php-bench

Benchmarks for the docker-library/php official image: our compiled PHP vs. the same distro's own packaged PHP, across versions/OSes, targeting the specific "PHP is slow" complaints on file there:

• CPU: docker-library/php#493, docker-library/php#1431
• Imagick resize: docker-library/php#1100
• TLS/CA-verification cost: docker-library/php#1431
• apache vs fpm throughput: docker-library/php#681 (mpm_event, discussed in docker-library/php#742, isn't one of the modes -- the official apache image is NTS and Apache refuses to load it under a threaded MPM; see .targets.sh)

Use

./run.sh list                  # show the target matrix
./run.sh ensure-images         # build/pull every image up front (its own step in CI)
./run.sh bench cpu             # or tls, imagick, throughput
./run.sh bench cpu --only 8.4  # filter targets by substring
./run.sh report cpu            # results/cpu/*.json -> CSV on stdout
./run.sh summarize             # fixed comparisons + version-consistency checks

--pull[=never|missing|always] and --build[=never|missing|always] are global flags (valid anywhere in argv, for any command, e.g. run.sh bench cpu --pull) controlling whether third-party base images and our own overlay/distro-pkg images get refreshed or trusted as-is; a bare --pull/--build means always. Also settable via PHP_BENCH_PULL_POLICY/PHP_BENCH_BUILD_POLICY in the environment. They default differently, on purpose: --pull defaults to missing, since docker pull is a mandatory registry round-trip on every call (even when already current) multiplied across dozens of targets, for freshness we rarely need mid-session. --build defaults to always, since docker build has a real local cache-hit path (near-free once warm) and it's the only thing that catches our own Dockerfile changes -- editing install-imagick.sh doesn't touch any base tag, so --pull alone would never rebuild the overlay that actually needs it. See PULL_POLICY's comment in run.sh for the full reasoning.

Requires Docker. First run of bench imagick renders the fixed test image (bench/imagick/assets/bench.jpg) if it isn't already there.

Layout

• .targets.sh -- the version/OS/SAPI matrix.
• images/ -- Dockerfiles: debian-pkg/alpine-pkg (vanilla distro + its own packaged PHP), overlay (adds hyperfine, and curl/imagick compiled against our PHP when needed), asset-gen (one-off, for the Imagick test image).
• bench/<suite>/ -- the PHP scripts being timed. cpu/tls/imagick each have a run.sh that hyperfine-invokes them inside the target container, redirecting hyperfine's own verbose per-benchmark output and printing a compact one-line-per-result summary instead via bench/summarize-hyperfine.php (reads the same --export-json file, rather than parsing hyperfine's human-readable text); throughput is multi-container (server + nginx/fpm + a curl_multi load driver), so its orchestration lives in run.sh at the repo root instead.
• results/<suite>/*.json -- one file per target: hyperfine's --export-json output for cpu/tls/imagick, load.php's own JSON for throughput.
• summarize.jq -- reads all of the above and prints six fixed comparisons (one per question above), each collapsed across PHP version with a spread/consistency check rather than silently averaging it away. Runnable directly (./summarize.jq results/*/*.json) or via run.sh summarize.