INTEGRATE_REVIEW=#21708 from olupton:numa 793fde0