Benchmarks

2026-10-08 — Per-domain counters (ADR-0042): A/B gate

Host: development VM (8 vCPU EPYC 7642, Ubuntu 24.04). One binary (release, --features http3), two configurations: metrics.hosts: false (off, via bench/hosts-off-wrapper.sh, which adds the setting to the bench configuration and execs the binary) and the default metrics.hosts: true (on). Variants alternate within every run (bench/multi.sh), so host drift affects both alike.

Saturated throughput (`MODE=release bench/multi.sh`, 5 × 30 s, server on CPUs 0-3)

Scenario	off req/s (median, CV)	on req/s (median, CV)	Δ req/s	off CPU µs/req	on CPU µs/req	RSS MiB off / on
respond	101 592 (1.2 %)	102 130 (1.2 %)	+0.5 %	25.59	26.39	37 / 37
static-small	95 148 (3.8 %)	95 920 (2.9 %)	+0.8 %	39.53	39.20	37 / 37
proxy	76 709 (5.2 %)	78 252 (4.3 %)	+2.0 %	51.60	50.66	39 / 40
PHP	27 953 (11.7 %)	29 176 (10.3 %)	+4.4 %	89.65	85.68	45 / 44

No throughput regression in any scenario (gate: ≤ 1 %, 2 % hard limit). CPU per request from a saturated closed loop is noisy (signs differ between scenarios); it is measured at a fixed rate below.

CPU per request at a fixed rate (`bench/cpu-fixed-rate.sh`, 20 000 req/s, 6 alternating rounds, server on CPUs 0-1)

Scenario	off µs/req (median, min–max)	on µs/req (median, min–max)	Δ
respond (HTTP/1.1)	31.29 (29.08–32.08)	30.88 (30.28–31.30)	−1.3 %
static-small (HTTP/1.1)	38.22 (37.58–38.68)	37.91 (37.35–38.83)	−0.8 %
TLS + HTTP/2 static	31.43 (26.35–32.38)	31.24 (30.95–33.43)	−0.6 %

No increase (the differences are within the spread of the rounds). Gate passed: metrics.hosts is on by default.

Allocations and memory

The request path gains no allocation: counters are preallocated per 64 domains and event loop on first use. RSS is unchanged within 1 MiB (table above).

Invalid runs (not used)

The first A/B attempt measured a leftover server that held port 8080 (the variants under test exited with “address in use”, yet the harness saw answers). bench/run.sh now checks that the server process it started is alive before measuring. The second attempt used a build without the http3 feature, which the bench configuration needs; the new check stopped it.