Benchmarks

Real machine, full methodology, every CSV linked. Reproduce yourself before quoting.

Three findings worth highlighting

Sequential head-to-head — same workload, every stack

Each runtime gets the full 12-core machine in isolation; we don't run them concurrently because that measures the scheduler instead of the framework. ab -n 50000 -c 200 -k -l, warmed up first.

vs Laravel 11: ~210× · vs Symfony 7: ~55× · vs Slim 4: ~28×

Concurrency sweep — ZealPHP solo across c = 1…1000

Same 4 workers, varying simultaneous connections. Shows where each endpoint saturates, how tail latency degrades, and whether throughput holds at heavy load.

/raw/bench — lean runtime, no demo middleware

/json — full PSR-15 stack (CORS · ETag · Range · sessions · reflection-injected handler)

Peak at c = 100, sustained well past it. Throughput holds within ~20% of peak at c = 1000 with zero failures — the framework degrades gracefully rather than falling over.
Low-concurrency throughput (c = 1, c = 10) is bounded by Docker localhost-network round-trip latency, not framework cost. Run on bare metal to see higher c = 1 numbers; the c ≥ 50 figures are unaffected.

Raw CSVs: /raw/bench · /json

Reproduce on your own machine

Numbers are hardware- and OS-bound. Published figures are a starting point, not a contract. Three harnesses ship with the repo; pick the one that matches the claim you want to verify.

One-line install (Ubuntu/Debian)

Goes from a fresh box to a benched-ready clone — installs PHP 8.3, OpenSwoole, uopz, composer, wrk, ab, then clones sibidharan/zealphp to ~/zealphp and runs composer install:

curl -fsSL https://php.zeal.ninja/bench-install.sh | sudo bash
# Prints the next-step bench command when it finishes.

Inspect before piping to sudo: curl -fsSL https://php.zeal.ninja/bench-install.sh | less

Manual install (macOS / inspect-friendly)

brew install wrk php composer node
pecl install openswoole uopz
git clone https://github.com/sibidharan/zealphp && cd zealphp && composer install

curl -fsSL https://php.zeal.ninja/install.sh | sudo bash   # PHP + openswoole + uopz + composer
sudo apt install -y wrk apache2-utils git
git clone https://github.com/sibidharan/zealphp && cd zealphp && composer install

Verify extensions loaded: php -m | grep -E 'openswoole|uopz'

Recipe 1 — single-stack concurrency sweep (matches the tables above)

scripts/bench.sh --tool ab --requests 50000 \
                 --workers 4 --threads 4 --task-workers 0 \
                 --paths /raw/bench,/json --p1000
# Output: bench/results/zealphp-<timestamp>.csv + per-c raw logs

Recipe 2 — ZealPHP vs raw Node (matches the head-to-head table)

scripts/bench_compare.sh --workers 4 --threads 4 --p1000 --duration 30s
# Or via Docker so versions don't matter:
mkdir -p bench/results && docker compose run --rm --build compare

Recipe 3 — 3-way with sample-to-sample variance (autocannon)

A single 30s run can hide 10–15% per-sample swings on noisy hardware. This harness runs 10 short samples per stack spread over time and reports mean ± stddev so you can see how stable each stack is.

cd /tmp && npm install autocannon express   # one-off
./bench/compare-3way/run.sh                 # ~10 min

Methodology

Endpoints under test

Caveats — read before quoting

• Single-machine numbers. Your hardware, OS limits, payload size, and middleware set will move these. Quote your own measurements.
• Docker localhost RTT. c = 1 and c = 10 throughput is bounded by Docker's localhost networking overhead, not framework cost. Bare-metal runs typically post c = 1 closer to 15k-20k req/s.
• 4 workers ≈ 4 cores. Deliberate baseline. The framework is multi-process; doubling workers on a wider machine scales further until you saturate I/O or coroutine context-switching.
• Express comparison is fair. Express runs with cors + etag + express-session + session-file-store + ejs + body-parser — middleware roughly equivalent to ZealPHP's built-in PSR-15 stack. We're not comparing bare Express to full-stack ZealPHP.
• "Other PHP frameworks" numbers are community benchmarks, not measured on this box. They're rough orders of magnitude; we don't claim 1.0% precision.

Source: PERF.md · Raw CSVs: bench/results/ryzen-sweep/ · Scripts: scripts/ · bench/compare-3way/