perf(io): add busy-wait completion mode and numpy batch-transfer mars…#444
Open
TianDi101 wants to merge 3 commits into
Open
perf(io): add busy-wait completion mode and numpy batch-transfer mars…#444TianDi101 wants to merge 3 commits into
TianDi101 wants to merge 3 commits into
Conversation
…haling WaitBusy() spins on the completion flag instead of blocking on the condition variable, trading a CPU core for lower wakeup latency. BatchRead/BatchWrite now accept numpy arrays for offsets/sizes, avoiding per-element Python-list conversion. Benchmark gains --busy-wait, --post-batch-size, --warmup-iters flags and separate launch/transfer latency reporting. Co-Authored-By: Claude Sonnet 5 <[email protected]>
Co-Authored-By: Claude Sonnet 5 <[email protected]>
TianDi101
force-pushed
the
perf/io-busy-wait-numpy
branch
from
July 23, 2026 03:44
a5d44cd to
bf3078f
Compare
…_once actually uses _validate_rdma() unconditionally assumed the strided (buffer_size+1) offset layout whenever --batch-contiguous was off, but that layout is only ever used by run_batch_once() (--enable-batch-transfer). The default path, run_single_once(), always issues transfers at contiguous offsets (buffer_size * i) regardless of --batch-contiguous, causing spurious AssertionErrors on batch_contiguous=False runs with batch > 1 even though the RDMA transfer itself succeeded. Gate the strided comparison on self.enable_batch_transfer and add a branch for the non-batched, non-contiguous-flag case that compares against the contiguous prefix that was actually transferred.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
…haling
WaitBusy() spins on the completion flag instead of blocking on the condition variable, trading a CPU core for lower wakeup latency. BatchRead/BatchWrite now accept numpy arrays for offsets/sizes, avoiding per-element Python-list conversion. Benchmark gains --busy-wait, --post-batch-size, --warmup-iters flags and separate launch/transfer latency reporting.
Motivation
Technical Details
Test Plan
Test Result
Submission Checklist