Skip to content

perf: run handlers queued behind the reactor first while that pays off - #416

Open
sgerbino wants to merge 1 commit into
developfrom
pr/recycling-free-list
Open

sgerbino wants to merge 1 commit into
developfrom
pr/recycling-free-list

Conversation

@sgerbino

@sgerbino sgerbino commented Nov 12, 2026 •

Copy link
Copy Markdown
Collaborator

When the reactor sentinel reached the front of the queue with handlers still behind it, the scheduler ran a non-blocking reactor poll before running them. Those handlers usually start the next I/O themselves, such as a coroutine an I/O completion woke, and that I/O often completes on the spot. Running them first makes many of those polls unnecessary.

This change cuts the polls a busy loop makes while handlers are queued. In pingpong on epoll, they drop by about half (485K → 264K). The saving is CPU per request under load. An idle or lightly loaded loop rarely has handlers queued when the sentinel comes up, so it behaves as before.

The sentinel now moves behind the queued handlers once, then polls on its next turn regardless, so handlers that keep reposting still cannot starve I/O. poll() and poll_one() are unchanged.

Waiting doesn't always pay. On macOS, loopback data arrives a little later than on Linux. A fixed deferral there let the handlers drain the queue, their speculative reads and writes hit EAGAIN about 8× as often and parked, and kqueue bidirectional/1048576 lost 13–14% in every run. reactor_defer_policy scores each deferral by what its turn found:

  • handlers still queued: a hit, score −1
  • queue empty: a miss, score +8, capped at 64
  • the sentinel defers only while the score is below 8; each turn that polls instead lowers it by 1, so a workload that changes gets tried again

A stopped run resets the score. A deferral it left pending stays marked, so the next run polls first, and the sentinel is never deferred twice in a row, even across runs. A workload made of many short run() calls therefore starts each run with a fresh score.

Tests:

  • reactor_defer_policy.cpp checks the scoring deterministically, with no timing.
  • testPostedWakeRunsBeforeNextPoll in the reactor fault suite counts the backend's wait calls to check that a posted wake runs before the next poll.
  • testStoppedDeferralPollsFirst checks that a run stopped with the sentinel deferred makes the next run poll first.
  • testRestartWhileRunningKeepsPolling and testStopPumpKeepsPolling check that a handler calling restart() every turn, and a loop of post-a-stop / run() / restart(), still let a parked read complete.

The recycling frame allocator change this used to pair with merged as cppalliance/capy#485, so both sides of the benchmark A/B build against capy develop.

Results

All numbers below are over loopback. A skipped poll saves well under a microsecond. That is a visible share of a 2–7 µs loopback round trip, so the latency rows overstate what a real network sees. The CPU saved per request doesn't depend on the network, so the http_server rows are the better guide for real deployments.

Benchmark report for cf9d859 vs develop 7cb81d1 (run 44195833544): 20 rows flagged, all improvements, none slower beyond noise on epoll or kqueue.

Area Change
local_socket_latency (epoll, kqueue) +6–7%
socket_latency (epoll) +2.5–3%
http_server (epoll, kqueue) +0–1.7%
fan_out ±0.5%
kqueue socket_throughput/bidirectional/1048576 +1.9%

uring and IOCP have their own schedulers and run the same code on both sides. Their rows still moved by up to ±10%, which shows the noise floor on those runners.

macOS ARM64 CI, kqueue, 7 interleaved reps, vs deferral off:

Row fixed deferral adaptive
bidirectional/1048576 −14.4% +1.3%
bidirectional_lockless/1048576 −13.5% +1.2%
bidirectional/262144 −0.3% −0.7%
unidirectional/1048576 −0.1% +0.4%
pingpong/64 ≈0 ≈0

Local epoll, one idle machine, 5 interleaved runs:

Row off fixed adaptive
pingpong/64 306.8K 328.4K 327.8K
concurrent_parents_lockless/16 15,430 15,698 15,474
fork_join_lockless/64 4586 4584 4586
bidirectional/1048576 8.54 8.54 8.44 ±1.9%

The adaptive rule keeps the poll savings and recovers the macOS bulk rows. It gives up fixed deferral's small concurrent_parents_lockless/16 gain: that workload's retries land at the end of a lap, where the queue empties.

Summary by CodeRabbit

  • New Features
    • During some operations with a nonzero timeout, queued handlers may run before reactor polling. Deferral is limited to one consecutive turn, and polling proceeds when deferral is not selected.

@cppalliance-bot

cppalliance-bot commented Nov 12, 2026 •

Copy link
Copy Markdown

GCOVR code coverage report https://416.corosio.prtest3.cppalliance.org/gcovr/index.html
LCOV code coverage report https://416.corosio.prtest3.cppalliance.org/genhtml/index.html
Coverage Diff Report https://416.corosio.prtest3.cppalliance.org/diff-report/index.html

Build time: 2026-11-18 19:45:29 UTC

@sgerbino
sgerbino force-pushed the pr/recycling-free-list branch from 7c6bd6d to 1e59c16 Compare November 12, 2026 15:07
@cppalliance-bot

cppalliance-bot commented Nov 12, 2026 •

Copy link
Copy Markdown

An automated preview of the documentation is available at https://416.corosio.prtest3.cppalliance.org/index.html

If more commits are pushed to the pull request, the docs will rebuild at the same URL.

2026-11-18 19:29:19 UTC

@github-actions

github-actions Bot commented Nov 12, 2026 •

Copy link
Copy Markdown

Benchmark report

7cb81d1a76c6 (base) vs 0b4dfacb9b3f (head)

  • ⚠️ linux — 14 flagged (epoll, uring)
  • ✅ windows — clean (iocp)
  • ⚠️ macos — 13 flagged (kqueue)

⚠️ Flagged

Platform Backend Benchmark Metric Base Head Δ Noise
linux epoll local_socket_latency/concurrent/16 latency_mean_ns 104.24 µs 101.69 µs 🟢▲ +2.40% 0.39%
linux epoll local_socket_latency/concurrent_lockless/1 latency_mean_ns 6.69 µs 6.33 µs 🟢▲ +5.54% 0.22%
linux epoll local_socket_latency/concurrent_lockless/16 latency_mean_ns 103.15 µs 100.96 µs 🟢▲ +2.23% 0.49%
linux epoll local_socket_latency/pingpong/64 latency_mean_ns 6.73 µs 6.36 µs 🟢▲ +5.45% 0.22%
linux epoll local_socket_latency/pingpong_lockless/1 latency_mean_ns 6.64 µs 6.29 µs 🟢▲ +5.44% 0.39%
linux epoll local_socket_throughput/unidirectional/1024 bytes_per_sec 303.2 MB/s 313.4 MB/s 🟢▲ +3.34% 0.52%
linux epoll local_socket_throughput/unidirectional_lockless/1024 bytes_per_sec 305.7 MB/s 314.0 MB/s 🟢▲ +2.74% 0.72%
linux epoll local_socket_throughput/unidirectional_lockless/4096 bytes_per_sec 1.00 GB/s 1.03 GB/s 🟢▲ +2.51% 0.47%
linux epoll socket_latency/concurrent/1 latency_mean_ns 14.46 µs 14.09 µs 🟢▲ +2.45% 0.27%
linux epoll socket_latency/pingpong/1024 latency_mean_ns 14.76 µs 14.33 µs 🟢▲ +2.92% 0.22%
linux epoll socket_latency/pingpong/64 latency_mean_ns 14.48 µs 14.06 µs 🟢▲ +2.66% 0.40%
linux epoll socket_latency/pingpong_lockless/1 latency_mean_ns 14.35 µs 13.98 µs 🟢▲ +2.69% 0.23%
linux epoll socket_latency/pingpong_lockless/1024 latency_mean_ns 14.69 µs 14.31 µs 🟢▲ +2.53% 0.37%
linux epoll socket_latency/pingpong_lockless/64 latency_mean_ns 14.41 µs 14.05 µs 🟢▲ +2.64% 0.28%
macos kqueue accept_churn/burst_lockless/10 ops_per_sec 4.62K ops/s 4.73K ops/s 🟢▲ +2.33% 0.52%
macos kqueue local_socket_latency/concurrent/1 latency_mean_ns 1.82 µs 1.71 µs 🟢▲ +6.17% 0.72%
macos kqueue local_socket_latency/concurrent/4 latency_mean_ns 6.81 µs 6.65 µs 🟢▲ +2.51% 0.34%
macos kqueue local_socket_latency/concurrent_lockless/1 latency_mean_ns 1.81 µs 1.69 µs 🟢▲ +6.77% 0.28%
macos kqueue local_socket_latency/concurrent_lockless/4 latency_mean_ns 6.75 µs 6.60 µs 🟢▲ +2.43% 0.32%
macos kqueue local_socket_latency/pingpong/1 latency_mean_ns 1.82 µs 1.71 µs 🟢▲ +6.44% 0.64%
macos kqueue local_socket_latency/pingpong/1024 latency_mean_ns 1.91 µs 1.80 µs 🟢▲ +5.89% 0.48%
macos kqueue local_socket_latency/pingpong/64 latency_mean_ns 1.82 µs 1.71 µs 🟢▲ +6.22% 0.64%
macos kqueue local_socket_latency/pingpong_lockless/1 latency_mean_ns 1.80 µs 1.69 µs 🟢▲ +6.62% 0.55%
macos kqueue local_socket_latency/pingpong_lockless/1024 latency_mean_ns 1.91 µs 1.78 µs 🟢▲ +6.52% 0.51%
macos kqueue local_socket_latency/pingpong_lockless/64 latency_mean_ns 1.81 µs 1.69 µs 🟢▲ +6.88% 0.56%
macos kqueue local_socket_throughput/unidirectional/1024 bytes_per_sec 1.28 GB/s 1.32 GB/s 🟢▲ +2.85% 0.21%
macos kqueue local_socket_throughput/unidirectional_lockless/1024 bytes_per_sec 1.29 GB/s 1.32 GB/s 🟢▲ +3.22% 0.49%
linux — full results (226 benchmarks, 5 iterations, 1.5s each)

accept_churn

Benchmark Backend Metric Base Head Δ Noise
burst/10 epoll ops_per_sec 2.02K ops/s 2.03K ops/s 🟢▲ +0.56% 0.26%
burst/10 uring ops_per_sec 2.42K ops/s 2.41K ops/s ⚪ 0.00% 1.15%
burst/100 epoll ops_per_sec 205.0 ops/s 205.1 ops/s 🔴▼ -0.01% 0.26%
burst/100 uring ops_per_sec 248.2 ops/s 248.4 ops/s 🟢▲ +0.01% 0.50%
burst_lockless/10 epoll ops_per_sec 2.03K ops/s 2.04K ops/s 🟢▲ +0.36% 0.14%
burst_lockless/10 uring ops_per_sec 2.41K ops/s 2.43K ops/s 🟢▲ +0.03% 1.07%
burst_lockless/100 epoll ops_per_sec 206.8 ops/s 207.3 ops/s 🟢▲ +0.25% 0.39%
burst_lockless/100 uring ops_per_sec 249.0 ops/s 249.1 ops/s 🔴▼ -0.04% 0.65%
concurrent/1 epoll ops_per_sec 16.05K ops/s 16.20K ops/s 🟢▲ +1.09% 0.77%
concurrent/1 uring ops_per_sec 17.56K ops/s 17.37K ops/s 🔴▼ -0.45% 1.26%
concurrent/16 epoll ops_per_sec 16.11K ops/s 16.20K ops/s 🟢▲ +0.51% 0.67%
concurrent/16 uring ops_per_sec 19.07K ops/s 18.92K ops/s 🟢▲ +0.10% 1.38%
concurrent/4 epoll ops_per_sec 16.08K ops/s 16.21K ops/s 🟢▲ +1.00% 0.55%
concurrent/4 uring ops_per_sec 18.13K ops/s 17.94K ops/s 🔴▼ -0.01% 1.55%
sequential epoll ops_per_sec 16.14K ops/s 16.23K ops/s 🟢▲ +0.67% 1.28%
sequential uring ops_per_sec 17.52K ops/s 17.37K ops/s 🔴▼ -0.33% 1.14%
sequential_lockless epoll ops_per_sec 16.19K ops/s 16.32K ops/s 🟢▲ +0.59% 0.92%
sequential_lockless uring ops_per_sec 17.71K ops/s 17.48K ops/s 🔴▼ -1.16% 1.20%

fan_out

Benchmark Backend Metric Base Head Δ Noise
concurrent_parents/1 epoll ops_per_sec 3.93K ops/s 3.96K ops/s 🟢▲ +0.67% 0.24%
concurrent_parents/1 uring ops_per_sec 4.06K ops/s 4.06K ops/s 🔴▼ -0.25% 0.30%
concurrent_parents/16 epoll ops_per_sec 3.84K ops/s 3.86K ops/s 🟢▲ +0.22% 0.31%
concurrent_parents/16 uring ops_per_sec 3.98K ops/s 3.98K ops/s 🔴▼ -0.07% 0.22%
concurrent_parents/4 epoll ops_per_sec 3.85K ops/s 3.87K ops/s 🟢▲ +0.47% 0.30%
concurrent_parents/4 uring ops_per_sec 4.02K ops/s 4.02K ops/s 🔴▼ -0.24% 0.27%
concurrent_parents_lockless/1 epoll ops_per_sec 3.95K ops/s 3.97K ops/s 🟢▲ +0.76% 0.23%
concurrent_parents_lockless/1 uring ops_per_sec 4.09K ops/s 4.10K ops/s 🟢▲ +0.10% 0.23%
concurrent_parents_lockless/16 epoll ops_per_sec 3.86K ops/s 3.88K ops/s 🟢▲ +0.40% 0.30%
concurrent_parents_lockless/16 uring ops_per_sec 4.02K ops/s 4.02K ops/s 🟢▲ +0.12% 0.76%
concurrent_parents_lockless/4 epoll ops_per_sec 3.87K ops/s 3.89K ops/s 🟢▲ +0.42% 0.30%
concurrent_parents_lockless/4 uring ops_per_sec 4.06K ops/s 4.07K ops/s 🟢▲ +0.20% 0.89%
fork_join/1 epoll ops_per_sec 57.86K ops/s 57.99K ops/s 🟢▲ +0.60% 1.04%
fork_join/1 uring ops_per_sec 55.75K ops/s 55.45K ops/s 🔴▼ -0.49% 0.28%
fork_join/16 epoll ops_per_sec 3.95K ops/s 3.95K ops/s 🟢▲ +0.52% 1.04%
fork_join/16 uring ops_per_sec 4.07K ops/s 4.07K ops/s 🔴▼ -0.20% 0.77%
fork_join/4 epoll ops_per_sec 15.58K ops/s 15.60K ops/s 🟢▲ +0.61% 1.03%
fork_join/4 uring ops_per_sec 15.29K ops/s 15.26K ops/s 🔴▼ -0.26% 0.29%
fork_join/64 epoll ops_per_sec 966.9 ops/s 969.4 ops/s 🟢▲ +0.69% 0.97%
fork_join/64 uring ops_per_sec 1.01K ops/s 1.01K ops/s 🔴▼ -0.03% 1.10%
fork_join_lockless/1 epoll ops_per_sec 58.43K ops/s 58.43K ops/s 🟢▲ +0.19% 0.98%
fork_join_lockless/1 uring ops_per_sec 56.96K ops/s 56.60K ops/s 🔴▼ -0.60% 1.19%
fork_join_lockless/16 epoll ops_per_sec 3.96K ops/s 3.97K ops/s 🟢▲ +0.19% 0.50%
fork_join_lockless/16 uring ops_per_sec 4.13K ops/s 4.11K ops/s 🔴▼ -0.35% 0.95%
fork_join_lockless/4 epoll ops_per_sec 15.69K ops/s 15.66K ops/s 🔴▼ -0.17% 1.02%
fork_join_lockless/4 uring ops_per_sec 15.53K ops/s 15.47K ops/s 🔴▼ -0.30% 1.18%
fork_join_lockless/64 epoll ops_per_sec 967.3 ops/s 972.2 ops/s 🟢▲ +0.47% 0.23%
fork_join_lockless/64 uring ops_per_sec 1.02K ops/s 1.02K ops/s 🔴▼ -0.03% 0.98%
nested/16 epoll ops_per_sec 960.4 ops/s 964.5 ops/s 🟢▲ +0.39% 0.31%
nested/16 uring ops_per_sec 1.01K ops/s 1.00K ops/s 🔴▼ -0.38% 0.49%
nested/4 epoll ops_per_sec 3.91K ops/s 3.93K ops/s 🟢▲ +0.46% 0.19%
nested/4 uring ops_per_sec 4.07K ops/s 4.05K ops/s 🔴▼ -0.53% 1.08%
nested_lockless/16 epoll ops_per_sec 965.1 ops/s 969.5 ops/s 🟢▲ +0.43% 0.39%
nested_lockless/16 uring ops_per_sec 1.01K ops/s 1.01K ops/s 🔴▼ -0.30% 0.30%
nested_lockless/4 epoll ops_per_sec 3.93K ops/s 3.95K ops/s 🟢▲ +1.00% 0.35%
nested_lockless/4 uring ops_per_sec 4.09K ops/s 4.07K ops/s 🔴▼ -0.47% 0.24%

http_server

Benchmark Backend Metric Base Head Δ Noise
concurrent/1 epoll ops_per_sec 59.13K ops/s 59.10K ops/s 🟢▲ +0.10% 0.26%
concurrent/1 uring ops_per_sec 55.55K ops/s 55.45K ops/s 🟢▲ +0.28% 1.24%
concurrent/16 epoll ops_per_sec 62.04K ops/s 62.03K ops/s 🟢▲ +0.05% 0.43%
concurrent/16 uring ops_per_sec 65.51K ops/s 65.33K ops/s 🟢▲ +0.14% 1.06%
concurrent/32 epoll ops_per_sec 61.56K ops/s 61.53K ops/s 🔴▼ -0.04% 0.25%
concurrent/32 uring ops_per_sec 65.39K ops/s 65.13K ops/s 🔴▼ -0.15% 0.83%
concurrent/4 epoll ops_per_sec 61.84K ops/s 61.69K ops/s 🔴▼ -0.23% 0.20%
concurrent/4 uring ops_per_sec 63.43K ops/s 63.16K ops/s 🔴▼ -0.19% 1.00%
multithread/1 epoll ops_per_sec 61.49K ops/s 61.56K ops/s 🟢▲ +0.19% 0.28%
multithread/1 uring ops_per_sec 65.34K ops/s 65.31K ops/s 🟢▲ +0.20% 1.10%
multithread/16 epoll ops_per_sec 172.6K ops/s 172.8K ops/s 🔴▼ -0.22% 0.67%
multithread/16 uring ops_per_sec 104.0K ops/s 103.4K ops/s 🔴▼ -0.23% 0.58%
multithread/2 epoll ops_per_sec 94.65K ops/s 94.41K ops/s 🔴▼ -0.08% 0.67%
multithread/2 uring ops_per_sec 78.83K ops/s 80.03K ops/s 🟢▲ +0.44% 2.60%
multithread/4 epoll ops_per_sec 169.5K ops/s 170.1K ops/s 🔴▼ -0.02% 1.27%
multithread/4 uring ops_per_sec 124.6K ops/s 124.8K ops/s 🟢▲ +0.38% 0.97%
multithread/8 epoll ops_per_sec 172.0K ops/s 172.1K ops/s 🟢▲ +0.36% 0.58%
multithread/8 uring ops_per_sec 116.2K ops/s 116.3K ops/s 🟢▲ +0.25% 0.40%
single_conn epoll ops_per_sec 59.31K ops/s 59.17K ops/s 🟢▲ +0.29% 1.00%
single_conn uring ops_per_sec 55.63K ops/s 55.49K ops/s 🟢▲ +0.13% 1.09%
single_conn_lockless epoll ops_per_sec 59.72K ops/s 59.53K ops/s 🟢▲ +0.10% 0.92%
single_conn_lockless uring ops_per_sec 56.69K ops/s 56.49K ops/s 🟢▲ +0.11% 1.09%

local_socket_latency

Benchmark Backend Metric Base Head Δ Noise
concurrent/1 epoll latency_mean_ns 6.79 µs 6.72 µs 🔴▼ -0.89% 5.06%
concurrent/1 uring latency_mean_ns 5.29 µs 5.43 µs 🔴▼ -0.28% 5.30%
concurrent/16 epoll latency_mean_ns 104.24 µs 101.69 µs 🟢▲ +2.40% 0.39%
concurrent/16 uring latency_mean_ns 87.15 µs 87.02 µs 🔴▼ -0.12% 0.81%
concurrent/4 epoll latency_mean_ns 26.47 µs 25.46 µs 🟢▲ +3.44% 1.31%
concurrent/4 uring latency_mean_ns 21.36 µs 22.02 µs 🔴▼ -3.84% 1.50%
concurrent_lockless/1 epoll latency_mean_ns 6.69 µs 6.33 µs 🟢▲ +5.54% 0.22%
concurrent_lockless/1 uring latency_mean_ns 5.40 µs 5.67 µs 🔴▼ -1.08% 6.04%
concurrent_lockless/16 epoll latency_mean_ns 103.15 µs 100.96 µs 🟢▲ +2.23% 0.49%
concurrent_lockless/16 uring latency_mean_ns 86.67 µs 87.15 µs 🔴▼ -0.68% 0.88%
concurrent_lockless/4 epoll latency_mean_ns 25.96 µs 25.26 µs 🟢▲ +2.87% 1.60%
concurrent_lockless/4 uring latency_mean_ns 21.59 µs 21.87 µs 🔴▼ -0.56% 1.41%
pingpong/1 epoll latency_mean_ns 7.09 µs 6.57 µs 🟢▲ +5.83% 5.60%
pingpong/1 uring latency_mean_ns 5.25 µs 5.30 µs 🟢▲ +0.02% 3.77%
pingpong/1024 epoll latency_mean_ns 7.36 µs 6.99 µs 🟢▲ +4.91% 5.15%
pingpong/1024 uring latency_mean_ns 5.62 µs 5.94 µs 🔴▼ -7.61% 6.29%
pingpong/64 epoll latency_mean_ns 6.73 µs 6.36 µs 🟢▲ +5.45% 0.22%
pingpong/64 uring latency_mean_ns 5.56 µs 5.43 µs 🟢▲ +0.22% 6.49%
pingpong_lockless/1 epoll latency_mean_ns 6.64 µs 6.29 µs 🟢▲ +5.44% 0.39%
pingpong_lockless/1 uring latency_mean_ns 5.37 µs 5.50 µs 🔴▼ -0.19% 6.48%
pingpong_lockless/1024 epoll latency_mean_ns 7.47 µs 6.96 µs 🟢▲ +5.26% 5.25%
pingpong_lockless/1024 uring latency_mean_ns 6.18 µs 5.93 µs 🟢▲ +0.33% 6.19%
pingpong_lockless/64 epoll latency_mean_ns 6.95 µs 6.46 µs 🟢▲ +5.42% 5.16%
pingpong_lockless/64 uring latency_mean_ns 5.27 µs 5.42 µs 🔴▼ -0.37% 5.66%

local_socket_throughput

Benchmark Backend Metric Base Head Δ Noise
bidirectional/1024 epoll bytes_per_sec 312.3 MB/s 319.1 MB/s 🟢▲ +1.83% 1.23%
bidirectional/1024 uring bytes_per_sec 363.0 MB/s 361.5 MB/s 🔴▼ -0.15% 0.94%
bidirectional/1048576 epoll bytes_per_sec 3.25 GB/s 3.19 GB/s 🟢▲ +0.33% 5.02%
bidirectional/1048576 uring bytes_per_sec 3.56 GB/s 3.58 GB/s 🟢▲ +0.54% 2.63%
bidirectional/16384 epoll bytes_per_sec 3.10 GB/s 3.18 GB/s 🟢▲ +1.48% 2.47%
bidirectional/16384 uring bytes_per_sec 3.03 GB/s 3.08 GB/s 🟢▲ +1.74% 2.00%
bidirectional/262144 epoll bytes_per_sec 3.25 GB/s 3.23 GB/s 🔴▼ -0.17% 2.77%
bidirectional/262144 uring bytes_per_sec 3.54 GB/s 3.58 GB/s 🔴▼ -0.31% 2.82%
bidirectional/4096 epoll bytes_per_sec 1.02 GB/s 1.04 GB/s 🟢▲ +1.82% 0.65%
bidirectional/4096 uring bytes_per_sec 1.15 GB/s 1.15 GB/s 🟢▲ +0.08% 0.53%
bidirectional/65536 epoll bytes_per_sec 3.03 GB/s 3.01 GB/s 🔴▼ -0.58% 1.00%
bidirectional/65536 uring bytes_per_sec 3.49 GB/s 3.52 GB/s 🔴▼ -0.43% 3.24%
bidirectional_lockless/1024 epoll bytes_per_sec 314.8 MB/s 318.3 MB/s 🟢▲ +1.07% 0.58%
bidirectional_lockless/1024 uring bytes_per_sec 362.7 MB/s 361.9 MB/s 🔴▼ -0.29% 0.72%
bidirectional_lockless/1048576 epoll bytes_per_sec 3.27 GB/s 3.21 GB/s 🔴▼ -0.33% 3.33%
bidirectional_lockless/1048576 uring bytes_per_sec 3.50 GB/s 3.57 GB/s 🟢▲ +2.69% 3.26%
bidirectional_lockless/16384 epoll bytes_per_sec 3.12 GB/s 3.16 GB/s 🟢▲ +2.15% 2.78%
bidirectional_lockless/16384 uring bytes_per_sec 3.07 GB/s 3.09 GB/s 🟢▲ +2.02% 2.25%
bidirectional_lockless/262144 epoll bytes_per_sec 3.27 GB/s 3.24 GB/s 🔴▼ -0.66% 2.63%
bidirectional_lockless/262144 uring bytes_per_sec 3.49 GB/s 3.59 GB/s 🟢▲ +1.04% 4.20%
bidirectional_lockless/4096 epoll bytes_per_sec 1.03 GB/s 1.04 GB/s 🟢▲ +0.70% 0.22%
bidirectional_lockless/4096 uring bytes_per_sec 1.15 GB/s 1.15 GB/s 🟢▲ +0.04% 0.92%
bidirectional_lockless/65536 epoll bytes_per_sec 3.05 GB/s 3.02 GB/s 🔴▼ -1.12% 0.72%
bidirectional_lockless/65536 uring bytes_per_sec 3.47 GB/s 3.52 GB/s 🟢▲ +0.64% 3.48%
unidirectional/1024 epoll bytes_per_sec 303.2 MB/s 313.4 MB/s 🟢▲ +3.34% 0.52%
unidirectional/1024 uring bytes_per_sec 361.7 MB/s 360.6 MB/s 🟢▲ +0.04% 0.49%
unidirectional/1048576 epoll bytes_per_sec 3.33 GB/s 3.29 GB/s 🔴▼ -1.15% 4.42%
unidirectional/1048576 uring bytes_per_sec 3.58 GB/s 3.55 GB/s 🔴▼ -0.33% 2.12%
unidirectional/16384 epoll bytes_per_sec 3.09 GB/s 3.12 GB/s 🟢▲ +1.73% 4.23%
unidirectional/16384 uring bytes_per_sec 3.05 GB/s 3.04 GB/s 🔴▼ -0.48% 2.53%
unidirectional/262144 epoll bytes_per_sec 3.36 GB/s 3.30 GB/s 🔴▼ -0.39% 2.75%
unidirectional/262144 uring bytes_per_sec 3.54 GB/s 3.58 GB/s 🟢▲ +0.40% 3.25%
unidirectional/4096 epoll bytes_per_sec 996.5 MB/s 1.03 GB/s 🟢▲ +2.75% 1.00%
unidirectional/4096 uring bytes_per_sec 1.15 GB/s 1.15 GB/s 🔴▼ -0.48% 0.74%
unidirectional/65536 epoll bytes_per_sec 3.31 GB/s 3.28 GB/s 🔴▼ -0.58% 1.63%
unidirectional/65536 uring bytes_per_sec 3.49 GB/s 3.52 GB/s 🟢▲ +0.19% 3.15%
unidirectional_lockless/1024 epoll bytes_per_sec 305.7 MB/s 314.0 MB/s 🟢▲ +2.74% 0.72%
unidirectional_lockless/1024 uring bytes_per_sec 364.0 MB/s 364.5 MB/s 🟢▲ +0.32% 1.23%
unidirectional_lockless/1048576 epoll bytes_per_sec 3.33 GB/s 3.30 GB/s 🔴▼ -1.57% 3.92%
unidirectional_lockless/1048576 uring bytes_per_sec 3.57 GB/s 3.57 GB/s 🟢▲ +0.98% 2.74%
unidirectional_lockless/16384 epoll bytes_per_sec 3.07 GB/s 3.18 GB/s 🟢▲ +3.77% 2.31%
unidirectional_lockless/16384 uring bytes_per_sec 3.07 GB/s 3.09 GB/s 🟢▲ +0.89% 2.50%
unidirectional_lockless/262144 epoll bytes_per_sec 3.37 GB/s 3.32 GB/s 🔴▼ -0.01% 2.91%
unidirectional_lockless/262144 uring bytes_per_sec 3.53 GB/s 3.61 GB/s 🟢▲ +0.39% 3.00%
unidirectional_lockless/4096 epoll bytes_per_sec 1.00 GB/s 1.03 GB/s 🟢▲ +2.51% 0.47%
unidirectional_lockless/4096 uring bytes_per_sec 1.15 GB/s 1.16 GB/s 🔴▼ -0.11% 1.02%
unidirectional_lockless/65536 epoll bytes_per_sec 3.33 GB/s 3.33 GB/s 🟢▲ +1.24% 3.38%
unidirectional_lockless/65536 uring bytes_per_sec 3.55 GB/s 3.55 GB/s 🔴▼ -0.75% 2.30%

socket_latency

Benchmark Backend Metric Base Head Δ Noise
concurrent/1 epoll latency_mean_ns 14.46 µs 14.09 µs 🟢▲ +2.45% 0.27%
concurrent/1 uring latency_mean_ns 13.21 µs 13.25 µs 🔴▼ -0.37% 1.29%
concurrent/16 epoll latency_mean_ns 227.45 µs 225.42 µs 🟢▲ +1.17% 1.09%
concurrent/16 uring latency_mean_ns 213.36 µs 214.36 µs 🔴▼ -0.15% 1.24%
concurrent/4 epoll latency_mean_ns 56.86 µs 56.10 µs 🟢▲ +1.37% 0.31%
concurrent/4 uring latency_mean_ns 53.01 µs 53.20 µs 🔴▼ -0.19% 1.13%
concurrent_lockless/1 epoll latency_mean_ns 14.34 µs 14.06 µs 🟢▲ +2.35% 1.17%
concurrent_lockless/1 uring latency_mean_ns 13.19 µs 13.23 µs 🔴▼ -0.13% 1.04%
concurrent_lockless/16 epoll latency_mean_ns 226.51 µs 224.58 µs 🟢▲ +1.34% 1.00%
concurrent_lockless/16 uring latency_mean_ns 214.13 µs 214.47 µs 🟢▲ +0.05% 1.13%
concurrent_lockless/4 epoll latency_mean_ns 56.53 µs 55.93 µs 🟢▲ +1.53% 1.06%
concurrent_lockless/4 uring latency_mean_ns 53.08 µs 53.13 µs 🔴▼ -0.16% 1.03%
pingpong/1 epoll latency_mean_ns 14.36 µs 13.95 µs 🟢▲ +2.75% 1.00%
pingpong/1 uring latency_mean_ns 13.21 µs 13.25 µs 🟢▲ +0.26% 1.15%
pingpong/1024 epoll latency_mean_ns 14.76 µs 14.33 µs 🟢▲ +2.92% 0.22%
pingpong/1024 uring latency_mean_ns 13.51 µs 13.56 µs 🔴▼ -0.05% 1.00%
pingpong/64 epoll latency_mean_ns 14.48 µs 14.06 µs 🟢▲ +2.66% 0.40%
pingpong/64 uring latency_mean_ns 13.28 µs 13.32 µs 🟢▲ +0.16% 1.14%
pingpong_lockless/1 epoll latency_mean_ns 14.35 µs 13.98 µs 🟢▲ +2.69% 0.23%
pingpong_lockless/1 uring latency_mean_ns 13.17 µs 13.22 µs 🔴▼ -0.11% 0.90%
pingpong_lockless/1024 epoll latency_mean_ns 14.69 µs 14.31 µs 🟢▲ +2.53% 0.37%
pingpong_lockless/1024 uring latency_mean_ns 13.43 µs 13.46 µs 🔴▼ -0.09% 1.24%
pingpong_lockless/64 epoll latency_mean_ns 14.41 µs 14.05 µs 🟢▲ +2.64% 0.28%
pingpong_lockless/64 uring latency_mean_ns 13.20 µs 13.30 µs 🔴▼ -0.54% 1.16%

socket_throughput

Benchmark Backend Metric Base Head Δ Noise
bidirectional/1024 epoll bytes_per_sec 135.9 MB/s 137.0 MB/s 🟢▲ +0.79% 1.24%
bidirectional/1024 uring bytes_per_sec 152.5 MB/s 151.6 MB/s 🔴▼ -0.18% 1.12%
bidirectional/1048576 epoll bytes_per_sec 3.36 GB/s 3.35 GB/s 🔴▼ -1.05% 2.85%
bidirectional/1048576 uring bytes_per_sec 3.38 GB/s 3.36 GB/s 🔴▼ -0.33% 1.01%
bidirectional/16384 epoll bytes_per_sec 1.65 GB/s 1.66 GB/s 🟢▲ +0.05% 1.22%
bidirectional/16384 uring bytes_per_sec 1.64 GB/s 1.64 GB/s 🟢▲ +0.40% 0.87%
bidirectional/262144 epoll bytes_per_sec 3.26 GB/s 3.25 GB/s 🔴▼ -0.98% 0.69%
bidirectional/262144 uring bytes_per_sec 3.28 GB/s 3.26 GB/s 🟢▲ +0.05% 1.18%
bidirectional/4096 epoll bytes_per_sec 512.1 MB/s 515.1 MB/s 🟢▲ +0.57% 1.64%
bidirectional/4096 uring bytes_per_sec 530.9 MB/s 528.3 MB/s 🔴▼ -0.53% 1.06%
bidirectional/65536 epoll bytes_per_sec 2.57 GB/s 2.58 GB/s 🟢▲ +1.70% 1.73%
bidirectional/65536 uring bytes_per_sec 2.88 GB/s 2.86 GB/s 🔴▼ -0.38% 2.22%
bidirectional_lockless/1024 epoll bytes_per_sec 135.3 MB/s 136.3 MB/s 🟢▲ +0.84% 0.15%
bidirectional_lockless/1024 uring bytes_per_sec 152.0 MB/s 152.1 MB/s 🟢▲ +0.18% 0.27%
bidirectional_lockless/1048576 epoll bytes_per_sec 3.30 GB/s 3.38 GB/s 🟢▲ +1.99% 1.50%
bidirectional_lockless/1048576 uring bytes_per_sec 3.31 GB/s 3.31 GB/s 🔴▼ -0.92% 1.40%
bidirectional_lockless/16384 epoll bytes_per_sec 1.65 GB/s 1.63 GB/s 🔴▼ -1.14% 1.85%
bidirectional_lockless/16384 uring bytes_per_sec 1.58 GB/s 1.58 GB/s 🔴▼ -0.05% 0.30%
bidirectional_lockless/262144 epoll bytes_per_sec 3.24 GB/s 3.25 GB/s 🟢▲ +0.55% 1.19%
bidirectional_lockless/262144 uring bytes_per_sec 3.23 GB/s 3.21 GB/s 🔴▼ -1.26% 1.23%
bidirectional_lockless/4096 epoll bytes_per_sec 509.4 MB/s 513.1 MB/s 🟢▲ +0.99% 0.51%
bidirectional_lockless/4096 uring bytes_per_sec 526.3 MB/s 526.7 MB/s 🟢▲ +0.04% 0.67%
bidirectional_lockless/65536 epoll bytes_per_sec 2.56 GB/s 2.59 GB/s 🟢▲ +1.09% 0.77%
bidirectional_lockless/65536 uring bytes_per_sec 2.79 GB/s 2.75 GB/s 🔴▼ -0.98% 1.99%
multithread/2 epoll bytes_per_sec 3.63 GB/s 3.65 GB/s 🟢▲ +0.29% 1.47%
multithread/2 uring bytes_per_sec 4.23 GB/s 4.21 GB/s 🔴▼ -0.34% 0.76%
multithread/4 epoll bytes_per_sec 5.78 GB/s 5.93 GB/s 🟢▲ +1.85% 1.66%
multithread/4 uring bytes_per_sec 6.75 GB/s 6.75 GB/s 🟢▲ +0.18% 0.69%
multithread/8 epoll bytes_per_sec 6.36 GB/s 6.43 GB/s 🟢▲ +1.26% 1.19%
multithread/8 uring bytes_per_sec 6.89 GB/s 6.89 GB/s 🟢▲ +0.28% 0.47%
unidirectional/1024 epoll bytes_per_sec 127.4 MB/s 128.3 MB/s 🟢▲ +1.14% 1.09%
unidirectional/1024 uring bytes_per_sec 150.2 MB/s 150.3 MB/s 🟢▲ +0.13% 0.15%
unidirectional/1048576 epoll bytes_per_sec 3.36 GB/s 3.38 GB/s 🟢▲ +0.62% 1.08%
unidirectional/1048576 uring bytes_per_sec 3.41 GB/s 3.41 GB/s 🟢▲ +0.56% 0.66%
unidirectional/16384 epoll bytes_per_sec 1.56 GB/s 1.57 GB/s 🔴▼ -0.63% 1.43%
unidirectional/16384 uring bytes_per_sec 1.73 GB/s 1.72 GB/s 🔴▼ -0.58% 0.81%
unidirectional/262144 epoll bytes_per_sec 3.37 GB/s 3.35 GB/s 🔴▼ -1.03% 0.80%
unidirectional/262144 uring bytes_per_sec 3.24 GB/s 3.27 GB/s 🟢▲ +0.79% 1.18%
unidirectional/4096 epoll bytes_per_sec 478.0 MB/s 481.4 MB/s 🟢▲ +0.94% 0.76%
unidirectional/4096 uring bytes_per_sec 527.4 MB/s 527.3 MB/s 🟢▲ +0.07% 1.31%
unidirectional/65536 epoll bytes_per_sec 2.74 GB/s 2.76 GB/s 🔴▼ -0.94% 4.12%
unidirectional/65536 uring bytes_per_sec 2.81 GB/s 2.84 GB/s 🟢▲ +0.62% 2.50%
unidirectional_lockless/1024 epoll bytes_per_sec 127.0 MB/s 128.1 MB/s 🟢▲ +0.84% 0.17%
unidirectional_lockless/1024 uring bytes_per_sec 151.2 MB/s 150.1 MB/s 🔴▼ -0.13% 0.94%
unidirectional_lockless/1048576 epoll bytes_per_sec 3.28 GB/s 3.37 GB/s 🟢▲ +3.08% 1.52%
unidirectional_lockless/1048576 uring bytes_per_sec 3.43 GB/s 3.40 GB/s 🔴▼ -0.94% 0.99%
unidirectional_lockless/16384 epoll bytes_per_sec 1.56 GB/s 1.58 GB/s 🟢▲ +0.74% 1.50%
unidirectional_lockless/16384 uring bytes_per_sec 1.74 GB/s 1.70 GB/s 🔴▼ -1.31% 2.65%
unidirectional_lockless/262144 epoll bytes_per_sec 3.35 GB/s 3.38 GB/s 🟢▲ +0.95% 0.89%
unidirectional_lockless/262144 uring bytes_per_sec 3.32 GB/s 3.29 GB/s 🔴▼ -0.48% 1.16%
unidirectional_lockless/4096 epoll bytes_per_sec 476.2 MB/s 481.4 MB/s 🟢▲ +1.06% 0.46%
unidirectional_lockless/4096 uring bytes_per_sec 528.1 MB/s 524.0 MB/s 🔴▼ -0.55% 1.17%
unidirectional_lockless/65536 epoll bytes_per_sec 2.79 GB/s 2.81 GB/s 🟢▲ +0.97% 4.15%
unidirectional_lockless/65536 uring bytes_per_sec 2.84 GB/s 2.82 GB/s 🔴▼ -0.33% 1.32%
windows — full results (77 benchmarks, 5 iterations, 1.5s each)

accept_churn

Benchmark Backend Metric Base Head Δ Noise
burst/10 iocp ops_per_sec 1.56K ops/s 1.55K ops/s 🔴▼ -0.23% 0.53%
burst/100 iocp ops_per_sec 156.6 ops/s 155.7 ops/s 🔴▼ -0.82% 0.85%
burst_lockless/10 iocp ops_per_sec 1.57K ops/s 1.56K ops/s 🔴▼ -0.03% 0.25%
burst_lockless/100 iocp ops_per_sec 158.5 ops/s 159.0 ops/s 🟢▲ +0.25% 0.94%
concurrent/1 iocp ops_per_sec 12.63K ops/s 12.56K ops/s 🔴▼ -0.51% 0.45%
concurrent/16 iocp ops_per_sec 13.48K ops/s 13.43K ops/s 🔴▼ -0.19% 0.23%
concurrent/4 iocp ops_per_sec 13.20K ops/s 13.14K ops/s 🔴▼ -0.34% 0.39%
sequential iocp ops_per_sec 12.56K ops/s 12.54K ops/s 🔴▼ -0.19% 0.38%
sequential_lockless iocp ops_per_sec 12.58K ops/s 12.59K ops/s 🔴▼ -0.14% 0.26%

fan_out

Benchmark Backend Metric Base Head Δ Noise
concurrent_parents/1 iocp ops_per_sec 7.37K ops/s 7.34K ops/s 🔴▼ -0.22% 0.31%
concurrent_parents/16 iocp ops_per_sec 6.83K ops/s 6.83K ops/s 🟢▲ +0.01% 1.19%
concurrent_parents/4 iocp ops_per_sec 7.29K ops/s 7.28K ops/s 🔴▼ -0.08% 0.83%
concurrent_parents_lockless/1 iocp ops_per_sec 7.38K ops/s 7.38K ops/s 🟢▲ +0.14% 0.24%
concurrent_parents_lockless/16 iocp ops_per_sec 6.84K ops/s 6.85K ops/s 🟢▲ +0.10% 1.40%
concurrent_parents_lockless/4 iocp ops_per_sec 7.29K ops/s 7.30K ops/s 🔴▼ -0.08% 0.75%
fork_join/1 iocp ops_per_sec 112.9K ops/s 112.7K ops/s 🔴▼ -0.33% 0.59%
fork_join/16 iocp ops_per_sec 7.36K ops/s 7.36K ops/s 🔴▼ -0.11% 0.31%
fork_join/4 iocp ops_per_sec 29.27K ops/s 29.21K ops/s 🔴▼ -0.17% 0.32%
fork_join/64 iocp ops_per_sec 1.82K ops/s 1.83K ops/s 🟢▲ +0.12% 1.64%
fork_join_lockless/1 iocp ops_per_sec 112.8K ops/s 113.4K ops/s 🟢▲ +0.54% 0.18%
fork_join_lockless/16 iocp ops_per_sec 7.35K ops/s 7.36K ops/s 🟢▲ +0.09% 0.38%
fork_join_lockless/4 iocp ops_per_sec 29.16K ops/s 29.24K ops/s 🟢▲ +0.23% 0.23%
fork_join_lockless/64 iocp ops_per_sec 1.82K ops/s 1.82K ops/s 🔴▼ -0.21% 1.62%
nested/16 iocp ops_per_sec 1.79K ops/s 1.80K ops/s 🟢▲ +0.15% 1.00%
nested/4 iocp ops_per_sec 7.24K ops/s 7.23K ops/s 🔴▼ -0.23% 0.42%
nested_lockless/16 iocp ops_per_sec 1.79K ops/s 1.80K ops/s 🔴▼ -0.02% 1.04%
nested_lockless/4 iocp ops_per_sec 7.24K ops/s 7.22K ops/s 🔴▼ -0.16% 0.24%

http_server

Benchmark Backend Metric Base Head Δ Noise
concurrent/1 iocp ops_per_sec 111.5K ops/s 111.7K ops/s 🟢▲ +0.26% 0.44%
concurrent/16 iocp ops_per_sec 108.7K ops/s 108.9K ops/s 🟢▲ +0.40% 0.41%
concurrent/32 iocp ops_per_sec 108.0K ops/s 108.3K ops/s 🟢▲ +0.22% 0.34%
concurrent/4 iocp ops_per_sec 109.4K ops/s 109.5K ops/s 🟢▲ +0.14% 0.21%
multithread/1 iocp ops_per_sec 107.9K ops/s 108.0K ops/s 🔴▼ -0.15% 0.34%
multithread/16 iocp ops_per_sec 264.8K ops/s 264.4K ops/s 🔴▼ -0.58% 0.87%
multithread/2 iocp ops_per_sec 180.8K ops/s 180.5K ops/s 🔴▼ -0.05% 0.39%
multithread/4 iocp ops_per_sec 223.8K ops/s 224.6K ops/s 🟢▲ +0.23% 0.58%
multithread/8 iocp ops_per_sec 264.9K ops/s 265.2K ops/s 🟢▲ +0.38% 0.80%
single_conn iocp ops_per_sec 111.4K ops/s 111.7K ops/s 🟢▲ +0.24% 0.42%
single_conn_lockless iocp ops_per_sec 111.9K ops/s 112.0K ops/s 🔴▼ -0.14% 0.42%

socket_latency

Benchmark Backend Metric Base Head Δ Noise
concurrent/1 iocp latency_mean_ns 7.59 µs 7.59 µs 🔴▼ -0.04% 0.52%
concurrent/16 iocp latency_mean_ns 129.50 µs 130.89 µs 🔴▼ -0.08% 4.25%
concurrent/4 iocp latency_mean_ns 30.45 µs 30.67 µs 🔴▼ -0.60% 0.38%
concurrent_lockless/1 iocp latency_mean_ns 7.59 µs 7.58 µs 🟢▲ +0.04% 0.48%
concurrent_lockless/16 iocp latency_mean_ns 121.96 µs 122.33 µs 🟢▲ +0.03% 1.67%
concurrent_lockless/4 iocp latency_mean_ns 30.40 µs 30.47 µs 🔴▼ -0.22% 0.44%
pingpong/1 iocp latency_mean_ns 7.58 µs 7.57 µs 🔴▼ -0.14% 0.43%
pingpong/1024 iocp latency_mean_ns 7.68 µs 7.69 µs 🔴▼ -0.27% 0.74%
pingpong/64 iocp latency_mean_ns 7.57 µs 7.58 µs 🔴▼ -0.41% 0.49%
pingpong_lockless/1 iocp latency_mean_ns 7.55 µs 7.55 µs 🔴▼ -0.13% 0.25%
pingpong_lockless/1024 iocp latency_mean_ns 7.69 µs 7.70 µs 🟢▲ +0.16% 0.51%
pingpong_lockless/64 iocp latency_mean_ns 7.57 µs 7.59 µs 🔴▼ -0.34% 0.42%

socket_throughput

Benchmark Backend Metric Base Head Δ Noise
bidirectional/1024 iocp bytes_per_sec 281.1 MB/s 280.8 MB/s 🔴▼ -0.23% 0.52%
bidirectional/1048576 iocp bytes_per_sec 4.58 GB/s 5.16 GB/s 🟢▲ +14.40% 8.88%
bidirectional/16384 iocp bytes_per_sec 3.76 GB/s 3.73 GB/s 🟢▲ +0.06% 0.75%
bidirectional/262144 iocp bytes_per_sec 9.46 GB/s 9.36 GB/s 🔴▼ -0.69% 1.73%
bidirectional/4096 iocp bytes_per_sec 1.08 GB/s 1.08 GB/s 🔴▼ -0.35% 0.30%
bidirectional/65536 iocp bytes_per_sec 9.07 GB/s 9.02 GB/s 🔴▼ -0.12% 1.55%
bidirectional_lockless/1024 iocp bytes_per_sec 281.3 MB/s 280.4 MB/s 🔴▼ -0.39% 0.56%
bidirectional_lockless/1048576 iocp bytes_per_sec 4.88 GB/s 5.29 GB/s 🟢▲ +6.12% 11.64%
bidirectional_lockless/16384 iocp bytes_per_sec 3.73 GB/s 3.75 GB/s 🟢▲ +0.99% 0.78%
bidirectional_lockless/262144 iocp bytes_per_sec 9.49 GB/s 9.46 GB/s 🔴▼ -0.94% 2.65%
bidirectional_lockless/4096 iocp bytes_per_sec 1.08 GB/s 1.08 GB/s 🟢▲ +0.31% 0.35%
bidirectional_lockless/65536 iocp bytes_per_sec 8.96 GB/s 9.01 GB/s 🟢▲ +0.63% 1.30%
multithread/2 iocp bytes_per_sec 12.54 GB/s 12.47 GB/s 🔴▼ -0.71% 1.20%
multithread/4 iocp bytes_per_sec 16.80 GB/s 16.83 GB/s 🔴▼ -0.48% 1.29%
multithread/8 iocp bytes_per_sec 18.36 GB/s 18.40 GB/s 🔴▼ -0.11% 0.65%
unidirectional/1024 iocp bytes_per_sec 283.8 MB/s 282.9 MB/s 🔴▼ -0.04% 0.38%
unidirectional/1048576 iocp bytes_per_sec 4.16 GB/s 4.16 GB/s 🟢▲ +2.41% 4.24%
unidirectional/16384 iocp bytes_per_sec 3.74 GB/s 3.78 GB/s 🔴▼ -0.38% 3.16%
unidirectional/262144 iocp bytes_per_sec 8.76 GB/s 8.55 GB/s 🔴▼ -0.87% 4.63%
unidirectional/4096 iocp bytes_per_sec 1.09 GB/s 1.09 GB/s 🔴▼ -0.18% 0.57%
unidirectional/65536 iocp bytes_per_sec 9.04 GB/s 9.23 GB/s 🟢▲ +3.07% 2.27%
unidirectional_lockless/1024 iocp bytes_per_sec 284.5 MB/s 283.9 MB/s 🔴▼ -0.15% 0.75%
unidirectional_lockless/1048576 iocp bytes_per_sec 4.39 GB/s 4.62 GB/s 🟢▲ +2.80% 4.66%
unidirectional_lockless/16384 iocp bytes_per_sec 3.81 GB/s 3.77 GB/s 🔴▼ -0.62% 1.70%
unidirectional_lockless/262144 iocp bytes_per_sec 9.54 GB/s 9.56 GB/s 🟢▲ +0.01% 1.83%
unidirectional_lockless/4096 iocp bytes_per_sec 1.09 GB/s 1.09 GB/s 🟢▲ +0.06% 0.62%
unidirectional_lockless/65536 iocp bytes_per_sec 9.34 GB/s 9.11 GB/s 🔴▼ -0.62% 3.60%
macos — full results (113 benchmarks, 5 iterations, 1.5s each)

accept_churn

Benchmark Backend Metric Base Head Δ Noise
burst/10 kqueue ops_per_sec 4.59K ops/s 4.71K ops/s 🟢▲ +2.80% 0.94%
burst/100 kqueue ops_per_sec 466.8 ops/s 468.6 ops/s 🟢▲ +0.26% 2.09%
burst_lockless/10 kqueue ops_per_sec 4.62K ops/s 4.73K ops/s 🟢▲ +2.33% 0.52%
burst_lockless/100 kqueue ops_per_sec 470.6 ops/s 469.3 ops/s 🔴▼ -0.14% 0.79%
concurrent/1 kqueue ops_per_sec 20.81K ops/s 20.60K ops/s 🔴▼ -1.06% 0.96%
concurrent/16 kqueue ops_per_sec 37.46K ops/s 37.74K ops/s 🔴▼ -0.01% 1.72%
concurrent/4 kqueue ops_per_sec 29.29K ops/s 29.35K ops/s 🟢▲ +0.40% 0.56%
sequential kqueue ops_per_sec 21.00K ops/s 21.01K ops/s 🟢▲ +0.53% 1.26%
sequential_lockless kqueue ops_per_sec 20.90K ops/s 20.69K ops/s 🔴▼ -1.27% 1.29%

fan_out

Benchmark Backend Metric Base Head Δ Noise
concurrent_parents/1 kqueue ops_per_sec 8.14K ops/s 8.15K ops/s 🟢▲ +0.33% 1.52%
concurrent_parents/16 kqueue ops_per_sec 7.11K ops/s 7.29K ops/s 🔴▼ -7.53% 21.56%
concurrent_parents/4 kqueue ops_per_sec 8.39K ops/s 8.31K ops/s 🔴▼ -0.20% 1.91%
concurrent_parents_lockless/1 kqueue ops_per_sec 8.24K ops/s 8.26K ops/s 🟢▲ +1.13% 1.29%
concurrent_parents_lockless/16 kqueue ops_per_sec 7.16K ops/s 7.11K ops/s 🔴▼ -1.06% 22.51%
concurrent_parents_lockless/4 kqueue ops_per_sec 8.48K ops/s 8.48K ops/s 🔴▼ -0.13% 2.25%
fork_join/1 kqueue ops_per_sec 49.12K ops/s 49.42K ops/s 🟢▲ +0.84% 0.81%
fork_join/16 kqueue ops_per_sec 8.15K ops/s 8.14K ops/s 🟢▲ +0.51% 1.44%
fork_join/4 kqueue ops_per_sec 23.59K ops/s 23.64K ops/s 🔴▼ -0.18% 0.78%
fork_join/64 kqueue ops_per_sec 2.12K ops/s 2.12K ops/s 🔴▼ -0.31% 2.40%
fork_join_lockless/1 kqueue ops_per_sec 50.34K ops/s 50.50K ops/s 🟢▲ +0.45% 0.47%
fork_join_lockless/16 kqueue ops_per_sec 8.25K ops/s 8.26K ops/s 🟢▲ +1.19% 1.42%
fork_join_lockless/4 kqueue ops_per_sec 23.71K ops/s 23.76K ops/s 🔴▼ -0.03% 0.65%
fork_join_lockless/64 kqueue ops_per_sec 2.14K ops/s 2.13K ops/s 🔴▼ -0.40% 2.39%
nested/16 kqueue ops_per_sec 2.11K ops/s 2.08K ops/s 🔴▼ -0.52% 2.41%
nested/4 kqueue ops_per_sec 8.08K ops/s 8.06K ops/s 🟢▲ +0.22% 1.38%
nested_lockless/16 kqueue ops_per_sec 2.13K ops/s 2.11K ops/s 🔴▼ -0.90% 2.50%
nested_lockless/4 kqueue ops_per_sec 8.16K ops/s 8.17K ops/s 🟢▲ +0.42% 1.41%

http_server

Benchmark Backend Metric Base Head Δ Noise
concurrent/1 kqueue ops_per_sec 49.62K ops/s 49.60K ops/s 🔴▼ -0.15% 0.27%
concurrent/16 kqueue ops_per_sec 130.4K ops/s 130.2K ops/s 🔴▼ -0.37% 0.71%
concurrent/32 kqueue ops_per_sec 130.1K ops/s 129.5K ops/s 🔴▼ -0.01% 1.39%
concurrent/4 kqueue ops_per_sec 89.84K ops/s 90.22K ops/s 🟢▲ +0.53% 0.78%
multithread/1 kqueue ops_per_sec 130.0K ops/s 128.8K ops/s 🔴▼ -0.68% 1.09%
multithread/16 kqueue ops_per_sec 133.2K ops/s 133.1K ops/s 🔴▼ -0.11% 0.98%
multithread/2 kqueue ops_per_sec 183.9K ops/s 184.8K ops/s 🟢▲ +0.42% 0.84%
multithread/4 kqueue ops_per_sec 162.4K ops/s 162.2K ops/s 🔴▼ -0.27% 0.98%
multithread/8 kqueue ops_per_sec 138.8K ops/s 137.8K ops/s 🔴▼ -0.75% 1.23%
single_conn kqueue ops_per_sec 49.66K ops/s 49.61K ops/s 🔴▼ -0.19% 0.29%
single_conn_lockless kqueue ops_per_sec 50.09K ops/s 49.96K ops/s 🟢▲ +0.10% 0.67%

local_socket_latency

Benchmark Backend Metric Base Head Δ Noise
concurrent/1 kqueue latency_mean_ns 1.82 µs 1.71 µs 🟢▲ +6.17% 0.72%
concurrent/16 kqueue latency_mean_ns 27.13 µs 26.76 µs 🟢▲ +1.41% 0.36%
concurrent/4 kqueue latency_mean_ns 6.81 µs 6.65 µs 🟢▲ +2.51% 0.34%
concurrent_lockless/1 kqueue latency_mean_ns 1.81 µs 1.69 µs 🟢▲ +6.77% 0.28%
concurrent_lockless/16 kqueue latency_mean_ns 26.91 µs 26.56 µs 🟢▲ +1.15% 0.32%
concurrent_lockless/4 kqueue latency_mean_ns 6.75 µs 6.60 µs 🟢▲ +2.43% 0.32%
pingpong/1 kqueue latency_mean_ns 1.82 µs 1.71 µs 🟢▲ +6.44% 0.64%
pingpong/1024 kqueue latency_mean_ns 1.91 µs 1.80 µs 🟢▲ +5.89% 0.48%
pingpong/64 kqueue latency_mean_ns 1.82 µs 1.71 µs 🟢▲ +6.22% 0.64%
pingpong_lockless/1 kqueue latency_mean_ns 1.80 µs 1.69 µs 🟢▲ +6.62% 0.55%
pingpong_lockless/1024 kqueue latency_mean_ns 1.91 µs 1.78 µs 🟢▲ +6.52% 0.51%
pingpong_lockless/64 kqueue latency_mean_ns 1.81 µs 1.69 µs 🟢▲ +6.88% 0.56%

local_socket_throughput

Benchmark Backend Metric Base Head Δ Noise
bidirectional/1024 kqueue bytes_per_sec 1.30 GB/s 1.32 GB/s 🟢▲ +1.43% 0.38%
bidirectional/1048576 kqueue bytes_per_sec 3.89 GB/s 3.93 GB/s 🟢▲ +1.22% 0.46%
bidirectional/16384 kqueue bytes_per_sec 3.91 GB/s 3.91 GB/s 🔴▼ -0.20% 0.60%
bidirectional/262144 kqueue bytes_per_sec 3.88 GB/s 3.91 GB/s 🟢▲ +0.63% 0.87%
bidirectional/4096 kqueue bytes_per_sec 2.94 GB/s 2.93 GB/s 🟢▲ +0.22% 0.57%
bidirectional/65536 kqueue bytes_per_sec 3.89 GB/s 3.91 GB/s 🟢▲ +0.42% 0.76%
bidirectional_lockless/1024 kqueue bytes_per_sec 1.31 GB/s 1.33 GB/s 🟢▲ +1.82% 0.63%
bidirectional_lockless/1048576 kqueue bytes_per_sec 4.00 GB/s 4.02 GB/s 🟢▲ +0.44% 0.59%
bidirectional_lockless/16384 kqueue bytes_per_sec 4.01 GB/s 4.00 GB/s 🔴▼ -0.07% 0.86%
bidirectional_lockless/262144 kqueue bytes_per_sec 4.00 GB/s 4.00 GB/s 🔴▼ -0.54% 0.68%
bidirectional_lockless/4096 kqueue bytes_per_sec 3.00 GB/s 2.99 GB/s 🔴▼ -0.13% 0.62%
bidirectional_lockless/65536 kqueue bytes_per_sec 4.01 GB/s 3.99 GB/s 🔴▼ -0.64% 0.43%
unidirectional/1024 kqueue bytes_per_sec 1.28 GB/s 1.32 GB/s 🟢▲ +2.85% 0.21%
unidirectional/1048576 kqueue bytes_per_sec 3.55 GB/s 3.56 GB/s 🟢▲ +0.64% 0.40%
unidirectional/16384 kqueue bytes_per_sec 3.56 GB/s 3.55 GB/s 🔴▼ -0.32% 0.65%
unidirectional/262144 kqueue bytes_per_sec 3.54 GB/s 3.56 GB/s 🟢▲ +0.54% 0.46%
unidirectional/4096 kqueue bytes_per_sec 2.72 GB/s 2.73 GB/s 🔴▼ -0.27% 0.63%
unidirectional/65536 kqueue bytes_per_sec 3.55 GB/s 3.56 GB/s 🟢▲ +0.28% 0.43%
unidirectional_lockless/1024 kqueue bytes_per_sec 1.29 GB/s 1.32 GB/s 🟢▲ +3.22% 0.49%
unidirectional_lockless/1048576 kqueue bytes_per_sec 3.63 GB/s 3.63 GB/s 🔴▼ -0.25% 0.51%
unidirectional_lockless/16384 kqueue bytes_per_sec 3.63 GB/s 3.64 GB/s 🟢▲ +0.12% 0.37%
unidirectional_lockless/262144 kqueue bytes_per_sec 3.62 GB/s 3.63 GB/s 🟢▲ +0.25% 0.72%
unidirectional_lockless/4096 kqueue bytes_per_sec 2.78 GB/s 2.77 GB/s 🔴▼ -0.45% 0.48%
unidirectional_lockless/65536 kqueue bytes_per_sec 3.62 GB/s 3.62 GB/s 🔴▼ -0.17% 0.45%

socket_latency

Benchmark Backend Metric Base Head Δ Noise
concurrent/1 kqueue latency_mean_ns 18.73 µs 18.73 µs 🟢▲ +0.44% 0.65%
concurrent/16 kqueue latency_mean_ns 120.85 µs 120.89 µs 🔴▼ -0.32% 0.92%
concurrent/4 kqueue latency_mean_ns 43.55 µs 43.05 µs 🟢▲ +2.10% 1.58%
concurrent_lockless/1 kqueue latency_mean_ns 18.67 µs 18.68 µs 🟢▲ +0.09% 0.37%
concurrent_lockless/16 kqueue latency_mean_ns 121.13 µs 120.73 µs 🟢▲ +0.13% 1.07%
concurrent_lockless/4 kqueue latency_mean_ns 42.83 µs 42.44 µs 🟢▲ +0.88% 1.21%
pingpong/1 kqueue latency_mean_ns 15.51 µs 15.57 µs 🔴▼ -0.56% 1.29%
pingpong/1024 kqueue latency_mean_ns 19.05 µs 19.04 µs 🟢▲ +0.03% 0.19%
pingpong/64 kqueue latency_mean_ns 18.79 µs 18.70 µs 🟢▲ +0.46% 0.52%
pingpong_lockless/1 kqueue latency_mean_ns 15.35 µs 15.35 µs ⚪ 0.00% 0.06%
pingpong_lockless/1024 kqueue latency_mean_ns 19.05 µs 19.05 µs 🔴▼ -0.07% 0.39%
pingpong_lockless/64 kqueue latency_mean_ns 18.67 µs 18.64 µs 🟢▲ +0.28% 0.23%

socket_throughput

Benchmark Backend Metric Base Head Δ Noise
bidirectional/1024 kqueue bytes_per_sec 475.6 MB/s 478.4 MB/s 🟢▲ +0.91% 0.88%
bidirectional/1048576 kqueue bytes_per_sec 7.85 GB/s 7.78 GB/s 🟢▲ +1.60% 5.47%
bidirectional/16384 kqueue bytes_per_sec 4.56 GB/s 4.55 GB/s 🟢▲ +0.29% 1.36%
bidirectional/262144 kqueue bytes_per_sec 8.77 GB/s 8.75 GB/s 🔴▼ -0.14% 3.24%
bidirectional/4096 kqueue bytes_per_sec 1.82 GB/s 1.82 GB/s 🟢▲ +0.55% 0.81%
bidirectional/65536 kqueue bytes_per_sec 7.73 GB/s 7.56 GB/s 🔴▼ -2.16% 1.39%
bidirectional_lockless/1024 kqueue bytes_per_sec 475.8 MB/s 480.0 MB/s 🟢▲ +0.78% 0.94%
bidirectional_lockless/1048576 kqueue bytes_per_sec 7.76 GB/s 7.74 GB/s 🟢▲ +0.82% 6.37%
bidirectional_lockless/16384 kqueue bytes_per_sec 4.56 GB/s 4.60 GB/s 🟢▲ +1.00% 1.20%
bidirectional_lockless/262144 kqueue bytes_per_sec 8.88 GB/s 8.77 GB/s 🔴▼ -0.44% 1.47%
bidirectional_lockless/4096 kqueue bytes_per_sec 1.82 GB/s 1.84 GB/s 🟢▲ +1.28% 0.90%
bidirectional_lockless/65536 kqueue bytes_per_sec 7.61 GB/s 7.58 GB/s 🟢▲ +1.22% 1.82%
multithread/2 kqueue bytes_per_sec 7.54 GB/s 7.43 GB/s 🟢▲ +0.29% 4.98%
multithread/4 kqueue bytes_per_sec 3.61 GB/s 3.66 GB/s 🟢▲ +2.61% 3.43%
multithread/8 kqueue bytes_per_sec 3.18 GB/s 3.06 GB/s 🔴▼ -7.93% 8.54%
unidirectional/1024 kqueue bytes_per_sec 457.5 MB/s 460.1 MB/s 🟢▲ +1.30% 2.72%
unidirectional/1048576 kqueue bytes_per_sec 9.03 GB/s 8.82 GB/s 🔴▼ -0.01% 6.32%
unidirectional/16384 kqueue bytes_per_sec 4.47 GB/s 4.42 GB/s 🔴▼ -1.65% 2.03%
unidirectional/262144 kqueue bytes_per_sec 8.88 GB/s 8.84 GB/s 🔴▼ -1.01% 2.23%
unidirectional/4096 kqueue bytes_per_sec 1.80 GB/s 1.81 GB/s 🟢▲ +0.23% 2.12%
unidirectional/65536 kqueue bytes_per_sec 7.39 GB/s 7.34 GB/s 🔴▼ -1.26% 1.00%
unidirectional_lockless/1024 kqueue bytes_per_sec 478.7 MB/s 485.6 MB/s 🟢▲ +0.65% 2.04%
unidirectional_lockless/1048576 kqueue bytes_per_sec 9.04 GB/s 8.87 GB/s 🔴▼ -0.15% 5.63%
unidirectional_lockless/16384 kqueue bytes_per_sec 4.50 GB/s 4.57 GB/s 🟢▲ +0.52% 3.11%
unidirectional_lockless/262144 kqueue bytes_per_sec 8.93 GB/s 8.96 GB/s 🔴▼ -0.30% 1.93%
unidirectional_lockless/4096 kqueue bytes_per_sec 1.82 GB/s 1.83 GB/s 🔴▼ -0.58% 2.42%
unidirectional_lockless/65536 kqueue bytes_per_sec 7.42 GB/s 7.41 GB/s 🔴▼ -1.00% 1.93%

Run & raw JSON artifacts · flag rule: |median Δ| > max(2%, 3×CV of the noisier side) · advisory only

@sgerbino
sgerbino force-pushed the pr/recycling-free-list branch from 1e59c16 to f4b0d0d Compare November 12, 2026 19:16
@coderabbitai

coderabbitai Bot commented Nov 12, 2026 •

Copy link
Copy Markdown

Explain Change →

No actionable comments were generated in the recent review.

Walkthrough

The reactor scheduler now uses a score-based policy to decide when to defer the sentinel behind queued handlers. The policy score resets when the scheduler is stopped. Calls with a zero timeout do not defer.

Changes: Reactor Sentinel Deferral

Layer / File(s) Summary
Deferral policy
include/boost/corosio/native/detail/reactor/reactor_defer_policy.hpp
Adds a policy that permits deferral below a score threshold and adjusts the score after deferred and polled turns.
Scheduler integration
include/boost/corosio/native/detail/reactor/reactor_scheduler.hpp
The scheduler tracks deferred sentinel turns and applies the policy in do_one. It resets the policy when stopped and does not defer on zero-timeout calls.

Estimated code review effort: 2 (Simple) | ~10 minutes

Change: Refactor

Suggested reviewers: mvandeberg

Merge Risk: Minimal · up to 0b4df

This change reorders when queued handlers run relative to non-blocking polls in the reactor, using an adaptive policy to preserve I/O fairness. No actionable merge-blocking risk was found, and the reported tests and benchmarks cover the changed behavior.

Pre-merge checks: Passed 4

Check name Status Explanation
Title check Passed The title names the main change: queued handlers run before a non-blocking poll while the adaptive policy shows benefit. The raw summary describes the reactor_defer_policy class and the scheduler defe…
Linked Issues check Passed Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check Passed Check skipped because no linked issues were found for this pull request.
Description Check Passed Check skipped - CodeRabbit’s high-level summary is enabled.
Full details: Title check

Explanation

The title names the main change: queued handlers run before a non-blocking poll while the adaptive policy shows benefit. The raw summary describes the reactor_defer_policy class and the scheduler deferral logic in reactor_scheduler.hpp, which implement exactly this behavior. The title is specific, short, and free of file lists or vague terms, so it matches the changeset.

Finishing Touches

Generate unit tests (beta)

  • Commit to this branch
  • Create a new PR
  • Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

A rabbit watches queues go by,
The sentinel waits, then hops nearby.
When scores rise high, it takes its turn,
When handlers wait, it lets them run.
At zero time, no detour stays,
Then resets scores and bounds away.

Recent review info

Run configuration

Commits

Reviewing files that changed from the base of the PR and between cf9d859 and 0b4dfac.

Files ignored due to path filters: 3

  • test/unit/fault/kqueue_faults.cpp is excluded by !**/test/**
  • test/unit/fault/reactor_faults.hpp is excluded by !**/test/**
  • test/unit/reactor_defer_policy.cpp is excluded by !**/test/**
Files selected for processing: 1

  • include/boost/corosio/native/detail/reactor/reactor_scheduler.hpp

Included review availability: This review used your included allowance. Your plan provides up to 1 included review per hour; 0 remain after this review.

@sgerbino
sgerbino force-pushed the pr/recycling-free-list branch from f4b0d0d to 97fbd80 Compare November 13, 2026 05:57
@sgerbino sgerbino added the benchmark Triggers benchmark suite label Nov 16, 2026
@github-actions github-actions Bot removed the benchmark Triggers benchmark suite label Nov 16, 2026
@sgerbino
sgerbino force-pushed the pr/recycling-free-list branch from 97fbd80 to 54d7f4b Compare November 17, 2026 22:48
@sgerbino sgerbino changed the title perf: run handlers queued behind the reactor before polling again perf: run handlers queued behind the reactor first while that pays off Nov 17, 2026
@sgerbino
sgerbino force-pushed the pr/recycling-free-list branch from 54d7f4b to cf9d859 Compare November 17, 2026 23:02
@sgerbino
sgerbino marked this pull request as ready for review November 18, 2026 17:06
When the reactor sentinel reached the front with handlers still queued,
the scheduler polled the reactor without blocking before running them.
Those handlers usually start the next I/O themselves, such as a
coroutine an I/O completion woke, so running them first can save a
reactor pass.

The sentinel now moves behind the queued handlers once and polls on its
next turn regardless, so handlers that keep reposting still cannot
starve I/O. Waiting does not always pay: when the handlers drain the
queue, the deferred turn has nothing left to overlap with, and where
loopback data arrives later their I/O parks instead of completing.
reactor_defer_policy scores each deferral by whether its turn still
found handlers queued, and stops deferring while deferrals that empty
the queue are frequent. Turns that poll instead let it try again.

restart() forgets a pending deferral and the score, so a run stopped
with the sentinel moved back doesn't cost the next run its own. poll()
and poll_one() are unchanged.
@sgerbino
sgerbino force-pushed the pr/recycling-free-list branch from cf9d859 to 0b4dfac Compare November 18, 2026 19:20

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

Status: Backlog

Development

Successfully merging this pull request may close these issues.

2 participants