NURL HTTP-server peer-comparison

Hardware: Intel Core i7-5930K @ 3.50 GHz (6 physical / 12 logical cores), Ubuntu 24.04 (kernel 6.17), Linux x86_64, all loopback. Compilers: Rust 1.82.0 (hyper 1.9.0, tokio 1.52, lto=fat), Node v24.15.0 (built-in http). Load generator: oha 1.8.0 (HTTP/1.1, keep-alive enabled).

>

Numbers are median of 3 × 10 s runs per cell. Reproduce with bench/run_http.sh --concurrencies "1 10 50 200" --duration 10 (ITERS=3 by default).

Headline numbers

The path under test is a GET / returning Hello, World!\n (14 bytes, text/plain; charset=utf-8). No router middleware, no JSON, no metrics. All three implementations accept a TCP connection, parse one HTTP/1.1 request, write the 14-byte body, keep-alive.

	Server	C = 1	C = 10	C = 50	C = 200
req/s	NURL	14 451	68 960	60 897	59 044
	Rust	14 507	47 703	86 699	114 694
	Node	8 708	16 726	17 108	15 555
p50 (ms)	NURL	0.06	0.08	0.10	0.11
	Rust	0.07	0.15	0.41	1.33
	Node	0.11	0.54	2.90	12.89
p99 (ms)	NURL	0.14	0.56	0.67	0.62
	Rust	0.11	1.16	2.82	6.19
	Node	0.22	1.98	6.13	20.95

(Best per row in bold. Higher is better for req/s; lower for latency.)

Reading the table

Single client (C = 1)

The three implementations are dominated by per-request CPU work. NURL and Rust hyper land within < 1 % of each other (14 451 vs 14 507). Node trails ~1.7×.

Light parallel (C = 10)

NURL's server_run_pool runs 10 worker threads (see bench/http_server.nu); 10 in-flight connections fit inside that pool. Rust's tokio multi-thread runtime defaults to one worker per logical core (12 here); at C = 10 the runtime is over-provisioned. NURL: 69 k/s, p99 0.56 ms. Rust: 48 k/s, p99 1.16 ms.

Moderate parallel (C = 50)

NURL's 10-worker pool is saturated (~62 k/s). Rust's multi-thread scheduler reaches 87 k/s. NURL p50 stays at 0.10 ms; Rust p50 0.41 ms.

High parallel (C = 200)

Rust hyper reaches 115 k/s by saturating every core. NURL holds 59 k/s — its 10-worker pool is the throughput ceiling. NURL p50 0.11 ms, p99 0.62 ms. Rust p99 6.19 ms.

Node plateaus at ~16 k/s with rising latency.

Configuration notes

• NURL runs 10 worker threads; Rust hyper runs one tokio worker per

logical core (12 here). Each implementation uses its conventional default configuration.

• Loopback HTTP/1.1 only. No HTTPS measured. No HTTP/2 measured.
• Go is not included. bench/run_http.sh reserves a lane for

bench/http_server.go.

• The hardware is 11 years old; absolute numbers will differ on a

modern CPU.