Benchmarks

Relative numbers only: compare two commits on the same machine

Protocol

composer bench                                          # the demo app, path /
php benchmarks/boot.php demos/app /simple-module/       # an app and a path
php benchmarks/boot.php --synthetic=100 /site/thing3/action2/
php benchmarks/boot.php --runs=50 --requests=5000       # more samples

Two execution models are measured, as medians:

• fpm: one process per request (autoload, boot, handle), like PHP-FPM. Opcache is warm through its file cache, timestamps validated.
• worker: boot once, then many requests in the same process, like FrankenPHP or RoadRunner. This process runs without opcache, so its boot includes compilation.

--synthetic=N generates an application of 5 modules sharing N controllers of 6 actions each, to check how the boot scales with the size of the application.

Results

Windows 11, PHP 8.3, 2026-09-29.

App	fpm autoload	fpm boot	fpm request	worker request
demo (5 modules)	2.7 ms	7.0 ms	2.9 ms	0.074 ms
synthetic, 100 controllers	2.8 ms	6.8 ms	2.6 ms	0.022 ms
synthetic, 100 controllers, before hierarchical routing	2.6 ms	21.5 ms	2.1 ms	0.07 ms

Before hierarchical routing, the boot scanned every controller of every module for route attributes. It now only depends on the modules and their config.php, never on the number of controllers. See Runtime for why Kaly optimizes with workers rather than with a cache.