Repository navigation
refactor(cuda): keep inflight pinned buffers unwrapped - #10878
Conversation
|
@claude review this pr in depth |
|
Claude finished @0ax1's task in 35s —— View job PR review: refactor(cuda): keep inflight pinned buffers unwrapped
VerdictI found no issues. The PR is ready to merge. This is a mechanical cleanup and nothing it touches behaves differently afterwards. Why the behaviour is unchanged
Optional follow-ups (outside this PR's scope)These are not blockers. Both are older than this PR, and this refactor makes them easier to spot.
ChecksI did not compile or run any tests, per the repo's guidance that verification is up to the user. This review is based on reading the code. CI covers |
Merging this PR will not alter performance1 warning
Comparing Footnotes
|
`InflightPinnedBuffer` held its completion event next to the buffer until #10713 reused the buffer's own event, leaving a wrapper around a single field. Store the buffers directly. Signed-off-by: Alexander Droste <alexander.droste@protonmail.com>
7180163 to
52dbd35
Compare
Summary
Simplifies inflight pinned-buffer tracking by removing InflightPinnedBuffer,
a wrapper containing only a PinnedByteBuffer.
The pool now stores PinnedByteBuffer directly. Transfer completion checks,
buffer reclamation, and synchronization are unchanged. The existing test
still checks that an inflight buffer retains its original allocation.
Changes
Benchmark
BenchPress comparison against develop, using 8,388,608 Vortex block rows.
Mainline: 9384250 (origin/develop)
HEAD: 52dbd35
Times are means in ms. Negative change = faster on HEAD.
Samples are measured reads per revision, excluding warmup.
Cold synthetic uses a 10-sample rerun. Cold numeric combines two
10-sample processes per revision in develop -> HEAD -> HEAD -> develop
order, with one excluded warmup per process. Other rows use the original
two-sample runs.
The original cold numeric run contained a 10.257 ms outlier. Subsequent
follow-ups varied in direction, so the final ABBA comparison above was
used to reduce run-order bias. It did not reproduce the slowdown:
combined means differ by -0.25%, and medians by +0.44%. The original
cold synthetic slowdown also did not reproduce in its longer rerun.
These measurements show no consistent performance regression, but do not
establish statistical significance or exact performance equivalence.
All benchmark result and cache checks passed. Cold/hot refers to the OS
file cache, not GPU or storage-controller caches.
GPU: GH200. CUDA: 13.1.
Vortex block rows: 8,388,608.
I/O: native. GPU allocator: cuda_async.
Lineitem SF10: 59,986,052 rows.