perf research: WebSocket

[crowlbot]

TL;DR

Honest finding: no high-impact architectural slowdown in Deno's WebSocket.

• Client side: Deno beats Node 22 by 6.5× at 64 KB binary frames and beats Bun by 2.2× at the same size. Small-text is 1.36× faster than Node and ~28% slower than Bun.
• Server side: Within 6–25% of Bun's server, with the biggest gap on small text. Not graduating — the gap is in unattributed native ticks (deno binary + libc, ~65% of total) and native flamegraphs are blocked on this host.

Headline ratios

Client side (msgs/s, higher is better), Deno reference server

Workload	Deno	Node 22	Node 23	Bun	Deno vs Node 22	Deno vs Bun
client_text_11 (11 B text)	111,128	81,646	66,595	153,162	1.36× faster	1.38× slower
client_bin_32 (32 B bin)	101,448	70,825	97,242	110,911	1.43× faster	1.09× slower
client_bin_4k (4 KB bin)	48,887	11,238	11,892	39,576	4.35× faster	1.23× faster
client_bin_64k (64 KB bin)	6,927	1,067	1,071	3,138	6.49× faster	2.21× faster

Server side (Deno client, Deno vs Bun server)

Workload	Deno server	Bun server	Bun vs Deno
client_text_11	111,128	136,218	Bun 1.23× faster
client_bin_32	101,448	123,858	Bun 1.22× faster
client_bin_4k	48,887	63,068	Bun 1.29× faster
client_bin_64k	6,927	7,368	Bun 1.06× faster

Where the time goes — V8 prof on Deno client

ticks  total  nonlib   name
 229   42.2%          /deno   (WS framing + mio/tokio)
 127   23.4%          libc    (recv() syscalls)
  11    2.0%   5.9%   Builtin: LoadIC
   7    1.3%   3.7%   Builtin: ArrayPrototypePush
   ... (no single JS hot path)

65% of ticks in native code. JS-side overhead is uneventful — no obvious hot path to attack.

Hypotheses considered

#	Hypothesis	Verdict
H1	Per-frame allocation dominates	Rejected — 64 KB frames are 6.5× faster than Node
H2	EventTarget dispatch overhead per message	Unranked — wrappedHandler at 0.4% of ticks
H3	Server-side small-message gap	Unranked — native attribution blocked

What's not here

• Node-as-server (would require ws npm package).
• Native flamegraph (perf_event_paranoid = 4, no sudo on host).
• Many-connection scaling.

Final ranking

none — no high-impact slowdown to graduate. The 25% small-text server gap is real but currently unattributed.

Layout

```
tools/perf_research/websocket/
README.md full report
micro/ws_server.js Deno echo server (reference)
micro/ws_server_bun.js Bun echo server (cross-check)
micro/ws_client.js universal WS client bench
profiles/ws_results.log raw bench output
profiles/ws_client.prof.txt V8 --prof for Deno client
profiles/versions.txt runtime versions + host caps
```