node:quic (HTTP/3)

Compares bun's node:quic against node's. Bun implements the API on lsquic, node on ngtcp2 + nghttp3; the JS surface is the same, so server.mjs and client.mjs run unmodified on either runtime and only the runtime executing them changes.

Two benchmarks, because either side can be the bottleneck:

• server bench — one client runtime drives both servers
• client bench — one server runtime answers both clients

Each request is an HTTP/3 GET answered with :status: 200 and a BODY_SIZE body, with CONCURRENCY requests in flight on a single session.

Running

Node needs to be built with QUIC — released builds report process.features.quic === false and node:quic throws ERR_UNKNOWN_BUILTIN_MODULE:

cd /path/to/node
./configure --experimental-quic
make -j$(nproc)

Then:

BUN=path/to/bun NODE=path/to/node/out/Release/node bun run.mjs

Knobs (env): COUNT (2000), CONCURRENCY (50), BODY_SIZE (0), ROUNDS (3).

Reports the best of ROUNDS runs per pairing: QUIC throughput is noisy, and the fastest round is the one least perturbed by an unrelated scheduler hiccup.

Notes

• key.pem/cert.pem are the same self-signed fixtures the node test suite uses; the client passes verifyPeer: "manual" so the handshake is measured rather than the certificate chain.
• The client counts a request as done when its response headers arrive.
• Single session, so this measures stream and packet handling rather than handshakes.