Skip to content

Improve IOCP performance - #420

Draft
TheLingeringWill wants to merge 8 commits into
cppalliance:developfrom
TheLingeringWill:pr/iocp-perf-pass
Draft

TheLingeringWill wants to merge 8 commits into
cppalliance:developfrom
TheLingeringWill:pr/iocp-perf-pass

Conversation

@TheLingeringWill

@TheLingeringWill TheLingeringWill commented Nov 14, 2026 •

Copy link
Copy Markdown
Contributor

Compared with asio's IOCP backend, corosio made several syscalls per connection that asio does not, and the Windows benchmark report showed corosio about 5% behind asio coroutines on accept_churn and 3% on socket_latency pingpong. This PR removes those calls and some smaller per-operation costs. The two allocation sites #383 names were already removed on develop by 8efad53.

Changes

  • Accept reads both endpoints from the AcceptEx buffer with GetAcceptExSockaddrs instead of calling getsockname and getpeername. If the provider doesn't hand out the function, it falls back to the old calls.
  • Connect no longer calls getsockname. The first local_endpoint() fetches and caches the endpoint, as the io_uring backend already does, and its docs now say so. A second connect on a connected socket still fails at ConnectEx with WSAEISCONN.
  • Close, cancel and open() ask the wait reactor to drop a wait only when it holds one, so they no longer wake its polling thread for nothing. A flag on each op records whether the reactor holds it.
  • Wait reactor: removing an entry moves the last entry into its slot instead of shifting the table, and draining the wakeup channel takes one recv instead of two.
  • Fix: update_timeout now reads the earliest expiry and arms the waitable timer under the dispatch lock. Two threads arming at once could leave the later deadline set, so the earlier timer fired late.
  • File wrappers move their stop token instead of copying it.

Measurements

Benchmark workflow (Windows x64 runner): all 9 accept_churn benchmarks are 3.5–4.4% faster, with 0.2–1.2% noise.

bench base head Δ
accept_churn/sequential 12.48K ops/s 12.96K ops/s +3.8%
accept_churn/concurrent/16 13.45K ops/s 13.93K ops/s +3.5%
accept_churn/burst/10 1.56K ops/s 1.62K ops/s +4.2%

Pingpong latency improves only 0.4–1.3%, so most of that gap to asio remains. Nothing else on Windows, Linux or macOS moves beyond noise.

Locally (Windows 11 ARM64, clang, Release, same run scheme), accept_churn is faster in 8 of 9 benchmarks (+1.9% to +7.7%). With 2–7% noise, nothing is flagged.

Testing

New iocp_faults tests:

  • Accept: endpoints come from the AcceptEx buffer with no getsockname or getpeername call, and come from those calls when GetAcceptExSockaddrs is missing.
  • Connect: no getsockname until the first local_endpoint(), and none after it. A second connect still fails with WSAEISCONN. The deferred endpoint is still right when first read after a reset or a close.
  • Wait reactor: closes and cancels with nothing parked send nothing to the polling thread. A wait moved by a removal stays parked and cancellable.

testWakeSendFails and testWaitReactorPollFails relied on a cancel with nothing parked to wake the polling thread. They now re-park or register a wait instead.

All CI checks pass. Locally on Windows 11 ARM64, all 122 ctest tests pass.

Part of #383

@coderabbitai

coderabbitai Bot commented Nov 14, 2026

Copy link
Copy Markdown

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

  • Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true
  • Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

@cppalliance-bot

cppalliance-bot commented Nov 14, 2026 •

Copy link
Copy Markdown

An automated preview of the documentation is available at https://420.corosio.prtest3.cppalliance.org/index.html

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

2026-11-16 21:36:49 UTC

@cppalliance-bot

cppalliance-bot commented Nov 14, 2026 •

Copy link
Copy Markdown

GCOVR code coverage report https://420.corosio.prtest3.cppalliance.org/gcovr/index.html
LCOV code coverage report https://420.corosio.prtest3.cppalliance.org/genhtml/index.html
Coverage Diff Report https://420.corosio.prtest3.cppalliance.org/diff-report/index.html

Build time: 2026-11-16 21:54:42 UTC

Each accept asked the kernel for both endpoints of the new connection,
one getsockname and one getpeername, although AcceptEx had already
written both addresses into the buffer it was given. The tcp service now
looks up GetAcceptExSockaddrs next to AcceptEx and the accept parses the
two addresses out of that buffer, which takes no syscall.

A provider that does not hand out GetAcceptExSockaddrs leaves the
accept asking the kernel as before.
A successful connect called getsockname to cache a local endpoint that
most connections never read. It now marks the endpoint pending, and the
first local_endpoint() call fetches and caches it, as the io_uring
backend does since it stopped fetching it eagerly. One reader claims
the fill and a reader racing it fetches its own copy, so concurrent
calls never read the cache while it is being written. A fetch that
fails caches nothing, and the next call asks again.

ConnectEx needs a bound socket, and an endpoint left unfetched says the
socket most likely is one. The mark can outlive the socket it described
when a late completion races a close and reopen, so a connect over it
tries the bind anyway and lets an "already bound" refusal through to
ConnectEx. A second connect on a connected socket still fails as it did
before, at ConnectEx.

The local_endpoint() documentation now says the endpoint may be cached
by the first call after a connect rather than at the connect.
Every socket close and cancel asked the wait reactor to drop the
socket's wait op, parked or not, and every open() closes first. Once the
polling thread exists, each ask took the reactor lock, sent a byte down
the wakeup channel, and made the thread wake, search its table and
rebuild its poll set, for a socket that in nearly every case had nothing
parked.

Ops now carry a flag the reactor sets when a register queues them and
clears before it completes them. A cancel reads it under the reactor
lock and returns when it is clear: the op is not held, so there is
nothing to drop. A cancel that comes before the register is still caught
there by the op's cancelled flag, and one that races the reactor letting
go of the op is dropped by the table search as before.

Two fault tests used a cancel with nothing parked as a wake that always
reaches the polling thread. testWakeSendFails now parks its wait again
after every cancel, and testWaitReactorPollFails wakes the thread with a
register.
Removing a fired, cancelled or refused wait erased it from the middle of
the table, moving every entry after it down, so k completions out of n
parked waits cost O(k*n) moves. The table has no order, so the last
entry now fills the hole; the walks that remove as they go run from the
back, so the entry moved was already examined and the slots below it
keep their poll set entries.

Draining the wakeup channel read until recv failed, which took a second
recv on every wake just to hear WSAEWOULDBLOCK. Only the sender that
turns wake_pending_ from false to true writes a byte, so the channel
holds one at a time and a short read has emptied it.
update_timeout read the earliest expiry and set the waitable timer with
no lock held. Two threads arming at once could read different expiries
and set the timer in the opposite order, leaving the later deadline set,
so the earlier timer fired late, at the later one's deadline. A thread
inserting a new earliest timer races the run thread re-arming after a
fire the same way.

The read and the arm now happen under the dispatch lock, as asio arms
its timer. A new earliest expiry is published before its notification
takes the lock, so the last thread through always arms the earliest
one.
The stream and random-access file wrappers passed their by-value stop
token on by copy, so every read and write took and dropped a reference
on the stop state. They now move it, as the stream handle already does.
@pgit

pgit commented Nov 14, 2026

Copy link
Copy Markdown

There is something weird with the unit tests in the CI:

https://github.com/cppalliance/corosio/actions/runs/43557619651/job/132334044957?pr=420#step:31:1168

          Start  87: boost.corosio.wolfssl_engine
   84/245 Test  #87: boost.corosio.wolfssl_engine ...............................   Passed    0.21 sec
          Start  88: boost.corosio.wolfssl_stream
   85/245 Test  #78: boost.corosio.udp_socket.iocp ..............................   Passed    1.49 sec
          Start  89: fault
   86/245 Test  #89: fault ......................................................   Passed    0.01 sec
          Start  90: harness:
   87/245 Test  #90: harness: ...................................................   Passed    0.01 sec
          Start  91: pthread_create
   88/245 Test  #91: pthread_create .............................................   Passed    0.01 sec

@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
@github-actions

Copy link
Copy Markdown

Benchmark report

13bcf933adf8 (base) vs c3e11482f0f4 (head)

  • ✅ linux — clean (epoll, uring)
  • ⚠️ windows — 9 flagged (iocp)
  • ✅ macos — clean (kqueue)

⚠️ Flagged

Platform Backend Benchmark Metric Base Head Δ Noise
windows iocp accept_churn/burst/10 ops_per_sec 1.56K ops/s 1.62K ops/s 🟢▲ +4.21% 0.34%
windows iocp accept_churn/burst/100 ops_per_sec 156.3 ops/s 162.2 ops/s 🟢▲ +3.62% 1.15%
windows iocp accept_churn/burst_lockless/10 ops_per_sec 1.56K ops/s 1.63K ops/s 🟢▲ +4.38% 0.36%
windows iocp accept_churn/burst_lockless/100 ops_per_sec 159.8 ops/s 166.0 ops/s 🟢▲ +4.09% 0.84%
windows iocp accept_churn/concurrent/1 ops_per_sec 12.53K ops/s 13.04K ops/s 🟢▲ +4.14% 0.38%
windows iocp accept_churn/concurrent/16 ops_per_sec 13.45K ops/s 13.93K ops/s 🟢▲ +3.46% 0.30%
windows iocp accept_churn/concurrent/4 ops_per_sec 13.17K ops/s 13.68K ops/s 🟢▲ +3.93% 0.41%
windows iocp accept_churn/sequential ops_per_sec 12.48K ops/s 12.96K ops/s 🟢▲ +3.83% 0.33%
windows iocp accept_churn/sequential_lockless ops_per_sec 12.56K ops/s 13.04K ops/s 🟢▲ +3.84% 0.22%
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 2.01K ops/s 2.01K ops/s 🔴▼ -0.11% 0.99%
burst/10 uring ops_per_sec 2.41K ops/s 2.41K ops/s 🔴▼ -0.06% 1.08%
burst/100 epoll ops_per_sec 204.9 ops/s 203.9 ops/s 🔴▼ -0.19% 1.19%
burst/100 uring ops_per_sec 248.3 ops/s 247.6 ops/s 🔴▼ -0.22% 0.65%
burst_lockless/10 epoll ops_per_sec 2.03K ops/s 2.03K ops/s 🟢▲ +0.02% 0.20%
burst_lockless/10 uring ops_per_sec 2.43K ops/s 2.41K ops/s 🔴▼ -0.28% 1.05%
burst_lockless/100 epoll ops_per_sec 206.5 ops/s 206.4 ops/s 🔴▼ -0.09% 0.28%
burst_lockless/100 uring ops_per_sec 249.3 ops/s 248.7 ops/s 🔴▼ -0.29% 0.67%
concurrent/1 epoll ops_per_sec 15.90K ops/s 15.86K ops/s 🟢▲ +0.21% 1.38%
concurrent/1 uring ops_per_sec 17.47K ops/s 17.24K ops/s 🔴▼ -1.18% 1.21%
concurrent/16 epoll ops_per_sec 16.01K ops/s 15.94K ops/s 🔴▼ -0.14% 0.98%
concurrent/16 uring ops_per_sec 18.87K ops/s 18.94K ops/s 🔴▼ -0.08% 1.39%
concurrent/4 epoll ops_per_sec 15.95K ops/s 15.90K ops/s 🔴▼ -0.22% 1.10%
concurrent/4 uring ops_per_sec 17.93K ops/s 17.89K ops/s 🟢▲ +0.02% 1.03%
sequential epoll ops_per_sec 15.85K ops/s 15.83K ops/s 🔴▼ -0.02% 0.77%
sequential uring ops_per_sec 17.33K ops/s 17.22K ops/s 🔴▼ -1.29% 1.12%
sequential_lockless epoll ops_per_sec 16.10K ops/s 16.03K ops/s 🔴▼ -0.12% 1.30%
sequential_lockless uring ops_per_sec 17.54K ops/s 17.40K ops/s 🔴▼ -1.33% 1.19%

fan_out

Benchmark Backend Metric Base Head Δ Noise
concurrent_parents/1 epoll ops_per_sec 3.86K ops/s 3.87K ops/s 🟢▲ +0.15% 1.07%
concurrent_parents/1 uring ops_per_sec 4.04K ops/s 4.00K ops/s 🔴▼ -0.56% 1.21%
concurrent_parents/16 epoll ops_per_sec 3.80K ops/s 3.79K ops/s 🔴▼ -0.12% 0.98%
concurrent_parents/16 uring ops_per_sec 3.94K ops/s 3.93K ops/s 🔴▼ -0.04% 1.02%
concurrent_parents/4 epoll ops_per_sec 3.79K ops/s 3.79K ops/s 🔴▼ -0.01% 1.07%
concurrent_parents/4 uring ops_per_sec 4.00K ops/s 3.96K ops/s 🔴▼ -0.47% 1.03%
concurrent_parents_lockless/1 epoll ops_per_sec 3.92K ops/s 3.89K ops/s 🔴▼ -0.21% 1.20%
concurrent_parents_lockless/1 uring ops_per_sec 4.06K ops/s 4.04K ops/s 🔴▼ -0.19% 0.96%
concurrent_parents_lockless/16 epoll ops_per_sec 3.83K ops/s 3.80K ops/s 🔴▼ -0.06% 1.10%
concurrent_parents_lockless/16 uring ops_per_sec 3.98K ops/s 3.96K ops/s 🔴▼ -0.13% 0.96%
concurrent_parents_lockless/4 epoll ops_per_sec 3.84K ops/s 3.82K ops/s 🔴▼ -0.07% 1.21%
concurrent_parents_lockless/4 uring ops_per_sec 4.02K ops/s 4.00K ops/s 🟢▲ +0.03% 0.96%
fork_join/1 epoll ops_per_sec 56.35K ops/s 56.33K ops/s 🟢▲ +0.03% 0.34%
fork_join/1 uring ops_per_sec 54.60K ops/s 54.29K ops/s 🔴▼ -0.78% 1.25%
fork_join/16 epoll ops_per_sec 3.85K ops/s 3.85K ops/s 🔴▼ -0.07% 1.01%
fork_join/16 uring ops_per_sec 4.02K ops/s 4.01K ops/s 🔴▼ -0.13% 1.19%
fork_join/4 epoll ops_per_sec 15.22K ops/s 15.17K ops/s 🟢▲ +0.08% 1.22%
fork_join/4 uring ops_per_sec 15.06K ops/s 15.01K ops/s 🔴▼ -0.18% 1.02%
fork_join/64 epoll ops_per_sec 947.4 ops/s 950.8 ops/s 🟢▲ +0.09% 0.93%
fork_join/64 uring ops_per_sec 996.1 ops/s 996.2 ops/s 🟢▲ +0.11% 1.00%
fork_join_lockless/1 epoll ops_per_sec 57.31K ops/s 57.47K ops/s 🟢▲ +0.19% 1.30%
fork_join_lockless/1 uring ops_per_sec 56.15K ops/s 56.02K ops/s 🔴▼ -0.25% 1.01%
fork_join_lockless/16 epoll ops_per_sec 3.89K ops/s 3.90K ops/s 🟢▲ +0.08% 1.21%
fork_join_lockless/16 uring ops_per_sec 4.05K ops/s 4.07K ops/s 🔴▼ -0.04% 0.98%
fork_join_lockless/4 epoll ops_per_sec 15.39K ops/s 15.40K ops/s 🔴▼ -0.23% 1.26%
fork_join_lockless/4 uring ops_per_sec 15.26K ops/s 15.28K ops/s 🟢▲ +0.03% 1.07%
fork_join_lockless/64 epoll ops_per_sec 953.4 ops/s 958.0 ops/s 🔴▼ -0.03% 1.10%
fork_join_lockless/64 uring ops_per_sec 1.00K ops/s 1.01K ops/s 🟢▲ +0.04% 1.05%
nested/16 epoll ops_per_sec 940.8 ops/s 945.2 ops/s 🟢▲ +0.16% 1.04%
nested/16 uring ops_per_sec 992.7 ops/s 987.9 ops/s 🔴▼ -0.11% 0.99%
nested/4 epoll ops_per_sec 3.81K ops/s 3.84K ops/s 🟢▲ +0.23% 1.14%
nested/4 uring ops_per_sec 4.00K ops/s 3.98K ops/s 🟢▲ +0.03% 1.07%
nested_lockless/16 epoll ops_per_sec 950.9 ops/s 955.2 ops/s 🟢▲ +0.15% 1.09%
nested_lockless/16 uring ops_per_sec 1.00K ops/s 997.2 ops/s 🔴▼ -0.29% 0.77%
nested_lockless/4 epoll ops_per_sec 3.86K ops/s 3.89K ops/s 🟢▲ +0.20% 1.10%
nested_lockless/4 uring ops_per_sec 4.04K ops/s 4.03K ops/s 🔴▼ -0.09% 0.99%

http_server

Benchmark Backend Metric Base Head Δ Noise
concurrent/1 epoll ops_per_sec 57.20K ops/s 56.96K ops/s 🔴▼ -0.35% 1.28%
concurrent/1 uring ops_per_sec 53.42K ops/s 53.19K ops/s 🔴▼ -0.47% 1.10%
concurrent/16 epoll ops_per_sec 60.30K ops/s 59.88K ops/s 🔴▼ -0.02% 1.32%
concurrent/16 uring ops_per_sec 63.69K ops/s 63.80K ops/s 🔴▼ -0.34% 1.21%
concurrent/32 epoll ops_per_sec 59.60K ops/s 59.48K ops/s 🟢▲ +0.09% 0.93%
concurrent/32 uring ops_per_sec 64.03K ops/s 63.89K ops/s 🔴▼ -0.54% 1.11%
concurrent/4 epoll ops_per_sec 59.47K ops/s 59.22K ops/s 🔴▼ -0.71% 1.10%
concurrent/4 uring ops_per_sec 61.24K ops/s 61.40K ops/s 🔴▼ -0.45% 1.19%
multithread/1 epoll ops_per_sec 59.48K ops/s 59.35K ops/s 🟢▲ +0.21% 1.12%
multithread/1 uring ops_per_sec 63.84K ops/s 63.49K ops/s 🔴▼ -0.88% 1.20%
multithread/16 epoll ops_per_sec 163.4K ops/s 163.6K ops/s 🟢▲ +0.15% 0.24%
multithread/16 uring ops_per_sec 100.7K ops/s 100.8K ops/s 🟢▲ +0.32% 0.57%
multithread/2 epoll ops_per_sec 87.53K ops/s 87.58K ops/s 🟢▲ +0.06% 0.45%
multithread/2 uring ops_per_sec 75.30K ops/s 75.22K ops/s 🔴▼ -0.22% 0.52%
multithread/4 epoll ops_per_sec 161.1K ops/s 161.3K ops/s 🔴▼ -0.12% 0.70%
multithread/4 uring ops_per_sec 120.3K ops/s 119.8K ops/s 🔴▼ -0.20% 0.54%
multithread/8 epoll ops_per_sec 162.3K ops/s 163.1K ops/s 🟢▲ +0.28% 0.57%
multithread/8 uring ops_per_sec 112.0K ops/s 112.0K ops/s 🔴▼ -0.08% 0.35%
single_conn epoll ops_per_sec 57.01K ops/s 56.93K ops/s 🔴▼ -0.31% 1.15%
single_conn uring ops_per_sec 53.23K ops/s 53.23K ops/s 🔴▼ -0.07% 0.99%
single_conn_lockless epoll ops_per_sec 57.62K ops/s 57.62K ops/s 🟢▲ +0.01% 1.04%
single_conn_lockless uring ops_per_sec 55.08K ops/s 54.99K ops/s 🔴▼ -0.52% 1.25%

local_socket_latency

Benchmark Backend Metric Base Head Δ Noise
concurrent/1 epoll latency_mean_ns 7.24 µs 7.35 µs 🔴▼ -0.30% 4.52%
concurrent/1 uring latency_mean_ns 5.66 µs 5.94 µs 🔴▼ -1.62% 6.48%
concurrent/16 epoll latency_mean_ns 110.23 µs 110.37 µs 🔴▼ -0.16% 0.28%
concurrent/16 uring latency_mean_ns 92.69 µs 92.73 µs 🔴▼ -0.09% 0.86%
concurrent/4 epoll latency_mean_ns 27.72 µs 28.06 µs 🔴▼ -0.36% 1.54%
concurrent/4 uring latency_mean_ns 23.13 µs 23.19 µs 🟢▲ +0.08% 1.40%
concurrent_lockless/1 epoll latency_mean_ns 7.02 µs 7.23 µs 🔴▼ -0.72% 4.16%
concurrent_lockless/1 uring latency_mean_ns 5.62 µs 5.82 µs 🔴▼ -0.43% 4.44%
concurrent_lockless/16 epoll latency_mean_ns 108.89 µs 108.82 µs 🔴▼ -0.03% 0.31%
concurrent_lockless/16 uring latency_mean_ns 92.12 µs 92.13 µs 🔴▼ -0.12% 0.92%
concurrent_lockless/4 epoll latency_mean_ns 27.37 µs 27.66 µs 🟢▲ +0.07% 1.36%
concurrent_lockless/4 uring latency_mean_ns 23.03 µs 23.16 µs 🔴▼ -0.16% 1.15%
pingpong/1 epoll latency_mean_ns 7.18 µs 7.09 µs 🔴▼ -0.07% 4.36%
pingpong/1 uring latency_mean_ns 5.73 µs 5.59 µs 🟢▲ +0.31% 4.30%
pingpong/1024 epoll latency_mean_ns 7.77 µs 7.88 µs 🔴▼ -0.02% 3.80%
pingpong/1024 uring latency_mean_ns 6.37 µs 6.40 µs 🔴▼ -1.49% 5.38%
pingpong/64 epoll latency_mean_ns 7.22 µs 7.26 µs 🔴▼ -0.22% 4.34%
pingpong/64 uring latency_mean_ns 5.64 µs 5.77 µs 🔴▼ -0.25% 4.18%
pingpong_lockless/1 epoll latency_mean_ns 7.09 µs 7.10 µs 🔴▼ -0.20% 4.45%
pingpong_lockless/1 uring latency_mean_ns 5.58 µs 5.72 µs 🔴▼ -0.30% 5.24%
pingpong_lockless/1024 epoll latency_mean_ns 7.38 µs 7.67 µs 🔴▼ -8.61% 4.42%
pingpong_lockless/1024 uring latency_mean_ns 6.04 µs 6.26 µs 🔴▼ -8.37% 5.84%
pingpong_lockless/64 epoll latency_mean_ns 7.13 µs 7.01 µs 🔴▼ -0.21% 4.47%
pingpong_lockless/64 uring latency_mean_ns 5.77 µs 5.77 µs 🟢▲ +0.04% 5.33%

local_socket_throughput

Benchmark Backend Metric Base Head Δ Noise
bidirectional/1024 epoll bytes_per_sec 308.4 MB/s 305.2 MB/s 🔴▼ -0.05% 1.56%
bidirectional/1024 uring bytes_per_sec 354.4 MB/s 355.6 MB/s 🟢▲ +0.41% 0.91%
bidirectional/1048576 epoll bytes_per_sec 3.20 GB/s 3.25 GB/s 🟢▲ +1.00% 3.28%
bidirectional/1048576 uring bytes_per_sec 3.56 GB/s 3.54 GB/s 🟢▲ +0.23% 2.92%
bidirectional/16384 epoll bytes_per_sec 3.13 GB/s 2.99 GB/s 🔴▼ -4.77% 4.06%
bidirectional/16384 uring bytes_per_sec 3.00 GB/s 2.94 GB/s 🔴▼ -2.32% 2.37%
bidirectional/262144 epoll bytes_per_sec 3.26 GB/s 3.25 GB/s 🔴▼ -0.72% 2.30%
bidirectional/262144 uring bytes_per_sec 3.54 GB/s 3.52 GB/s 🔴▼ -2.16% 2.36%
bidirectional/4096 epoll bytes_per_sec 1.01 GB/s 1.01 GB/s 🟢▲ +0.24% 0.96%
bidirectional/4096 uring bytes_per_sec 1.13 GB/s 1.13 GB/s 🔴▼ -0.30% 0.81%
bidirectional/65536 epoll bytes_per_sec 3.01 GB/s 3.02 GB/s 🟢▲ +0.84% 2.90%
bidirectional/65536 uring bytes_per_sec 3.49 GB/s 3.45 GB/s 🔴▼ -0.79% 1.98%
bidirectional_lockless/1024 epoll bytes_per_sec 309.9 MB/s 309.0 MB/s 🔴▼ -0.76% 1.16%
bidirectional_lockless/1024 uring bytes_per_sec 354.9 MB/s 353.9 MB/s 🔴▼ -0.02% 0.80%
bidirectional_lockless/1048576 epoll bytes_per_sec 3.24 GB/s 3.25 GB/s 🟢▲ +0.84% 2.59%
bidirectional_lockless/1048576 uring bytes_per_sec 3.53 GB/s 3.57 GB/s 🟢▲ +0.66% 4.00%
bidirectional_lockless/16384 epoll bytes_per_sec 3.10 GB/s 3.07 GB/s 🔴▼ -0.99% 2.35%
bidirectional_lockless/16384 uring bytes_per_sec 3.01 GB/s 2.99 GB/s 🔴▼ -0.96% 1.66%
bidirectional_lockless/262144 epoll bytes_per_sec 3.28 GB/s 3.22 GB/s 🔴▼ -2.70% 2.52%
bidirectional_lockless/262144 uring bytes_per_sec 3.55 GB/s 3.53 GB/s 🔴▼ -2.15% 2.36%
bidirectional_lockless/4096 epoll bytes_per_sec 1.01 GB/s 1.01 GB/s 🔴▼ -0.35% 0.61%
bidirectional_lockless/4096 uring bytes_per_sec 1.14 GB/s 1.13 GB/s 🔴▼ -0.41% 0.67%
bidirectional_lockless/65536 epoll bytes_per_sec 3.01 GB/s 3.01 GB/s 🔴▼ -0.30% 2.44%
bidirectional_lockless/65536 uring bytes_per_sec 3.51 GB/s 3.50 GB/s 🟢▲ +0.76% 2.54%
unidirectional/1024 epoll bytes_per_sec 298.1 MB/s 297.9 MB/s 🔴▼ -0.34% 1.17%
unidirectional/1024 uring bytes_per_sec 355.2 MB/s 353.8 MB/s 🔴▼ -0.49% 0.47%
unidirectional/1048576 epoll bytes_per_sec 3.31 GB/s 3.32 GB/s 🟢▲ +1.29% 2.44%
unidirectional/1048576 uring bytes_per_sec 3.53 GB/s 3.56 GB/s 🟢▲ +0.94% 4.12%
unidirectional/16384 epoll bytes_per_sec 3.10 GB/s 2.97 GB/s 🔴▼ -3.70% 3.32%
unidirectional/16384 uring bytes_per_sec 2.98 GB/s 2.99 GB/s 🔴▼ -0.14% 1.80%
unidirectional/262144 epoll bytes_per_sec 3.35 GB/s 3.30 GB/s 🔴▼ -1.92% 1.92%
unidirectional/262144 uring bytes_per_sec 3.55 GB/s 3.51 GB/s 🔴▼ -1.87% 2.04%
unidirectional/4096 epoll bytes_per_sec 985.0 MB/s 981.2 MB/s 🔴▼ -0.03% 0.71%
unidirectional/4096 uring bytes_per_sec 1.13 GB/s 1.13 GB/s 🔴▼ -0.02% 0.60%
unidirectional/65536 epoll bytes_per_sec 3.24 GB/s 3.27 GB/s 🟢▲ +0.63% 4.23%
unidirectional/65536 uring bytes_per_sec 3.49 GB/s 3.51 GB/s 🟢▲ +2.20% 3.34%
unidirectional_lockless/1024 epoll bytes_per_sec 300.8 MB/s 302.9 MB/s 🟢▲ +1.27% 1.40%
unidirectional_lockless/1024 uring bytes_per_sec 356.6 MB/s 354.2 MB/s 🔴▼ -0.46% 0.69%
unidirectional_lockless/1048576 epoll bytes_per_sec 3.36 GB/s 3.36 GB/s 🟢▲ +0.67% 2.57%
unidirectional_lockless/1048576 uring bytes_per_sec 3.52 GB/s 3.60 GB/s 🟢▲ +2.07% 4.23%
unidirectional_lockless/16384 epoll bytes_per_sec 3.10 GB/s 3.04 GB/s 🔴▼ -1.90% 2.53%
unidirectional_lockless/16384 uring bytes_per_sec 3.03 GB/s 3.03 GB/s 🟢▲ +0.05% 1.46%
unidirectional_lockless/262144 epoll bytes_per_sec 3.38 GB/s 3.33 GB/s 🔴▼ -1.55% 2.44%
unidirectional_lockless/262144 uring bytes_per_sec 3.56 GB/s 3.57 GB/s 🔴▼ -1.07% 2.46%
unidirectional_lockless/4096 epoll bytes_per_sec 998.8 MB/s 995.7 MB/s 🔴▼ -0.17% 1.15%
unidirectional_lockless/4096 uring bytes_per_sec 1.14 GB/s 1.14 GB/s 🔴▼ -0.02% 0.98%
unidirectional_lockless/65536 epoll bytes_per_sec 3.26 GB/s 3.27 GB/s 🔴▼ -0.16% 2.92%
unidirectional_lockless/65536 uring bytes_per_sec 3.56 GB/s 3.59 GB/s 🟢▲ +0.92% 2.11%

socket_latency

Benchmark Backend Metric Base Head Δ Noise
concurrent/1 epoll latency_mean_ns 14.81 µs 14.76 µs ⚪ 0.00% 0.67%
concurrent/1 uring latency_mean_ns 13.62 µs 13.62 µs 🔴▼ -0.29% 1.21%
concurrent/16 epoll latency_mean_ns 233.51 µs 233.68 µs 🔴▼ -0.08% 0.23%
concurrent/16 uring latency_mean_ns 220.05 µs 220.03 µs 🔴▼ -0.14% 1.25%
concurrent/4 epoll latency_mean_ns 58.30 µs 58.44 µs 🔴▼ -0.27% 0.27%
concurrent/4 uring latency_mean_ns 54.65 µs 54.65 µs 🔴▼ -0.15% 1.17%
concurrent_lockless/1 epoll latency_mean_ns 14.71 µs 14.73 µs 🔴▼ -0.39% 0.60%
concurrent_lockless/1 uring latency_mean_ns 13.61 µs 13.60 µs 🔴▼ -0.20% 1.22%
concurrent_lockless/16 epoll latency_mean_ns 231.04 µs 231.31 µs 🔴▼ -0.30% 1.18%
concurrent_lockless/16 uring latency_mean_ns 220.93 µs 220.00 µs 🔴▼ -0.09% 1.22%
concurrent_lockless/4 epoll latency_mean_ns 57.84 µs 57.73 µs 🔴▼ -0.13% 1.18%
concurrent_lockless/4 uring latency_mean_ns 54.82 µs 54.53 µs 🟢▲ +0.13% 1.10%
pingpong/1 epoll latency_mean_ns 14.67 µs 14.63 µs ⚪ 0.00% 1.46%
pingpong/1 uring latency_mean_ns 13.51 µs 13.51 µs 🔴▼ -0.16% 1.18%
pingpong/1024 epoll latency_mean_ns 15.06 µs 14.97 µs 🟢▲ +0.13% 1.31%
pingpong/1024 uring latency_mean_ns 13.89 µs 13.85 µs 🔴▼ -0.30% 1.28%
pingpong/64 epoll latency_mean_ns 14.78 µs 14.71 µs 🟢▲ +0.05% 1.39%
pingpong/64 uring latency_mean_ns 13.62 µs 13.56 µs 🟢▲ +0.01% 1.27%
pingpong_lockless/1 epoll latency_mean_ns 14.58 µs 14.56 µs 🟢▲ +0.07% 1.08%
pingpong_lockless/1 uring latency_mean_ns 13.49 µs 13.52 µs 🔴▼ -0.30% 1.11%
pingpong_lockless/1024 epoll latency_mean_ns 14.91 µs 14.92 µs 🔴▼ -0.45% 1.19%
pingpong_lockless/1024 uring latency_mean_ns 13.82 µs 13.86 µs 🔴▼ -0.33% 1.42%
pingpong_lockless/64 epoll latency_mean_ns 14.64 µs 14.63 µs 🟢▲ +0.10% 1.12%
pingpong_lockless/64 uring latency_mean_ns 13.52 µs 13.59 µs 🔴▼ -0.24% 1.31%

socket_throughput

Benchmark Backend Metric Base Head Δ Noise
bidirectional/1024 epoll bytes_per_sec 134.0 MB/s 134.5 MB/s 🔴▼ -0.16% 1.39%
bidirectional/1024 uring bytes_per_sec 149.8 MB/s 149.9 MB/s 🔴▼ -0.09% 0.36%
bidirectional/1048576 epoll bytes_per_sec 3.32 GB/s 3.26 GB/s 🔴▼ -1.93% 2.14%
bidirectional/1048576 uring bytes_per_sec 3.35 GB/s 3.33 GB/s 🔴▼ -0.28% 0.81%
bidirectional/16384 epoll bytes_per_sec 1.62 GB/s 1.63 GB/s 🟢▲ +1.50% 1.76%
bidirectional/16384 uring bytes_per_sec 1.60 GB/s 1.62 GB/s 🟢▲ +0.49% 2.11%
bidirectional/262144 epoll bytes_per_sec 3.24 GB/s 3.23 GB/s 🔴▼ -0.36% 1.84%
bidirectional/262144 uring bytes_per_sec 3.29 GB/s 3.28 GB/s 🟢▲ +0.08% 0.73%
bidirectional/4096 epoll bytes_per_sec 504.8 MB/s 508.6 MB/s 🟢▲ +0.59% 1.20%
bidirectional/4096 uring bytes_per_sec 520.8 MB/s 520.9 MB/s 🟢▲ +0.09% 0.35%
bidirectional/65536 epoll bytes_per_sec 2.64 GB/s 2.57 GB/s 🔴▼ -2.83% 4.04%
bidirectional/65536 uring bytes_per_sec 2.86 GB/s 2.83 GB/s 🔴▼ -1.62% 1.97%
bidirectional_lockless/1024 epoll bytes_per_sec 134.2 MB/s 135.0 MB/s 🟢▲ +0.01% 1.07%
bidirectional_lockless/1024 uring bytes_per_sec 150.9 MB/s 149.9 MB/s 🔴▼ -0.21% 1.02%
bidirectional_lockless/1048576 epoll bytes_per_sec 3.29 GB/s 3.29 GB/s 🟢▲ +0.71% 2.19%
bidirectional_lockless/1048576 uring bytes_per_sec 3.31 GB/s 3.33 GB/s 🟢▲ +0.74% 0.69%
bidirectional_lockless/16384 epoll bytes_per_sec 1.63 GB/s 1.63 GB/s 🔴▼ -0.54% 2.13%
bidirectional_lockless/16384 uring bytes_per_sec 1.61 GB/s 1.62 GB/s 🟢▲ +0.70% 1.29%
bidirectional_lockless/262144 epoll bytes_per_sec 3.23 GB/s 3.27 GB/s 🟢▲ +0.64% 1.18%
bidirectional_lockless/262144 uring bytes_per_sec 3.21 GB/s 3.27 GB/s 🟢▲ +1.85% 0.90%
bidirectional_lockless/4096 epoll bytes_per_sec 506.1 MB/s 507.9 MB/s ⚪ 0.00% 1.13%
bidirectional_lockless/4096 uring bytes_per_sec 523.6 MB/s 522.3 MB/s 🔴▼ -0.43% 1.15%
bidirectional_lockless/65536 epoll bytes_per_sec 2.57 GB/s 2.63 GB/s 🟢▲ +5.38% 2.91%
bidirectional_lockless/65536 uring bytes_per_sec 2.88 GB/s 2.90 GB/s 🟢▲ +0.31% 2.35%
multithread/2 epoll bytes_per_sec 3.68 GB/s 3.67 GB/s 🔴▼ -0.47% 1.05%
multithread/2 uring bytes_per_sec 4.22 GB/s 4.22 GB/s 🔴▼ -0.12% 0.43%
multithread/4 epoll bytes_per_sec 5.72 GB/s 5.76 GB/s 🟢▲ +0.76% 1.09%
multithread/4 uring bytes_per_sec 6.75 GB/s 6.73 GB/s 🔴▼ -0.80% 0.75%
multithread/8 epoll bytes_per_sec 6.37 GB/s 6.31 GB/s 🔴▼ -1.37% 1.18%
multithread/8 uring bytes_per_sec 6.86 GB/s 6.87 GB/s 🟢▲ +0.17% 0.40%
unidirectional/1024 epoll bytes_per_sec 126.0 MB/s 126.7 MB/s 🟢▲ +0.20% 1.45%
unidirectional/1024 uring bytes_per_sec 148.5 MB/s 148.3 MB/s 🔴▼ -0.02% 0.25%
unidirectional/1048576 epoll bytes_per_sec 3.34 GB/s 3.35 GB/s 🟢▲ +1.54% 1.64%
unidirectional/1048576 uring bytes_per_sec 3.40 GB/s 3.39 GB/s 🔴▼ -0.04% 0.79%
unidirectional/16384 epoll bytes_per_sec 1.54 GB/s 1.55 GB/s 🟢▲ +2.22% 1.79%
unidirectional/16384 uring bytes_per_sec 1.71 GB/s 1.70 GB/s 🔴▼ -0.24% 1.27%
unidirectional/262144 epoll bytes_per_sec 3.34 GB/s 3.37 GB/s 🟢▲ +0.92% 1.05%
unidirectional/262144 uring bytes_per_sec 3.27 GB/s 3.24 GB/s 🔴▼ -1.23% 0.85%
unidirectional/4096 epoll bytes_per_sec 472.8 MB/s 472.9 MB/s 🟢▲ +0.04% 1.20%
unidirectional/4096 uring bytes_per_sec 518.3 MB/s 519.3 MB/s 🟢▲ +0.20% 0.59%
unidirectional/65536 epoll bytes_per_sec 2.72 GB/s 2.78 GB/s 🔴▼ -0.14% 4.61%
unidirectional/65536 uring bytes_per_sec 2.84 GB/s 2.83 GB/s 🔴▼ -1.13% 1.91%
unidirectional_lockless/1024 epoll bytes_per_sec 126.0 MB/s 126.4 MB/s 🟢▲ +0.26% 0.22%
unidirectional_lockless/1024 uring bytes_per_sec 148.7 MB/s 148.5 MB/s 🔴▼ -0.21% 0.16%
unidirectional_lockless/1048576 epoll bytes_per_sec 3.33 GB/s 3.33 GB/s 🔴▼ -0.13% 1.41%
unidirectional_lockless/1048576 uring bytes_per_sec 3.37 GB/s 3.37 GB/s 🟢▲ +0.16% 1.11%
unidirectional_lockless/16384 epoll bytes_per_sec 1.56 GB/s 1.55 GB/s 🔴▼ -0.90% 1.26%
unidirectional_lockless/16384 uring bytes_per_sec 1.71 GB/s 1.70 GB/s 🔴▼ -0.49% 0.99%
unidirectional_lockless/262144 epoll bytes_per_sec 3.33 GB/s 3.33 GB/s 🔴▼ -0.72% 1.67%
unidirectional_lockless/262144 uring bytes_per_sec 3.28 GB/s 3.28 GB/s 🔴▼ -0.14% 1.21%
unidirectional_lockless/4096 epoll bytes_per_sec 474.1 MB/s 474.1 MB/s 🟢▲ +0.07% 0.55%
unidirectional_lockless/4096 uring bytes_per_sec 520.0 MB/s 522.5 MB/s 🟢▲ +0.46% 1.50%
unidirectional_lockless/65536 epoll bytes_per_sec 2.75 GB/s 2.82 GB/s 🟢▲ +0.82% 3.87%
unidirectional_lockless/65536 uring bytes_per_sec 2.77 GB/s 2.78 GB/s 🟢▲ +0.08% 1.19%
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 1.62K ops/s 🟢▲ +4.21% 0.34%
burst/100 iocp ops_per_sec 156.3 ops/s 162.2 ops/s 🟢▲ +3.62% 1.15%
burst_lockless/10 iocp ops_per_sec 1.56K ops/s 1.63K ops/s 🟢▲ +4.38% 0.36%
burst_lockless/100 iocp ops_per_sec 159.8 ops/s 166.0 ops/s 🟢▲ +4.09% 0.84%
concurrent/1 iocp ops_per_sec 12.53K ops/s 13.04K ops/s 🟢▲ +4.14% 0.38%
concurrent/16 iocp ops_per_sec 13.45K ops/s 13.93K ops/s 🟢▲ +3.46% 0.30%
concurrent/4 iocp ops_per_sec 13.17K ops/s 13.68K ops/s 🟢▲ +3.93% 0.41%
sequential iocp ops_per_sec 12.48K ops/s 12.96K ops/s 🟢▲ +3.83% 0.33%
sequential_lockless iocp ops_per_sec 12.56K ops/s 13.04K ops/s 🟢▲ +3.84% 0.22%

fan_out

Benchmark Backend Metric Base Head Δ Noise
concurrent_parents/1 iocp ops_per_sec 7.32K ops/s 7.35K ops/s 🟢▲ +0.32% 0.31%
concurrent_parents/16 iocp ops_per_sec 6.81K ops/s 6.81K ops/s 🔴▼ -0.16% 1.37%
concurrent_parents/4 iocp ops_per_sec 7.24K ops/s 7.28K ops/s 🟢▲ +0.42% 0.73%
concurrent_parents_lockless/1 iocp ops_per_sec 7.34K ops/s 7.35K ops/s 🟢▲ +0.12% 0.28%
concurrent_parents_lockless/16 iocp ops_per_sec 6.81K ops/s 6.83K ops/s 🟢▲ +0.61% 1.37%
concurrent_parents_lockless/4 iocp ops_per_sec 7.25K ops/s 7.27K ops/s 🟢▲ +0.23% 0.68%
fork_join/1 iocp ops_per_sec 111.9K ops/s 112.0K ops/s 🟢▲ +0.20% 0.37%
fork_join/16 iocp ops_per_sec 7.32K ops/s 7.33K ops/s 🟢▲ +0.09% 0.47%
fork_join/4 iocp ops_per_sec 29.03K ops/s 29.00K ops/s 🔴▼ -0.18% 0.63%
fork_join/64 iocp ops_per_sec 1.80K ops/s 1.82K ops/s 🟢▲ +0.45% 2.06%
fork_join_lockless/1 iocp ops_per_sec 112.2K ops/s 112.5K ops/s 🟢▲ +0.26% 0.65%
fork_join_lockless/16 iocp ops_per_sec 7.32K ops/s 7.33K ops/s 🟢▲ +0.07% 0.42%
fork_join_lockless/4 iocp ops_per_sec 29.07K ops/s 29.01K ops/s 🔴▼ -0.23% 0.24%
fork_join_lockless/64 iocp ops_per_sec 1.81K ops/s 1.82K ops/s 🟢▲ +0.47% 1.09%
nested/16 iocp ops_per_sec 1.79K ops/s 1.80K ops/s 🟢▲ +0.36% 1.06%
nested/4 iocp ops_per_sec 7.20K ops/s 7.20K ops/s 🔴▼ -0.14% 0.42%
nested_lockless/16 iocp ops_per_sec 1.79K ops/s 1.80K ops/s 🟢▲ +0.52% 0.84%
nested_lockless/4 iocp ops_per_sec 7.20K ops/s 7.21K ops/s 🟢▲ +0.07% 0.26%

http_server

Benchmark Backend Metric Base Head Δ Noise
concurrent/1 iocp ops_per_sec 111.0K ops/s 110.7K ops/s 🟢▲ +0.05% 0.60%
concurrent/16 iocp ops_per_sec 107.9K ops/s 107.8K ops/s 🟢▲ +0.05% 0.33%
concurrent/32 iocp ops_per_sec 107.2K ops/s 107.4K ops/s 🟢▲ +0.19% 0.34%
concurrent/4 iocp ops_per_sec 108.5K ops/s 108.3K ops/s 🔴▼ -0.12% 0.48%
multithread/1 iocp ops_per_sec 107.0K ops/s 107.4K ops/s 🟢▲ +0.37% 0.28%
multithread/16 iocp ops_per_sec 263.2K ops/s 262.7K ops/s 🔴▼ -0.16% 0.67%
multithread/2 iocp ops_per_sec 180.1K ops/s 179.7K ops/s 🔴▼ -0.38% 0.61%
multithread/4 iocp ops_per_sec 223.4K ops/s 224.6K ops/s 🟢▲ +0.53% 0.73%
multithread/8 iocp ops_per_sec 262.3K ops/s 263.0K ops/s 🟢▲ +0.02% 0.53%
single_conn iocp ops_per_sec 110.9K ops/s 110.7K ops/s 🟢▲ +0.14% 0.73%
single_conn_lockless iocp ops_per_sec 111.0K ops/s 110.6K ops/s 🔴▼ -0.07% 0.62%

socket_latency

Benchmark Backend Metric Base Head Δ Noise
concurrent/1 iocp latency_mean_ns 7.71 µs 7.63 µs 🟢▲ +0.95% 0.45%
concurrent/16 iocp latency_mean_ns 137.13 µs 130.86 µs 🟢▲ +2.55% 5.71%
concurrent/4 iocp latency_mean_ns 30.82 µs 30.61 µs 🟢▲ +0.32% 0.60%
concurrent_lockless/1 iocp latency_mean_ns 7.69 µs 7.65 µs 🟢▲ +0.55% 0.27%
concurrent_lockless/16 iocp latency_mean_ns 123.43 µs 122.59 µs 🟢▲ +0.40% 0.84%
concurrent_lockless/4 iocp latency_mean_ns 30.79 µs 30.50 µs 🟢▲ +0.86% 0.20%
pingpong/1 iocp latency_mean_ns 7.70 µs 7.64 µs 🟢▲ +0.96% 0.74%
pingpong/1024 iocp latency_mean_ns 7.79 µs 7.74 µs 🟢▲ +0.70% 0.57%
pingpong/64 iocp latency_mean_ns 7.69 µs 7.63 µs 🟢▲ +0.63% 0.37%
pingpong_lockless/1 iocp latency_mean_ns 7.66 µs 7.61 µs 🟢▲ +0.36% 0.45%
pingpong_lockless/1024 iocp latency_mean_ns 7.79 µs 7.71 µs 🟢▲ +1.27% 0.53%
pingpong_lockless/64 iocp latency_mean_ns 7.69 µs 7.64 µs 🟢▲ +0.40% 0.46%

socket_throughput

Benchmark Backend Metric Base Head Δ Noise
bidirectional/1024 iocp bytes_per_sec 277.6 MB/s 279.9 MB/s 🟢▲ +0.82% 0.67%
bidirectional/1048576 iocp bytes_per_sec 4.86 GB/s 5.03 GB/s 🟢▲ +1.61% 13.85%
bidirectional/16384 iocp bytes_per_sec 3.72 GB/s 3.74 GB/s 🟢▲ +0.22% 0.76%
bidirectional/262144 iocp bytes_per_sec 9.19 GB/s 9.70 GB/s 🟢▲ +1.11% 8.39%
bidirectional/4096 iocp bytes_per_sec 1.07 GB/s 1.08 GB/s 🟢▲ +0.84% 0.44%
bidirectional/65536 iocp bytes_per_sec 9.00 GB/s 9.14 GB/s 🟢▲ +1.40% 0.94%
bidirectional_lockless/1024 iocp bytes_per_sec 277.3 MB/s 281.3 MB/s 🟢▲ +1.22% 0.71%
bidirectional_lockless/1048576 iocp bytes_per_sec 4.96 GB/s 5.26 GB/s 🟢▲ +0.37% 14.71%
bidirectional_lockless/16384 iocp bytes_per_sec 3.68 GB/s 3.72 GB/s 🟢▲ +1.27% 0.64%
bidirectional_lockless/262144 iocp bytes_per_sec 9.28 GB/s 9.68 GB/s 🟢▲ +1.02% 9.21%
bidirectional_lockless/4096 iocp bytes_per_sec 1.06 GB/s 1.08 GB/s 🟢▲ +1.70% 0.34%
bidirectional_lockless/65536 iocp bytes_per_sec 8.84 GB/s 9.03 GB/s 🟢▲ +2.55% 1.08%
multithread/2 iocp bytes_per_sec 12.65 GB/s 12.64 GB/s 🟢▲ +0.36% 0.85%
multithread/4 iocp bytes_per_sec 16.78 GB/s 16.84 GB/s 🔴▼ -0.03% 0.78%
multithread/8 iocp bytes_per_sec 18.37 GB/s 18.45 GB/s 🟢▲ +0.59% 0.73%
unidirectional/1024 iocp bytes_per_sec 281.1 MB/s 285.6 MB/s 🟢▲ +1.69% 0.68%
unidirectional/1048576 iocp bytes_per_sec 4.15 GB/s 3.97 GB/s 🔴▼ -4.21% 3.78%
unidirectional/16384 iocp bytes_per_sec 3.78 GB/s 3.81 GB/s 🟢▲ +0.78% 0.56%
unidirectional/262144 iocp bytes_per_sec 8.55 GB/s 8.95 GB/s 🟢▲ +3.16% 9.54%
unidirectional/4096 iocp bytes_per_sec 1.08 GB/s 1.09 GB/s 🟢▲ +0.82% 0.48%
unidirectional/65536 iocp bytes_per_sec 9.20 GB/s 9.30 GB/s 🟢▲ +0.48% 1.29%
unidirectional_lockless/1024 iocp bytes_per_sec 282.3 MB/s 284.7 MB/s 🟢▲ +0.52% 0.53%
unidirectional_lockless/1048576 iocp bytes_per_sec 4.51 GB/s 4.50 GB/s 🟢▲ +0.70% 4.15%
unidirectional_lockless/16384 iocp bytes_per_sec 3.76 GB/s 3.74 GB/s 🔴▼ -0.35% 1.73%
unidirectional_lockless/262144 iocp bytes_per_sec 9.34 GB/s 9.81 GB/s 🟢▲ +1.34% 8.31%
unidirectional_lockless/4096 iocp bytes_per_sec 1.08 GB/s 1.09 GB/s 🟢▲ +0.93% 0.87%
unidirectional_lockless/65536 iocp bytes_per_sec 9.19 GB/s 9.21 GB/s 🟢▲ +0.92% 2.46%
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 4.61K ops/s 🟢▲ +0.10% 1.25%
burst/100 kqueue ops_per_sec 465.7 ops/s 466.4 ops/s 🟢▲ +0.30% 0.61%
burst_lockless/10 kqueue ops_per_sec 4.66K ops/s 4.64K ops/s 🔴▼ -0.27% 0.78%
burst_lockless/100 kqueue ops_per_sec 469.7 ops/s 467.6 ops/s 🔴▼ -0.34% 0.81%
concurrent/1 kqueue ops_per_sec 20.72K ops/s 20.67K ops/s 🔴▼ -0.25% 1.72%
concurrent/16 kqueue ops_per_sec 37.86K ops/s 37.59K ops/s 🔴▼ -0.75% 1.05%
concurrent/4 kqueue ops_per_sec 29.55K ops/s 29.45K ops/s 🟢▲ +0.35% 0.95%
sequential kqueue ops_per_sec 20.98K ops/s 20.98K ops/s 🟢▲ +0.12% 0.49%
sequential_lockless kqueue ops_per_sec 20.83K ops/s 20.79K ops/s 🔴▼ -0.21% 1.25%

fan_out

Benchmark Backend Metric Base Head Δ Noise
concurrent_parents/1 kqueue ops_per_sec 8.19K ops/s 8.17K ops/s 🔴▼ -0.20% 1.15%
concurrent_parents/16 kqueue ops_per_sec 6.63K ops/s 6.56K ops/s 🔴▼ -3.26% 19.34%
concurrent_parents/4 kqueue ops_per_sec 8.45K ops/s 8.41K ops/s 🔴▼ -0.13% 1.42%
concurrent_parents_lockless/1 kqueue ops_per_sec 8.29K ops/s 8.30K ops/s 🟢▲ +0.28% 1.32%
concurrent_parents_lockless/16 kqueue ops_per_sec 6.53K ops/s 6.53K ops/s 🔴▼ -1.45% 20.46%
concurrent_parents_lockless/4 kqueue ops_per_sec 8.54K ops/s 8.53K ops/s 🔴▼ -0.24% 1.39%
fork_join/1 kqueue ops_per_sec 49.10K ops/s 49.01K ops/s 🔴▼ -0.14% 0.91%
fork_join/16 kqueue ops_per_sec 8.21K ops/s 8.20K ops/s 🟢▲ +0.29% 1.30%
fork_join/4 kqueue ops_per_sec 23.57K ops/s 23.60K ops/s 🔴▼ -0.02% 0.57%
fork_join/64 kqueue ops_per_sec 2.15K ops/s 2.13K ops/s 🔴▼ -0.16% 2.51%
fork_join_lockless/1 kqueue ops_per_sec 50.53K ops/s 50.34K ops/s 🔴▼ -0.34% 0.63%
fork_join_lockless/16 kqueue ops_per_sec 8.32K ops/s 8.30K ops/s 🟢▲ +0.40% 1.37%
fork_join_lockless/4 kqueue ops_per_sec 23.73K ops/s 23.67K ops/s 🔴▼ -0.09% 0.57%
fork_join_lockless/64 kqueue ops_per_sec 2.17K ops/s 2.15K ops/s 🔴▼ -0.28% 2.37%
nested/16 kqueue ops_per_sec 2.13K ops/s 2.11K ops/s 🔴▼ -0.09% 2.37%
nested/4 kqueue ops_per_sec 8.12K ops/s 8.10K ops/s 🔴▼ -0.12% 1.26%
nested_lockless/16 kqueue ops_per_sec 2.15K ops/s 2.13K ops/s 🔴▼ -0.47% 2.37%
nested_lockless/4 kqueue ops_per_sec 8.25K ops/s 8.23K ops/s 🟢▲ +0.62% 1.49%

http_server

Benchmark Backend Metric Base Head Δ Noise
concurrent/1 kqueue ops_per_sec 49.79K ops/s 49.72K ops/s 🔴▼ -0.10% 0.35%
concurrent/16 kqueue ops_per_sec 130.2K ops/s 130.4K ops/s 🟢▲ +0.29% 0.54%
concurrent/32 kqueue ops_per_sec 129.7K ops/s 129.4K ops/s 🔴▼ -0.84% 1.10%
concurrent/4 kqueue ops_per_sec 89.95K ops/s 90.32K ops/s 🟢▲ +0.03% 1.09%
multithread/1 kqueue ops_per_sec 129.8K ops/s 129.5K ops/s 🔴▼ -0.03% 1.00%
multithread/16 kqueue ops_per_sec 132.6K ops/s 132.9K ops/s 🟢▲ +0.28% 0.51%
multithread/2 kqueue ops_per_sec 184.1K ops/s 184.7K ops/s 🟢▲ +0.57% 0.67%
multithread/4 kqueue ops_per_sec 162.9K ops/s 162.7K ops/s 🔴▼ -0.13% 0.29%
multithread/8 kqueue ops_per_sec 138.5K ops/s 137.9K ops/s 🔴▼ -0.41% 1.35%
single_conn kqueue ops_per_sec 49.84K ops/s 49.82K ops/s 🔴▼ -0.19% 0.29%
single_conn_lockless kqueue ops_per_sec 50.20K ops/s 49.87K ops/s 🔴▼ -0.90% 0.59%

local_socket_latency

Benchmark Backend Metric Base Head Δ Noise
concurrent/1 kqueue latency_mean_ns 1.83 µs 1.83 µs 🔴▼ -0.63% 0.67%
concurrent/16 kqueue latency_mean_ns 27.26 µs 27.29 µs 🔴▼ -0.33% 0.61%
concurrent/4 kqueue latency_mean_ns 6.87 µs 6.85 µs 🟢▲ +0.21% 0.44%
concurrent_lockless/1 kqueue latency_mean_ns 1.81 µs 1.81 µs 🔴▼ -0.87% 0.57%
concurrent_lockless/16 kqueue latency_mean_ns 27.09 µs 27.11 µs 🔴▼ -0.20% 0.79%
concurrent_lockless/4 kqueue latency_mean_ns 6.79 µs 6.81 µs 🔴▼ -0.15% 0.92%
pingpong/1 kqueue latency_mean_ns 1.83 µs 1.82 µs 🟢▲ +0.18% 0.43%
pingpong/1024 kqueue latency_mean_ns 1.93 µs 1.92 µs 🟢▲ +0.49% 0.41%
pingpong/64 kqueue latency_mean_ns 1.83 µs 1.83 µs 🔴▼ -0.36% 0.54%
pingpong_lockless/1 kqueue latency_mean_ns 1.81 µs 1.81 µs 🔴▼ -0.30% 0.57%
pingpong_lockless/1024 kqueue latency_mean_ns 1.91 µs 1.91 µs 🟢▲ +0.18% 0.25%
pingpong_lockless/64 kqueue latency_mean_ns 1.81 µs 1.81 µs 🔴▼ -0.30% 0.62%

local_socket_throughput

Benchmark Backend Metric Base Head Δ Noise
bidirectional/1024 kqueue bytes_per_sec 1.30 GB/s 1.30 GB/s 🔴▼ -0.12% 0.78%
bidirectional/1048576 kqueue bytes_per_sec 3.92 GB/s 3.92 GB/s 🟢▲ +0.04% 0.68%
bidirectional/16384 kqueue bytes_per_sec 3.92 GB/s 3.91 GB/s 🔴▼ -0.04% 0.92%
bidirectional/262144 kqueue bytes_per_sec 3.92 GB/s 3.92 GB/s 🟢▲ +0.07% 0.72%
bidirectional/4096 kqueue bytes_per_sec 2.94 GB/s 2.94 GB/s 🟢▲ +0.08% 0.45%
bidirectional/65536 kqueue bytes_per_sec 3.89 GB/s 3.92 GB/s 🟢▲ +0.94% 0.80%
bidirectional_lockless/1024 kqueue bytes_per_sec 1.30 GB/s 1.30 GB/s 🟢▲ +0.23% 0.59%
bidirectional_lockless/1048576 kqueue bytes_per_sec 3.98 GB/s 3.98 GB/s 🟢▲ +0.63% 1.07%
bidirectional_lockless/16384 kqueue bytes_per_sec 3.99 GB/s 3.99 GB/s 🟢▲ +0.39% 0.75%
bidirectional_lockless/262144 kqueue bytes_per_sec 4.00 GB/s 3.99 GB/s 🔴▼ -0.13% 0.80%
bidirectional_lockless/4096 kqueue bytes_per_sec 2.99 GB/s 2.98 GB/s 🔴▼ -0.05% 0.52%
bidirectional_lockless/65536 kqueue bytes_per_sec 3.99 GB/s 3.99 GB/s 🟢▲ +0.07% 0.79%
unidirectional/1024 kqueue bytes_per_sec 1.27 GB/s 1.28 GB/s 🟢▲ +0.13% 0.61%
unidirectional/1048576 kqueue bytes_per_sec 3.54 GB/s 3.55 GB/s 🟢▲ +0.29% 0.53%
unidirectional/16384 kqueue bytes_per_sec 3.55 GB/s 3.55 GB/s 🟢▲ +0.07% 0.25%
unidirectional/262144 kqueue bytes_per_sec 3.54 GB/s 3.54 GB/s 🔴▼ -0.11% 0.46%
unidirectional/4096 kqueue bytes_per_sec 2.73 GB/s 2.73 GB/s 🟢▲ +0.25% 0.41%
unidirectional/65536 kqueue bytes_per_sec 3.55 GB/s 3.55 GB/s 🔴▼ -0.27% 0.40%
unidirectional_lockless/1024 kqueue bytes_per_sec 1.28 GB/s 1.28 GB/s 🟢▲ +0.18% 0.37%
unidirectional_lockless/1048576 kqueue bytes_per_sec 3.61 GB/s 3.61 GB/s 🟢▲ +0.03% 0.53%
unidirectional_lockless/16384 kqueue bytes_per_sec 3.63 GB/s 3.62 GB/s 🔴▼ -0.41% 0.46%
unidirectional_lockless/262144 kqueue bytes_per_sec 3.61 GB/s 3.62 GB/s 🔴▼ -0.08% 0.31%
unidirectional_lockless/4096 kqueue bytes_per_sec 2.78 GB/s 2.78 GB/s 🔴▼ -0.10% 0.55%
unidirectional_lockless/65536 kqueue bytes_per_sec 3.62 GB/s 3.62 GB/s 🟢▲ +0.08% 0.46%

socket_latency

Benchmark Backend Metric Base Head Δ Noise
concurrent/1 kqueue latency_mean_ns 18.66 µs 18.66 µs 🟢▲ +0.01% 0.18%
concurrent/16 kqueue latency_mean_ns 121.11 µs 120.61 µs 🟢▲ +0.46% 1.02%
concurrent/4 kqueue latency_mean_ns 43.15 µs 43.14 µs 🔴▼ -0.62% 1.06%
concurrent_lockless/1 kqueue latency_mean_ns 18.64 µs 18.67 µs 🟢▲ +0.03% 0.48%
concurrent_lockless/16 kqueue latency_mean_ns 121.10 µs 120.56 µs 🔴▼ -0.03% 1.19%
concurrent_lockless/4 kqueue latency_mean_ns 42.45 µs 42.55 µs 🔴▼ -0.42% 1.02%
pingpong/1 kqueue latency_mean_ns 15.53 µs 15.37 µs 🟢▲ +0.10% 1.43%
pingpong/1024 kqueue latency_mean_ns 19.01 µs 19.00 µs 🟢▲ +0.09% 0.15%
pingpong/64 kqueue latency_mean_ns 18.72 µs 18.66 µs 🟢▲ +0.31% 0.27%
pingpong_lockless/1 kqueue latency_mean_ns 15.34 µs 15.35 µs 🔴▼ -0.08% 0.12%
pingpong_lockless/1024 kqueue latency_mean_ns 19.06 µs 19.01 µs 🟢▲ +0.16% 0.36%
pingpong_lockless/64 kqueue latency_mean_ns 18.67 µs 18.64 µs 🟢▲ +0.07% 0.36%

socket_throughput

Benchmark Backend Metric Base Head Δ Noise
bidirectional/1024 kqueue bytes_per_sec 475.1 MB/s 474.4 MB/s 🟢▲ +0.02% 0.73%
bidirectional/1048576 kqueue bytes_per_sec 7.80 GB/s 7.66 GB/s 🔴▼ -0.77% 4.44%
bidirectional/16384 kqueue bytes_per_sec 4.55 GB/s 4.53 GB/s 🔴▼ -0.51% 1.89%
bidirectional/262144 kqueue bytes_per_sec 8.81 GB/s 8.73 GB/s 🔴▼ -0.92% 2.06%
bidirectional/4096 kqueue bytes_per_sec 1.82 GB/s 1.82 GB/s 🟢▲ +0.13% 1.02%
bidirectional/65536 kqueue bytes_per_sec 7.59 GB/s 7.52 GB/s 🔴▼ -1.01% 2.28%
bidirectional_lockless/1024 kqueue bytes_per_sec 475.1 MB/s 473.5 MB/s 🔴▼ -0.66% 0.98%
bidirectional_lockless/1048576 kqueue bytes_per_sec 7.85 GB/s 7.68 GB/s 🔴▼ -0.61% 4.90%
bidirectional_lockless/16384 kqueue bytes_per_sec 4.56 GB/s 4.56 GB/s 🔴▼ -0.21% 1.26%
bidirectional_lockless/262144 kqueue bytes_per_sec 8.87 GB/s 8.78 GB/s 🔴▼ -1.05% 1.63%
bidirectional_lockless/4096 kqueue bytes_per_sec 1.83 GB/s 1.82 GB/s 🔴▼ -0.96% 0.75%
bidirectional_lockless/65536 kqueue bytes_per_sec 7.57 GB/s 7.63 GB/s 🟢▲ +2.11% 2.67%
multithread/2 kqueue bytes_per_sec 7.54 GB/s 7.53 GB/s 🔴▼ -0.55% 5.32%
multithread/4 kqueue bytes_per_sec 3.60 GB/s 3.66 GB/s 🔴▼ -0.84% 4.52%
multithread/8 kqueue bytes_per_sec 3.19 GB/s 3.33 GB/s 🟢▲ +2.32% 7.67%
unidirectional/1024 kqueue bytes_per_sec 459.8 MB/s 456.9 MB/s 🔴▼ -0.48% 2.52%
unidirectional/1048576 kqueue bytes_per_sec 8.97 GB/s 8.75 GB/s 🟢▲ +0.03% 6.22%
unidirectional/16384 kqueue bytes_per_sec 4.47 GB/s 4.45 GB/s 🟢▲ +0.98% 2.83%
unidirectional/262144 kqueue bytes_per_sec 8.87 GB/s 8.76 GB/s 🟢▲ +1.15% 3.30%
unidirectional/4096 kqueue bytes_per_sec 1.80 GB/s 1.80 GB/s 🟢▲ +1.08% 2.10%
unidirectional/65536 kqueue bytes_per_sec 7.40 GB/s 7.39 GB/s 🟢▲ +1.09% 2.67%
unidirectional_lockless/1024 kqueue bytes_per_sec 478.8 MB/s 475.6 MB/s 🔴▼ -0.40% 2.41%
unidirectional_lockless/1048576 kqueue bytes_per_sec 9.02 GB/s 8.79 GB/s 🔴▼ -0.49% 5.25%
unidirectional_lockless/16384 kqueue bytes_per_sec 4.47 GB/s 4.50 GB/s 🟢▲ +1.57% 2.88%
unidirectional_lockless/262144 kqueue bytes_per_sec 9.01 GB/s 8.89 GB/s 🔴▼ -1.07% 4.02%
unidirectional_lockless/4096 kqueue bytes_per_sec 1.81 GB/s 1.80 GB/s 🔴▼ -0.16% 1.14%
unidirectional_lockless/65536 kqueue bytes_per_sec 7.45 GB/s 7.36 GB/s 🔴▼ -1.41% 2.05%

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

The polling thread takes queued register and cancel requests by
swapping its own buffer with the one callers fill. It swaps twice per
wait in the common case, so callers kept filling the same buffer and
the other one stayed empty. The first request to land on the empty one
allocated on the caller's thread, whenever the polling thread happened
to be slow to come back around, which could be long after the reactor
had settled.

The buffer handed back now gets the capacity of the one taken, so both
grow together on the polling thread the first time a size is reached.
The zero-allocation suite ran on every backend except IOCP, so nothing
checked that accept, connect, read, write, close, timers, random-access
file I/O and tcp_server churn stay allocation-free on Windows. It is
now registered there, along with the post(continuation&) check, which
IOCP has satisfied since it began posting the continuation's address
as the completion packet.

A new test covers readiness waits on every backend. It waits in both
directions because IOCP serves a read wait with a zero-byte WSARecv and
a write wait through the wait reactor.

A Windows executable cannot count what a corosio DLL allocates, so the
suites skip with a notice in shared builds there instead of passing
without looking.

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.

4 participants