Performance & benchmarks

This page is a dated performance snapshot, not a timeless promise. It records what PAM 2.1.0 produced on GitHub-hosted runners on August 25, 2026, using source commit c9b09cfe.

The complete raw rounds, host metadata, runtime identities, logs, reports, and SHA-256 evidence manifest are attached to the public workflow runs:

• HTTP comparison: evidence suite 1
• 30-minute soak: evidence suite 3
• PAM 2.1.0 release

Headline

Across the four equivalent dynamic Laravel scenarios, PAM delivered 11.3% to 15.3% more median throughput than FrankenPHP and 32.1% to 36.8% more than OpenSwoole, with zero recorded request errors. The opt-in PAM edge cache served 74,620.54 requests/second at 2.84 ms median p99.

The edge-cache result is deliberately reported separately: it avoids full Laravel execution and is not equivalent to an uncached dynamic request.

Dynamic Laravel results

Each value is the median of five measured 15-second rounds after a 5-second warm-up per scenario. The load used 128 connections, four load threads, one application worker, two CPUs for the server, and two separate CPUs for wrk.

Scenario	Runtime	Median req/s	Median p99	Errors
Uncached	PAM 2.1.0	1,605.83	84.19 ms	0
Uncached	FrankenPHP	1,443.25	104.36 ms	0
Uncached	OpenSwoole	1,174.21	120.60 ms	0
Blade	PAM 2.1.0	1,231.76	108.29 ms	0
Blade	FrankenPHP	1,085.67	129.93 ms	0
Blade	OpenSwoole	916.31	150.95 ms	0
SQLite	PAM 2.1.0	1,242.56	106.46 ms	0
SQLite	FrankenPHP	1,111.64	141.32 ms	0
SQLite	OpenSwoole	940.38	145.57 ms	0
Large JSON	PAM 2.1.0	1,508.98	89.59 ms	0
Large JSON	FrankenPHP	1,308.18	109.31 ms	0
Large JSON	OpenSwoole	1,104.51	125.20 ms	0

PAM’s throughput ratio against FrankenPHP was 1.113× uncached, 1.135× for Blade, 1.118× for SQLite, and 1.153× for large JSON. Against OpenSwoole it was 1.368×, 1.344×, 1.321×, and 1.366× respectively.

Opt-in edge cache

Runtime path	Median req/s	Median p99	Errors
PAM edge cache	74,620.54	2.84 ms	0
FrankenPHP full Laravel execution	1,249.12	113.72 ms	0
OpenSwoole full Laravel execution	1,029.33	133.64 ms	0

This result passed the repository’s 5× edge-cache gate. It shows the value of explicitly cacheable responses; it must not be presented as a general 59× or 72× runtime claim.

30-minute stability soak

The soak ran PAM 2.1.0 with four workers, four load threads, and 256 connections for 30 minutes.

Measurement	Result	Budget
Throughput	3,680.25 req/s	Informational
Request errors	0	0
Resource samples	1,432	—
Baseline RSS	493.98 MiB	—
Final RSS	514.22 MiB	—
RSS growth	20.23 MiB	≤ 64 MiB
Soak gate	Passed	Required

RSS covers the supervised runtime process tree on the runner; it is not the memory cost of a single idle worker.

Test environment

The HTTP comparison ran on a four-vCPU GitHub-hosted Ubuntu runner backed by an AMD EPYC 7763 with approximately 16 GiB RAM. PAM used embedded PHP 8.5.9, OPcache, and JIT. The pinned FrankenPHP and OpenSwoole comparison images used PHP 8.4.24, also with OPcache and JIT.

The soak ran on another four-vCPU GitHub-hosted runner backed by an AMD EPYC 9V74 with approximately 16 GiB RAM. Cloud-runner results naturally contain host variance; use the raw rounds and relative MAD values rather than treating a single number as guaranteed capacity.

Read the limitations before quoting

The comparison is transparent but not a same-PHP-version certification: PAM used PHP 8.5.9 while the two pinned competitor images used PHP 8.4.24. The workflow metadata also records the source tree as dirty because the pinned PAM Native SDK is checked out inside the workspace during CI. The measured PAM binary and report are nevertheless tied to commit c9b09cfe and its SHA-256.

Two PAM scenarios, Blade and SQLite, exceeded the repository’s aspirational 100 ms dynamic-p99 gate at 108.29 ms and 106.46 ms. Therefore the evidence run succeeded and produced valid artifacts, but the report’s complete measurement gate did not pass. Do not describe this run as if every performance budget passed.

Reproduce it

From a checkout of push-in/pam at the measured commit:

# Five-round PAM, FrankenPHP, and OpenSwoole comparison
PAM_EVIDENCE_SUITE_ID=1 benchmarks/octane/run.sh

# Thirty-minute PAM stability run
PAM_SOAK_DURATION=30m benchmarks/octane/soak.sh

The CI workflow also installs the pinned tools, builds the measured release binary, verifies the evidence manifest, and retains the result bundle:

gh workflow run evidence.yml --repo push-in/pam --ref main -f suite_id=1
gh workflow run evidence.yml --repo push-in/pam --ref main -f suite_id=3

For capacity planning, rerun on the intended production CPU, kernel, PHP extensions, application code, database, and concurrency. See the Laravel benchmark laboratory for the broader fairness contract and captured metrics.