Network WebSocket Benchmarks

Numbers measured 2026-07-29. The tables report the median of three back‑to‑back runs on the same host.

Refresh these numbers after a substantive WebSocket transport change or dependency upgrade. Absolute numbers vary by machine; only same‑machine deltas are meaningful.

Environment

Measurement controls

• CPU governor: performance on all 128 logical CPUs.
• ASLR: disabled per process with setarch "$(uname -m)" -R.
• CPU scheduling: SMT and boost enabled; benchmark thread not pinned.
• Latency sampling: 1,000 warmup messages, then 50,000 measured messages.
• Throughput sampling: Criterion default warmup and 100 samples.
• Aggregation: median of three back‑to‑back runs per table cell.

How to reproduce

sudo cpupower frequency-set -g performance
for run in 1 2 3; do
    CARGO_BUILD_JOBS=16 setarch "$(uname -m)" -R \
        cargo bench -p nautilus-network --profile bench-lto \
        --bench websocket_transport -- --save-baseline "ws_run_$run" --noplot
    CARGO_BUILD_JOBS=16 NAUTILUS_WS_LATENCY_MESSAGES=50000 \
        setarch "$(uname -m)" -R \
        cargo bench -p nautilus-network --profile bench-lto \
        --bench websocket_latency
done
sudo cpupower frequency-set -g powersave

For policy and the general noise‑reduction recipe, see BENCHMARKING.md at the repository root.

Methodology

The benchmarks compare tokio-tungstenite 0.30.0 and sockudo-ws 2.0.1 in the same binary and measurement session.

• Both use established, uncompressed streams over identical 1 MiB in‑memory Tokio duplex transports and a current‑thread runtime through Sink and Stream.
• Sockudo enables simd, fastrand, tokio-runtime, and rustls-webpki-roots; auto_ping and idle_timeout are disabled to isolate frame transport.
• Throughput processes 10,000 text messages per Criterion iteration.
• Round‑trip latency spans client send through echo receive.
• One‑way burst latency timestamps each binary message from a continuous sender and includes in‑memory queueing and receiver backpressure.
• Each p99.9 value covers 50 observations per run; it is useful but noisier than p50, p95, or p99.

The benchmark excludes DNS, TCP connect, TLS, HTTP upgrade, kernel network I/O, and external network latency. It also excludes Compio, sockudo's native split‑stream driver, compression, and keepalive traffic.

Round-trip text latency

Lower is better. Values are microseconds.

Across these payloads, sockudo-ws 2.0.1 reduces p99 latency by 73-82% relative to tokio-tungstenite 0.30.0.

One-way binary burst latency

Lower is better. Values are microseconds and include queueing within the in‑memory transport.

At 512 B, sockudo-ws 2.0.1 reduces p99 burst latency by 15% relative to tokio-tungstenite 0.30.0.

Text throughput

Higher is better. Values are millions of messages per second.

At 512 B, sockudo-ws 2.0.1 processes 18% more receives, 13% more sends, and 250% more round trips than tokio-tungstenite 0.30.0.