Repository navigation
feat(subsystembenchmarks): add Ray checkpoint load benchmark - #1047
Merged
zhixiangli merged 7 commits intoSep 17, 2026
Merged
zhixiangli merged 7 commits into
zhixiangli merged 7 commits into
Conversation
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Code Review
This pull request introduces a new subsystem benchmark suite for Ray checkpointing (ray_data), which includes drivers, configurations, and tests for both checkpoint read and write scenarios. It also updates the Cloud Build configuration to support a configurable boot disk size. The review feedback suggests two improvements: logging warnings instead of silently suppressing exceptions when verifying staging directories in common.py, and simplifying the redundant parameter resolution logic in read/driver.py by directly accessing attributes already handled by CheckpointParameters.
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## main #1047 +/- ##
=======================================
Coverage 90.25% 90.25%
=======================================
Files 16 16
Lines 3755 3755
=======================================
Hits 3389 3389
Misses 366 366 ☔ View full report in Codecov by Harness. 🚀 New features to boost your workflow:
|
Yonghui-Lee
force-pushed
the
ray-subsystembenchmark-checkpointing-load
branch
from
September 8, 2026 05:09
27c7ee2 to
225597f
Compare
zhixiangli
reviewed
Sep 11, 2026
… ray_pytorch and scope save timing to GCS - Rename checkpointing/ray_data -> checkpointing/ray_pytorch; the group uses Ray core actors with PyTorch DCP, not Ray Data. - Time only the staged-to-GCS transfer, so the published write throughput measures GCS bandwidth rather than state dict gathering and local staging. - Reserve one CPU per save worker, matching the Ray macrobenchmark.
Yonghui-Lee
force-pushed
the
ray-subsystembenchmark-checkpointing-load
branch
from
September 14, 2026 05:32
225597f to
deb0c8b
Compare
yuxin00j
reviewed
Sep 16, 2026
# Conflicts: # gcsfs/tests/perf/subsystembenchmarks/tests/test_run_groups.py
yuxin00j
approved these changes
Sep 16, 2026
zhixiangli
changed the title
# Conflicts: # gcsfs/tests/perf/subsystembenchmarks/checkpointing/ray_pytorch/common.py # gcsfs/tests/perf/subsystembenchmarks/checkpointing/ray_pytorch/configs.yaml # gcsfs/tests/perf/subsystembenchmarks/checkpointing/ray_pytorch/tests/test_configs.py # gcsfs/tests/perf/subsystembenchmarks/checkpointing/ray_pytorch/write/driver.py
zhixiangli
approved these changes
Sep 17, 2026
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
3 participants
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.
Depends on #1043
This PR adds the checkpoint load subsystem benchmark for Ray Data and Ray Train PyTorch workloads.
It benchmarks loading single-node and distributed checkpoints across FSDP, DDP, and Tensor Parallel + Data Parallel strategies, including cross-topology resharding.
Key Changes
Ray Checkpoint Load Driver:
RayCheckpointLoadWorkeractor executing distributed checkpoint load operations via PyTorch Distributed Checkpoint (DCP) andray.train.Checkpoint.RayCheckpointSetupWorkeractor that pre-populates the source checkpoint on GCS using the case's setup topology (setup_world_size,setup_tensor_parallel_size,setup_data_parallel_size).RayCheckpointReadDrivermanaging cluster initialization, fixture setup, and benchmark execution.