Skip to content

[C] Add ringzero-ws - WebSocket#815

Merged
MDA2AV merged 2 commits into
mainfrom
add-ringzero-ws
Jun 6, 2026
Merged

[C] Add ringzero-ws - WebSocket#815
MDA2AV merged 2 commits into
mainfrom
add-ringzero-ws

Conversation

@MDA2AV
Copy link
Copy Markdown
Owner

@MDA2AV MDA2AV commented Jun 6, 2026

Description

A WebSocket echo server written directly on raw io_uring (liburing), with the WebSocket protocol hand-rolled — no tokio-tungstenite, no library WS stack. It's the io_uring sibling of the ringzero HTTP engine.

Added as an engine-tier entry: the lowest-level way to serve echo-ws on the modern Linux completion-based I/O path.

The "ring-zero" I/O path

  • One io_uring per core, each with its own SO_REUSEPORT listener — kernel-sharded accept, no shared queue, no cross-core work-stealing.
  • Multishot accept — one SQE yields every new connection.
  • Multishot recv + provided buffer ring — the kernel writes incoming bytes straight into a registered slab (io_uring_setup_buf_ring); frames are parsed in place and the buffer is recycled immediately (zero-copy ingest).
  • IORING_SETUP_SINGLE_ISSUER | DEFER_TASKRUN for the single-thread-per-ring fast path (with a fallback for older kernels). Echoes are batched into a per-connection write queue; the in-flight chunk is sealed so the kernel never sees a reallocated buffer.

Hand-rolled WebSocket

  • RFC 6455 handshake — request parsing, Sec-WebSocket-Accept derivation, 101 reply, with from-scratch SHA-1 + base64. liburing is the only dependency.
  • Frame codec — 7/16/64-bit lengths, client→server unmasking, partial frames carried across reads. Echoes re-emitted as unmasked server frames preserving FIN + opcode. Ping→Pong, Close echoed.

Tests

Subscribes to echo-ws and echo-ws-pipeline. Passes validate-ws.py (7/7), plus 200 KB-frame (multi-buffer reassembly), 2,000-frame pipelined, and 50-connection concurrent stress tests via the Docker build.

🤖 Generated with Claude Code

@MDA2AV @claude
[C] Add ringzero-ws - WebSocket
3f00819 
Hand-rolled WebSocket echo server on raw io_uring (liburing), with no
WebSocket library. Multishot accept + multishot recv into a provided
buffer ring give zero-copy ingest; the RFC 6455 handshake (with
from-scratch SHA-1 + base64), frame parser/masking, and batched echo
write path are implemented directly against io_uring completions. One
ring per core with SO_REUSEPORT, matching the ringzero engine.

Subscribes to echo-ws and echo-ws-pipeline. Passes validate-ws.py (7/7)
plus large-frame / pipelined / concurrent stress tests via the Docker
build.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
@MDA2AV
Copy link
Copy Markdown
Owner Author

MDA2AV commented Jun 6, 2026

/benchmark -f ringzero-ws

@github-actions
Copy link
Copy Markdown
Contributor

github-actions Bot commented Jun 6, 2026

👋 /benchmark request received. A collaborator will review and approve the run.

@github-actions
Copy link
Copy Markdown
Contributor

github-actions Bot commented Jun 6, 2026

Benchmark Results

Framework: ringzero-ws | Test: all tests

Test Conn RPS CPU Mem Δ RPS Δ Mem
echo-ws 512 4,360,772 6143.2% 211MiB NEW NEW
echo-ws 4096 4,481,908 6181.8% 230MiB NEW NEW
echo-ws 16384 4,123,360 6405.9% 337MiB NEW NEW
echo-ws-pipeline 512 66,785,414 6294.6% 207MiB NEW NEW
echo-ws-pipeline 4096 68,784,556 6408.1% 232MiB NEW NEW
echo-ws-pipeline 16384 61,210,198 6399.0% 336MiB NEW NEW
Full log 
  Thread Stats   Avg      p50      p90      p99    p99.9
    Latency    122us    117us    176us    220us    257us

  333805776 frames sent in 5.00s, 333927072 frames received
  Throughput: 66.75M req/s
  Bandwidth:  445.42MB/s
  WS upgrades: 512
  WS frames:   333927072
  Latency samples: 333852573 / 333927072 responses (100.0%)
[info] CPU 6294.6% | Mem 207MiB

[run 3/3]
gcannon v0.5.3 [WS]
  Target:    localhost:8080/ws
  Threads:   64
  Conns:     512 (8/thread)
  Pipeline:  16
  Req/conn:  unlimited (keep-alive)
  Expected:  200
  Duration:  5s


  Thread Stats   Avg      p50      p90      p99    p99.9
    Latency    122us    118us    173us    225us    273us

  332576368 frames sent in 5.00s, 332576368 frames received
  Throughput: 66.50M req/s
  Bandwidth:  443.91MB/s
  WS upgrades: 512
  WS frames:   332576368
  Latency samples: 332574832 / 332576368 responses (100.0%)
[info] CPU 6246.1% | Mem 202MiB

=== Best: 66785414 req/s (CPU: 6294.6%, Mem: 207MiB) ===
[info] saved results/echo-ws-pipeline/512/ringzero-ws.json
httparena-bench-ringzero-ws
httparena-bench-ringzero-ws

==============================================
=== ringzero-ws / echo-ws-pipeline / 4096c (tool=gcannon) ===
==============================================
[info] ws-only framework — skipping HTTP probe (sleep 2s for startup)

[run 1/3]
gcannon v0.5.3 [WS]
  Target:    localhost:8080/ws
  Threads:   64
  Conns:     4096 (64/thread)
  Pipeline:  16
  Req/conn:  unlimited (keep-alive)
  Expected:  200
  Duration:  5s


  Thread Stats   Avg      p50      p90      p99    p99.9
    Latency    944us    891us   1.18ms   1.30ms   1.49ms

  343834384 frames sent in 5.00s, 343771920 frames received
  Throughput: 68.73M req/s
  Bandwidth:  458.92MB/s
  WS upgrades: 4096
  WS frames:   343771920
  Latency samples: 343771920 / 343771920 responses (100.0%)
[info] CPU 6100.5% | Mem 230MiB

[run 2/3]
gcannon v0.5.3 [WS]
  Target:    localhost:8080/ws
  Threads:   64
  Conns:     4096 (64/thread)
  Pipeline:  16
  Req/conn:  unlimited (keep-alive)
  Expected:  200
  Duration:  5s


  Thread Stats   Avg      p50      p90      p99    p99.9
    Latency    951us    896us   1.18ms   1.33ms   1.48ms

  343953520 frames sent in 5.00s, 343922784 frames received
  Throughput: 68.76M req/s
  Bandwidth:  459.10MB/s
  WS upgrades: 4096
  WS frames:   343922784
  Latency samples: 343920288 / 343922784 responses (100.0%)
[info] CPU 6408.1% | Mem 232MiB

[run 3/3]
gcannon v0.5.3 [WS]
  Target:    localhost:8080/ws
  Threads:   64
  Conns:     4096 (64/thread)
  Pipeline:  16
  Req/conn:  unlimited (keep-alive)
  Expected:  200
  Duration:  5s


  Thread Stats   Avg      p50      p90      p99    p99.9
    Latency    963us    955us   1.17ms   1.31ms   1.44ms

  339809088 frames sent in 5.00s, 339808912 frames received
  Throughput: 67.94M req/s
  Bandwidth:  453.54MB/s
  WS upgrades: 4096
  WS frames:   339808912
  Latency samples: 339859056 / 339808912 responses (100.0%)
[info] CPU 6246.4% | Mem 231MiB

=== Best: 68784556 req/s (CPU: 6408.1%, Mem: 232MiB) ===
[info] saved results/echo-ws-pipeline/4096/ringzero-ws.json
httparena-bench-ringzero-ws
httparena-bench-ringzero-ws

==============================================
=== ringzero-ws / echo-ws-pipeline / 16384c (tool=gcannon) ===
==============================================
[info] ws-only framework — skipping HTTP probe (sleep 2s for startup)

[run 1/3]
gcannon v0.5.3 [WS]
  Target:    localhost:8080/ws
  Threads:   64
  Conns:     16384 (256/thread)
  Pipeline:  16
  Req/conn:  unlimited (keep-alive)
  Expected:  200
  Duration:  5s


  Thread Stats   Avg      p50      p90      p99    p99.9
    Latency   4.30ms   4.33ms   4.73ms   5.10ms   10.00ms

  298734448 frames sent in 5.00s, 298475520 frames received
  Throughput: 59.68M req/s
  Bandwidth:  398.79MB/s
  WS upgrades: 16384
  WS frames:   298475520
  Latency samples: 298475520 / 298475520 responses (100.0%)
[info] CPU 5989.3% | Mem 337MiB

[run 2/3]
gcannon v0.5.3 [WS]
  Target:    localhost:8080/ws
  Threads:   64
  Conns:     16384 (256/thread)
  Pipeline:  16
  Req/conn:  unlimited (keep-alive)
  Expected:  200
  Duration:  5s


  Thread Stats   Avg      p50      p90      p99    p99.9
    Latency   4.27ms   4.32ms   4.69ms   5.19ms   7.69ms

  306313136 frames sent in 5.00s, 306050992 frames received
  Throughput: 61.18M req/s
  Bandwidth:  408.85MB/s
  WS upgrades: 16384
  WS frames:   306050992
  Latency samples: 306050992 / 306050992 responses (100.0%)
[info] CPU 6399.0% | Mem 336MiB

[run 3/3]
gcannon v0.5.3 [WS]
  Target:    localhost:8080/ws
  Threads:   64
  Conns:     16384 (256/thread)
  Pipeline:  16
  Req/conn:  unlimited (keep-alive)
  Expected:  200
  Duration:  5s


  Thread Stats   Avg      p50      p90      p99    p99.9
    Latency   4.28ms   4.31ms   4.71ms   5.33ms   7.64ms

  304985024 frames sent in 5.00s, 304726976 frames received
  Throughput: 60.92M req/s
  Bandwidth:  407.07MB/s
  WS upgrades: 16384
  WS frames:   304726976
  Latency samples: 304726976 / 304726976 responses (100.0%)
[info] CPU 6147.6% | Mem 336MiB

=== Best: 61210198 req/s (CPU: 6399.0%, Mem: 336MiB) ===
[info] saved results/echo-ws-pipeline/16384/ringzero-ws.json
httparena-bench-ringzero-ws
httparena-bench-ringzero-ws
[info] rebuilding site/data/*.json
[updated] /home/diogo/actions-runner/_work/HttpArena/HttpArena/site/data/frameworks.json
[updated] /home/diogo/actions-runner/_work/HttpArena/HttpArena/site/data/echo-ws-16384.json
[updated] /home/diogo/actions-runner/_work/HttpArena/HttpArena/site/data/echo-ws-4096.json
[updated] /home/diogo/actions-runner/_work/HttpArena/HttpArena/site/data/echo-ws-512.json
[updated] /home/diogo/actions-runner/_work/HttpArena/HttpArena/site/data/echo-ws-pipeline-16384.json
[updated] /home/diogo/actions-runner/_work/HttpArena/HttpArena/site/data/echo-ws-pipeline-4096.json
[updated] /home/diogo/actions-runner/_work/HttpArena/HttpArena/site/data/echo-ws-pipeline-512.json
[updated] /home/diogo/actions-runner/_work/HttpArena/HttpArena/site/data/current.json
[info] done
[info] restoring loopback MTU to 65536

@MDA2AV
Copy link
Copy Markdown
Owner Author

MDA2AV commented Jun 6, 2026

/benchmark -f ringzero-ws --save

@github-actions
Copy link
Copy Markdown
Contributor

github-actions Bot commented Jun 6, 2026

👋 /benchmark request received. A collaborator will review and approve the run.

@github-actions
Copy link
Copy Markdown
Contributor

github-actions Bot commented Jun 6, 2026

Benchmark Results

Framework: ringzero-ws | Test: all tests

Test Conn RPS CPU Mem Δ RPS Δ Mem
echo-ws 512 4,353,147 6200.6% 210MiB NEW NEW
echo-ws 4096 4,429,548 6408.1% 231MiB NEW NEW
echo-ws 16384 4,066,852 6398.4% 336MiB NEW NEW
echo-ws-pipeline 512 66,683,702 6289.2% 203MiB NEW NEW
echo-ws-pipeline 4096 68,438,198 6160.2% 228MiB NEW NEW
echo-ws-pipeline 16384 61,261,692 6136.3% 337MiB NEW NEW
Full log 
  Thread Stats   Avg      p50      p90      p99    p99.9
    Latency    123us    118us    177us    224us    294us

  330936768 frames sent in 5.00s, 330974784 frames received
  Throughput: 66.17M req/s
  Bandwidth:  441.71MB/s
  WS upgrades: 512
  WS frames:   330974784
  Latency samples: 331043631 / 330974784 responses (100.0%)
[info] CPU 6284.1% | Mem 208MiB

[run 3/3]
gcannon v0.5.3 [WS]
  Target:    localhost:8080/ws
  Threads:   64
  Conns:     512 (8/thread)
  Pipeline:  16
  Req/conn:  unlimited (keep-alive)
  Expected:  200
  Duration:  5s


  Thread Stats   Avg      p50      p90      p99    p99.9
    Latency    122us    118us    173us    218us    253us

  333369664 frames sent in 5.00s, 333418511 frames received
  Throughput: 66.66M req/s
  Bandwidth:  444.97MB/s
  WS upgrades: 512
  WS frames:   333418511
  Latency samples: 333368045 / 333418511 responses (100.0%)
[info] CPU 6289.2% | Mem 203MiB

=== Best: 66683702 req/s (CPU: 6289.2%, Mem: 203MiB) ===
[info] saved results/echo-ws-pipeline/512/ringzero-ws.json
httparena-bench-ringzero-ws
httparena-bench-ringzero-ws

==============================================
=== ringzero-ws / echo-ws-pipeline / 4096c (tool=gcannon) ===
==============================================
[info] ws-only framework — skipping HTTP probe (sleep 2s for startup)

[run 1/3]
gcannon v0.5.3 [WS]
  Target:    localhost:8080/ws
  Threads:   64
  Conns:     4096 (64/thread)
  Pipeline:  16
  Req/conn:  unlimited (keep-alive)
  Expected:  200
  Duration:  5s


  Thread Stats   Avg      p50      p90      p99    p99.9
    Latency    954us    899us   1.19ms   1.42ms   1.52ms

  342253456 frames sent in 5.00s, 342190992 frames received
  Throughput: 68.42M req/s
  Bandwidth:  456.82MB/s
  WS upgrades: 4096
  WS frames:   342190992
  Latency samples: 342190992 / 342190992 responses (100.0%)
[info] CPU 6160.2% | Mem 228MiB

[run 2/3]
gcannon v0.5.3 [WS]
  Target:    localhost:8080/ws
  Threads:   64
  Conns:     4096 (64/thread)
  Pipeline:  16
  Req/conn:  unlimited (keep-alive)
  Expected:  200
  Duration:  5s


  Thread Stats   Avg      p50      p90      p99    p99.9
    Latency    958us    908us   1.19ms   1.38ms   1.50ms

  341772192 frames sent in 5.00s, 341772176 frames received
  Throughput: 68.33M req/s
  Bandwidth:  456.13MB/s
  WS upgrades: 4096
  WS frames:   341772176
  Latency samples: 341798352 / 341772176 responses (100.0%)
[info] CPU 6395.0% | Mem 230MiB

[run 3/3]
gcannon v0.5.3 [WS]
  Target:    localhost:8080/ws
  Threads:   64
  Conns:     4096 (64/thread)
  Pipeline:  16
  Req/conn:  unlimited (keep-alive)
  Expected:  200
  Duration:  5s


  Thread Stats   Avg      p50      p90      p99    p99.9
    Latency    957us    911us   1.17ms   1.31ms   1.47ms

  341946544 frames sent in 5.00s, 341944480 frames received
  Throughput: 68.36M req/s
  Bandwidth:  456.37MB/s
  WS upgrades: 4096
  WS frames:   341944480
  Latency samples: 341943424 / 341944480 responses (100.0%)
[info] CPU 6219.4% | Mem 231MiB

=== Best: 68438198 req/s (CPU: 6160.2%, Mem: 228MiB) ===
[info] saved results/echo-ws-pipeline/4096/ringzero-ws.json
httparena-bench-ringzero-ws
httparena-bench-ringzero-ws

==============================================
=== ringzero-ws / echo-ws-pipeline / 16384c (tool=gcannon) ===
==============================================
[info] ws-only framework — skipping HTTP probe (sleep 2s for startup)

[run 1/3]
gcannon v0.5.3 [WS]
  Target:    localhost:8080/ws
  Threads:   64
  Conns:     16384 (256/thread)
  Pipeline:  16
  Req/conn:  unlimited (keep-alive)
  Expected:  200
  Duration:  5s


  Thread Stats   Avg      p50      p90      p99    p99.9
    Latency   4.30ms   4.34ms   4.72ms   5.09ms   5.95ms

  297993056 frames sent in 5.00s, 297730912 frames received
  Throughput: 59.52M req/s
  Bandwidth:  397.75MB/s
  WS upgrades: 16384
  WS frames:   297730912
  Latency samples: 297730912 / 297730912 responses (100.0%)
[info] CPU 5979.2% | Mem 337MiB

[run 2/3]
gcannon v0.5.3 [WS]
  Target:    localhost:8080/ws
  Threads:   64
  Conns:     16384 (256/thread)
  Pipeline:  16
  Req/conn:  unlimited (keep-alive)
  Expected:  200
  Duration:  5s


  Thread Stats   Avg      p50      p90      p99    p99.9
    Latency   4.26ms   4.31ms   4.68ms   5.06ms   7.53ms

  306212352 frames sent in 5.00s, 305950208 frames received
  Throughput: 61.16M req/s
  Bandwidth:  408.72MB/s
  WS upgrades: 16384
  WS frames:   305950208
  Latency samples: 305950208 / 305950208 responses (100.0%)
[info] CPU 6400.2% | Mem 337MiB

[run 3/3]
gcannon v0.5.3 [WS]
  Target:    localhost:8080/ws
  Threads:   64
  Conns:     16384 (256/thread)
  Pipeline:  16
  Req/conn:  unlimited (keep-alive)
  Expected:  200
  Duration:  5s


  Thread Stats   Avg      p50      p90      p99    p99.9
    Latency   4.25ms   4.28ms   4.67ms   5.06ms   10.50ms

  306566512 frames sent in 5.00s, 306308464 frames received
  Throughput: 61.23M req/s
  Bandwidth:  409.18MB/s
  WS upgrades: 16384
  WS frames:   306308464
  Latency samples: 306308464 / 306308464 responses (100.0%)
[info] CPU 6136.3% | Mem 337MiB

=== Best: 61261692 req/s (CPU: 6136.3%, Mem: 337MiB) ===
[info] saved results/echo-ws-pipeline/16384/ringzero-ws.json
httparena-bench-ringzero-ws
httparena-bench-ringzero-ws
[info] rebuilding site/data/*.json
[updated] /home/diogo/actions-runner/_work/HttpArena/HttpArena/site/data/frameworks.json
[updated] /home/diogo/actions-runner/_work/HttpArena/HttpArena/site/data/echo-ws-16384.json
[updated] /home/diogo/actions-runner/_work/HttpArena/HttpArena/site/data/echo-ws-4096.json
[updated] /home/diogo/actions-runner/_work/HttpArena/HttpArena/site/data/echo-ws-512.json
[updated] /home/diogo/actions-runner/_work/HttpArena/HttpArena/site/data/echo-ws-pipeline-16384.json
[updated] /home/diogo/actions-runner/_work/HttpArena/HttpArena/site/data/echo-ws-pipeline-4096.json
[updated] /home/diogo/actions-runner/_work/HttpArena/HttpArena/site/data/echo-ws-pipeline-512.json
[updated] /home/diogo/actions-runner/_work/HttpArena/HttpArena/site/data/current.json
[info] done
[info] restoring loopback MTU to 65536

@MDA2AV MDA2AV merged commit 5a3f45a into main Jun 6, 2026
@MDA2AV MDA2AV deleted the add-ringzero-ws branch June 7, 2026 14:31
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@Kaliumhexacyanoferrat Kaliumhexacyanoferrat Awaiting requested review from Kaliumhexacyanoferrat Kaliumhexacyanoferrat is a code owner

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@MDA2AV