lang-bench

A Node.js / Go / Rust performance comparison toolkit — equivalent three-language implementations of the same benchmark suite, a unified runner, and a visualization page.

中文  |  Benchmarks · Quick Start · Metrics · Fairness

---

Benchmarks

Benchmark	Type	Scenario	Key metrics
fib	CLI	Recursive Fibonacci, pure CPU	Run time
json-parse	CLI	Parse + serialize ~9 MB event JSON	Run time, memory
vector-topk	CLI	RAG: brute-force cosine top-10 over 20k × 256-dim vectors	Run time
text-chunk	CLI	RAG: sliding-window token chunking + hashing	Run time, memory
http-server	Server	Agent tool-call gateway: POST JSON → parse → compute → JSON	req/s, p50/p99 latency
sse-stream	Server	LLM token stream relay (SSE) consumed concurrently	events/s, TTFB

Every CLI benchmark prints the same verification value (checksum / hash / sum) in all three languages — confirm they match before trusting any numbers. "Fast" never means "computed the wrong thing."

---

Quick Start

Requirements

Tool	Version	Notes
Node.js	≥ 20	Runner + visualization
Go	≥ 1.22
Rust	rustup (stable)
hyperfine	optional	brew install hyperfine — more rigorous timing

npm install          # installs autocannon for HTTP load testing
npm run bench:quick  # fast pass with small inputs (first Rust build may be slow)
npm run serve        # open http://localhost:4173 for the chart page
npm run bench        # full run with real parameters

Run a subset

# specific benchmarks
node runner/run.mjs fib vector-topk --lang=go,rust

# quick mode
node runner/run.mjs http-server --quick

# custom run count
node runner/run.mjs fib --runs=20

Results are written to results/run-<timestamp>.json and mirrored to results/latest.json (which the visualization reads by default). You can also drag any results JSON onto the page. The page UI supports Chinese/English via the toggle in the top-right corner.

---

Metrics

Metric	Description
Run time	Median of multiple runs after warmup. Hover any bar for mean / stddev / min / max.
Peak memory	CLI: max RSS via /usr/bin/time -l. Servers: max of ps samples every 200ms during load test.
Build time	Wall time of go build / cargo build --release. First Rust build includes all dependencies — cargo clean first for a cold-build measurement.
Artifact size	Stripped binaries for Go/Rust. Node has no binary, so the source file size is recorded (the runtime itself is ~80 MB — keep that in mind when comparing).

---

Fairness Notes

Read this before drawing conclusions.

• Rust builds with --release + strip. Go with -ldflags="-s -w". Node relies on the V8 JIT — warmup rounds ensure hot code is compiled.
• JSON benchmark: Go/Rust use typed structs (idiomatic). JS uses dynamic objects — that's how each language is actually used, not an unfairness.
• HTTP / SSE: Node and Go use their standard libraries. Rust's standard library has no HTTP server, so it uses the de-facto community standard axum (+ tokio).
• Load-test clients (autocannon / custom SSE client) are Node.js, applied identically to all three servers.
• Same machine, plugged in, heavy background programs closed. The results JSON records CPU / memory / language versions — never mix numbers from different machines.

---

Project Structure

lang-bench/
├── benchmarks/           # equivalent 3-language implementations
│   ├── fib/              #   {node,go,rust}/main.{mjs,go,rs}
│   ├── json-parse/
│   ├── vector-topk/
│   ├── text-chunk/
│   ├── http-server/
│   └── sse-stream/
├── runner/               # runner
│   ├── run.mjs           #   build → warmup → time → collect → JSON
│   ├── config.mjs        #   benchmark registry
│   ├── gen-fixtures.mjs  #   fixture generator
│   └── sse-client.mjs    #   concurrent SSE client
├── web/                  # visualization
│   ├── index.html        #   bar-chart comparison + i18n toggle
│   └── serve.mjs         #   minimal static server
├── fixtures/             # test data (auto-generated)
├── results/              # benchmark result JSONs
├── bin/                  # compiled Go binaries
└── package.json

---

Adding a Benchmark

1. Write three equivalent implementations under benchmarks/<name>/{node,go,rust}:
   • CLI type: args in → verification value printed to stdout.
   • Server type: read the PORT env var, expose GET /ping, print READY <port> to stderr once listening.
2. Register it in runner/config.mjs (type: 'cli' | 'http' | 'sse', default args and quick args).
3. Add a title line to I18N.*.bench in web/index.html and it will show up.

---

License

MIT