Repository navigation
feat: sleep when blocked, SPSC bulk transfer, MPMC pow2 indexing - #10
Merged
Merged
Conversation
Blocking push/pop yield briefly then sleep with doubling backoff (cq/backoff.hpp): 2s blocked drops from 1975ms to 10ms of CPU, wakeup bounded at 1ms. Two designs were measured and rejected on the way — a futex gate cost 70% of SPSC pair throughput (publisher-side check drains the store buffer), and inlining the sleep machinery into push/pop cost 4x on the round trip (inliner budget) — hence the noinline cold Backoff::wait(). Remaining cost: ~9% on the continuous 1p+1c rendezvous; round trip unchanged at 1.80G ops/s. try_push_n/try_pop_n move a batch with one index publish: 1.69G ops/s at batch 64, 3.1x the single-op pair. MpmcQueue maps ticket->slot with an AND when capacity is a power of two; within noise on M2, kept as strictly cheaper. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…ve-only test Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…ructure PR #11 moved per-milestone results to docs/results.md and rewrote the README around the three queues, so this branch's README section conflicts. Resolution: - The new results writeup moves to docs/results.md as a v3.1 section, with a row in the milestone map and an explicit note on the round-trip control level differing across sessions. - README (main's version) is updated where this branch made it stale: the "A waiting thread" row now says the rings spin briefly then sleep, the three-queues table gains a bulk-ops row, and Performance gains a bullet for the ~9% sleeping residual and the 1.69G bulk figure. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0133qvk9V3aREYsxvMCbZ9Am
…hygiene
- try_push_n/try_pop_n take std::span<T> instead of pointer + length: the
span carries both, call sites shrink, and the A/B benchmark shows no
cost (span 1.64G ± 0.03 vs pointer 1.60G ± 0.04 ops/s at batch 64).
- Backoff::wait() moves out of the class body per STYLE.md ("no function
bodies in class definitions") and gains the /// contract comment
include/cq/ requires; the constants get /// one-liners.
- slot_index's comment claimed the mask is set iff capacity is a power of
two — false at capacity 1 (2^0, mask 0, modulo path). Comment now says
what happens there; the code was already right.
- Drop includes the switch from yield loops to Backoff left unused:
<chrono>/<thread> in spsc_queue.ipp, <algorithm>/<chrono>/<thread> in
mpmc_queue.ipp.
Verified: 65/65 under TSan, 66/66 in Release; clang-format clean;
clang-tidy 0 findings; bulk benchmark within 1 sigma of the published
1.69G ± 0.09G.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0133qvk9V3aREYsxvMCbZ9Am
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
Three features closing the gap to the industrial queues, with the two designs that measured worse documented as findings:
push/popyield briefly, then sleep with doubling backoff (newcq/backoff.hpp): 2 s blocked = 10 ms of CPU, down from 1975 ms, wakeup bounded at 1 ms. Rejected on measurement: a futex gate (C++20 atomic wait/notify eventcount) cost 70% of SPSC pair throughput — the publisher-side waiter check after a release store drains the store buffer; and inlining the sleep machinery intopush/popcost 4× on the round trip by blowing the inliner budget — hence thenoinlinecoldBackoff::wait(). Honest residual: continuous 1p+1c rendezvous runs ~9% below pure spin (547M vs 602M ops/s); round trip unchanged at 1.80G ops/s.try_push_n/try_pop_n: up to astd::span's worth of items per call, one index publish per batch — the lever the results' cached-index section named. 1.69G ops/s at batch 64, 3.1× the single-op pair.Files changed & why
include/cq/backoff.hppnoinline wait()keeps callers inlinableinclude/cq/spsc_queue.hpp/.ippBackoff;try_push_n/try_pop_n(span API, single publish per batch); docsinclude/cq/mpmc_queue.hpp/.ippBackoff;slot_index()with pow2 mask; docstests/spsc_queue_test.cppbench/queue_bench.cppSpscQueue/bulk64_throughputbenchmarkREADME.mddocs/results.mdPost-review updates (after the #11 restructure landed on main)
mainin and resolved the README conflict: the results writeup now lives indocs/results.md(v3.1 section + milestone-map row); the restructured README got the integration edits listed above instead of a results section.std::span<T>instead of pointer + length (A/B benched: span 1.64G ± 0.03 vs pointer 1.60G ± 0.04 — no cost);Backoff::wait()moved out of the class body per STYLE.md and the header gained the required///docs;slot_index's comment corrected for capacity 1 (pow2 with mask 0 — code was already right); unused<algorithm>/<chrono>/<thread>includes dropped from the.ipps.Test plan
🤖 Generated with Claude Code