Skip to content
PAM

Benchmark laboratory

PAM’s Laravel laboratory compares the same locked application route across:

  • PAM;
  • PHP-FPM plus Nginx;
  • Laravel Octane plus Swoole;
  • FrankenPHP; and
  • RoadRunner.

It does not compare unrelated empty handlers.

Run the complete matrix

Section titled “Run the complete matrix”
Terminal window
scripts/benchmark-laravel.sh

The runner builds every container, starts one runtime at a time, warms it, executes three measured rounds, and writes raw plus aggregated JSON under:

benchmarks/results/<UTC timestamp>

Override workload parameters without editing the protocol:

Terminal window
PAM_BENCH_ROUNDS=5 \
PAM_BENCH_WARMUP_SECONDS=20 \
PAM_BENCH_DURATION_SECONDS=60 \
PAM_BENCH_THREADS=4 \
PAM_BENCH_CONNECTIONS=128 \
PAM_BENCH_ENDPOINT=/api/ping \
scripts/benchmark-laravel.sh

Fixed contract

Section titled “Fixed contract”
  • same machine, kernel, CPU governor, and open-file limits;
  • same Laravel release, lockfile, route, environment, and response bytes;
  • production dependencies with authoritative autoloading;
  • four application workers, two application CPUs, and 1 GiB total memory;
  • no Xdebug, debug pages, access logs, tracing, or profilers;
  • same HTTP and TLS mode; and
  • load generation on a separate recorded CPU set when the host permits it.

FPM and Nginx split the same combined CPU and memory budget. Images, PHP, Composer, and runtime releases are pinned.

Captured evidence

Section titled “Captured evidence”

The vendored wrk.lua records:

  • requests per second;
  • p50, p75, p90, p95, p99, and maximum latency;
  • connection, read, write, timeout, and HTTP status errors;
  • container memory;
  • host metadata; and
  • the exact Git commit.

Repeat at multiple concurrency levels. Peak throughput is not useful when tail latency or errors collapse.

Memory stability

Section titled “Memory stability”

The Laravel integration test warms a worker, verifies request isolation, sends another 2,000 requests, and enforces bounded RSS growth:

Terminal window
cargo test --test laravel -- --nocapture

For a release candidate, extend the soak to millions of requests with the real database, authentication, logging, and package graph. Set --max-requests from measured RSS.