Generated 2026-10-22T06:07:03Z by bench/run_http3.sh. Do not edit by hand — the next run overwrites it.
The HTTP/3 companion of HTTP_RESULTS.md (HTTP/1.1) and HTTP2_RESULTS.md (HTTP/2), which stay as they are. Each implementation terminates QUIC (RFC 9000/9001/9002) on a UDP socket, speaks HTTP/3 (RFC 9114 + QPACK, RFC 9204) and answers every request on every stream with the same 14-byte Hello, World!\n body (text/plain), over a self-signed EC (P-256) certificate the generator accepts without verification.
The NURL server is bench/http_server.nu unchanged from the HTTP/1.1 and HTTP/2 benchmarks: the packages/http HttpApp facade's TLS listener also binds its port over UDP and serves HTTP/3 there through the same routes; the QUIC transport, the TLS 1.3 handshake inside it, QPACK and the HTTP/3 framing are all pure NURL. The Rust peer is quinn + h3 (h3-quinn) on tokio with rustls — the stack behind most Rust HTTP/3 servers.
Cells are C x M: C client connections, each with M concurrent streams (h2load --h3 -c C -m M), so C x M requests are in flight. The same closed-loop caveat as the HTTP/2 report applies: at high in-flight counts read req/s, not the latency columns; cells whose effective concurrency (req/s x mean-latency) falls well short of C x M are marked ‡ and left un-bold.
| Item | Value |
|---|---|
| Host | GitHub Actions ubuntu-latest runner |
| Kernel | Linux 7.0.0-1017-azure x86_64 |
| CPU | AMD EPYC 9V45 96-Core Processor (4 logical cores) |
| Memory | 16367520 KiB |
| Commit | 4cc2a80a00753b0ab965b79977e142b17d2cf2e9 |
| CI run | https://github.com/nurl-lang/nurl/actions/runs/39731345494 |
| NURL | v0.76.0-32-g4cc2a80a |
| Rust | rustc 1.99.1 (58cf8f9c3 2026-10-13) |
| Load generator | h2load nghttp2/1.70.0 (docker nghttp2-h3) |
| Setting | Value |
|---|---|
| Throughput/latency | median of 3 x 10 s closed-loop runs |
| Cells (C x M) | 1x1 , 1x10 , 1x100 , 10x1 , 10x10 , 50x1 , 50x10 |
| Connection setup | 100 fresh connections, one request each, median of 3 runs |
| TLS cert | self-signed EC P-256, CN=localhost, ALPN h3 |
| Server | 1 x 1 | 1 x 10 | 1 x 100 | 10 x 1 | 10 x 10 | 50 x 1 | 50 x 10 | |
|---|---|---|---|---|---|---|---|---|
| req/s | NURL | 25 089 | 141 088 | 261 060 | 68 802 | 215 718 | 66 656 | 213 784 |
| Rust | 44 | 50 401 | 90 049 | 404 | 64 700 | 1 925 | 23 591 | |
| p50 (ms) | NURL | 0.04 | 0.07 | 0.28‡ | 0.14 | 0.45 | 0.69 | 2.24 |
| Rust | 26.35 | 0.18 | 0.86‡ | 26.20 | 0.20 | 26.29 | 26.16 | |
| p99 (ms) | NURL | 0.05 | 0.09 | 0.41‡ | 0.21 | 0.70 | 1.15 | 3.87 |
| Rust | 26.63 | 0.51 | 1.60‡ | 27.76 | 26.05 | 27.23 | 27.65 |
‡ closed-loop starved (NURL 1x100: ~75.7 in flight; Rust 1x100: ~77.4 in flight).
The same binary and the same host:port — QUIC over UDP for HTTP/3, TLS over TCP with ALPN h2 for HTTP/2 — driven by the same h2load. The gap is the transports' own cost (packet protection and per-packet framing for QUIC, records for TLS/TCP) with everything else held equal. HTTP/1.1 on this listener is in HTTP_RESULTS.md (oha): h2load's HTTP/1.1 mode paces itself and would misreport it.
| C | HTTP/3 req/s | HTTP/2 req/s | HTTP/3 / HTTP/2 | HTTP/3 p50 (ms) | HTTP/2 p50 (ms) |
|---|---|---|---|---|---|
| 1 | 25 089 | 1 000 | 25.09x | 0.04 | 0.04 |
| 10 | 68 802 | 10 000 | 6.88x | 0.14 | 0.12 |
| 50 | 66 656 | 44 560 | 1.50x | 0.69 | 0.57 |
100 clients each open a fresh connection, make one request and close: a QUIC handshake (Initial + Handshake, one round trip, keys derived on both ends) for HTTP/3, a TLS 1.3 handshake over a new TCP connection for HTTP/2. Connections per second, median of the runs.
| Server | Protocol | conn/s |
|---|---|---|
| NURL | HTTP/3 (QUIC) | 1 665 |
| NURL | HTTP/2 (TLS+TCP) | 1 466 |
| Rust | HTTP/3 (QUIC) | 1 620 |
(Best per column in bold; latency winners are chosen only among non-starved cells. ‡ = closed-loop starved. n/a = tool absent; FAIL = the server did not complete that cell.)
h2load built with ngtcp2 + nghttp3 (the distribution package is HTTP/2-only), run from a docker image built from nghttp2's own docker/Dockerfile. oha, the generator of the other two reports, has no HTTP/3 client.bench/rust_http3_server/ (quinn + h3 + rustls), run with quinn's default transport settings.max_ack_delay (25 ms) later, and ngtcp2's client does not open its next request stream until the packet that carried the previous one is acknowledged — so with few streams in flight every request costs ~25 ms regardless of the server's work. NURL bundles the pending ACK into any packet it sends (RFC 9000 §13.2.1), which is why its columns do not show it. The cells with many streams per connection are the ones where both servers are actually busy.tools/h3spec_gate.sh runs h3spec (49/49: 34 QUIC-transport + 15 HTTP/3 and QPACK error cases) against the same NURL HttpApp listener in CI. A fast server that fails h3spec would not be listed as a win.tc netem loss/delay) — the QUIC recovery machinery is not exercised on loopback.recvmmsg / GSO) and core isolation, as in the torture harness.