Numbers

Every figure on this site is read from bench/site-numbers.json. Each row names the command that produced it.

Linux

bench/linux-three-way.sh, p16s, Core Ultra 7 165H, Linux 6.17, Docker 29.1.3, Podman 4.9.3 rootless, 2026-11-07. Source: docs/PERFORMANCE.md.

ImageMetriccorailDockerPodman
alpinewarm run --rm /bin/true, median of 206 ms214 ms110 ms
alpineexec /bin/true, median of 202 ms53 ms105 ms
alpinehost footprint of 50 idle sleeps54 MB1.2 GB107 MB
busyboxwarm run --rm /bin/true, median of 207 ms289 ms147 ms
busyboxexec /bin/true, median of 204 ms69 ms119 ms
busyboxhost footprint of 50 idle sleeps53 MB1.2 GB106 MB
debianwarm run --rm /bin/true, median of 206 ms265 ms135 ms
debianexec /bin/true, median of 202 ms61 ms116 ms
debianhost footprint of 50 idle sleeps53 MB1.2 GB106 MB

macOS

bench/mac-image-start.py, bench/mac-image-rest.py, bench/mac-image-idle.py, Apple M4 Pro, macOS 26.6.2, Docker Desktop 4.98.0 (engine 29.9.2), Podman 6.1.4 applehv, 4 vCPU, 2 GiB, 2026-11-16. Source: docs/PERFORMANCE.md, section macOS three-way, refreshed 2026-11-16.

ImageMetriccorailDockerPodman
alpinewarm run --rm /bin/true, median of 205.4 ms132 ms152 ms
alpineexec /bin/true, median of 205.2 ms33 ms115 ms
alpinehost footprint of 50 idle sleeps196 MB1.5 GB1.3 GB
busyboxwarm run --rm /bin/true, median of 205.2 ms140 ms155 ms
busyboxexec /bin/true, median of 204.2 ms31 ms114 ms
busyboxhost footprint of 50 idle sleeps190 MB1.4 GB1.3 GB
debianwarm run --rm /bin/true, median of 206.6 ms145 ms159 ms
debianexec /bin/true, median of 204.3 ms31 ms113 ms
debianhost footprint of 50 idle sleeps340 MB1.4 GB1.3 GB

Targets and the control plane

A target is a target, not a result. The measured column is from just bench-check with SITE_NUMBERS=1; a row says "not re-recorded" when that run has no figure for it.

MetricMeasuredTargetCommand, commit
warm run --rm /bin/true, median of 105 ms≤ 9 msjust bench-check (run --rm alpine -- /bin/true, median of 10), p16s d21ab26
exec /bin/true, median of 102 ms≤ 15 ms (guard)just bench-check (exec /bin/true, median of 10), p16s d21ab26
recify apply to Running6 ms≤ 50 ms, guardedbench/coraild.sh, p16s d21ab26
kill to recovered8 ms≤ 300 ms, guardedbench/coraild.sh, p16s d21ab26
coraild RSS at 50 atolls3.6 MB< 16 MBbench/coraild.sh, p16s d21ab26

The comparison against Kubernetes (bench/vs-k8s.sh) has not been run for this site, so no Kubernetes figure appears here.

How we measure

  • Tools are sampled interleaved, and the tool that goes first rotates each round.
  • Three warm-up rounds are discarded. Latency rows use at least 20 samples and report the median.
  • The image is pulled on every side before timing starts.
  • A verdict needs a paired sign test, |z| ≥ 1.96, with at least 80% agreement (bench/ab-verdict.sh).
  • Memory is resident size summed over the daemon and each container's supervisor.
  • A network-bound figure, such as a cold pull, is only an absolute hang guard. It measures the link, not the code.
  • Rows where corail is worst are flagged as findings, not dropped.

Rules in full: bench/results/latest-p16s.md and docs/PERFORMANCE.md.