Document the benchmark harness's methodology gap - #8
Merged
Conversation
Ran the fixed SqpollEngine head-to-head against POSIX pread/pwrite, libaio, and plain io_uring on the same NVMe box, to check whether Phase 5's original "trailing every baseline" finding still held. It did, at every thread count, but the comparison itself is invalid: the scratch file never opens with O_DIRECT and its 64 MiB working set fits in the page cache, so the syscall baselines are measuring DRAM speed, not disk speed, and every backend including SqpollEngine runs at queue depth 1, which never gives an async batching design anything to amortize its syscall-avoidance over. Neither problem is a regression from today's other changes. Both were already true of every prior run of this benchmark. Recorded as unstarted work: O_DIRECT with aligned buffers so the baselines hit the real device, and a queue-depth sweep across all four backends being compared, not just the SQPOLL ones.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
Docs-only. Records why the fixed
SqpollEnginehead-to-head against POSIXpread/pwrite,libaio, and plainio_uringdoesn't settle the paper's core hypothesis, and what the harness needs before it can.What happened
Ran all five backends on the same
n2-standard-4box to check whether Phase 5's original "SqpollEnginetrailing every baseline" finding still held after the shared-poller and backoff fixes. It did, by a wide margin at every thread count. Before writing that up as a finding, checked why the baseline numbers looked too fast to be real disk I/O and found two problems inbenchmarks/iops_throughput_benchmark.cpp:MakeScratchFile()never opens withO_DIRECT, and the working set is 64 MiB. On a box with much more RAM than that, every backend's I/O after the first pass comes from the page cache, not the drive. POSIXpread/pwritehit 698K-1.3M IOPS in this run, which is DRAM speed.SqpollEngineincluded, runs at queue depth 1: submit one request, block-wait for that exact completion, submit the next. An async batching design has nothing to amortize its syscall-avoidance over at depth 1, so losing here is expected, not a regression.Neither problem is new; both were already true of every prior run of this suite. This PR just writes that down instead of letting the "trailing every baseline" line stand unqualified.
Next (not in this PR)
Two items recorded as unstarted work in
ROADMAP.md:O_DIRECTwith sector-aligned buffers so the baselines hit the actual device, and a queue-depth sweep (1/8/32/64/128) across all four backends being compared. Planned for a follow-up branch, not started here.Testing
None needed;
ROADMAP.mdonly.