Seven PHP frameworks under identical load: the gap narrows as soon as the request does real work

./bench/build.sh && STAMP=$(date -u +%F) ./bench/run.sh --all

Most PHP framework comparisons publish a number without publishing how it was produced: which version, which hardware, how many repetitions. What this record contributes is not the ranking but the fact that the ranking can be re-derived. Every figure below comes out of bench/run.sh and every raw run is committed in the repository.

What was measured

Each candidate ran at its latest stable release, from its official skeleton, in its own documented production set-up. Exactly one thing varies: the framework.

Candidate	Version	Data layer	Templates
Slim	4.15.2	none — raw PDO	none — slim/php-view
Yii2	2.0.55	ActiveRecord	Yii View
Phalcon	5.19.0	Phalcon Model	Volt
Symfony	8.1.4	Doctrine ORM	Twig
Laminas	3.8.0	laminas-db TableGateway	laminas-view
CodeIgniter	4.7.4	CI Model	CI View
Laravel	13.26.1	Eloquent	Blade

All seven register exactly fifty routes, with the parameterised route always at position 47, so route resolution is measured in a realistic table rather than a one-entry one. The scenarios: plain text, JSON, a parameterised route, a five-layer middleware chain, a hundred-row template (no database), a single row by primary key, and a twenty-row list.

What the numbers say

Data table

Best of 9 repetitions, 64 connections, 15 s — bench/run.sh output

Series	Slim	Yii2	Phalcon	Symfony	Laminas	CodeIgniter	Laravel
Empty route	25,975	19,834	15,021	13,067	8,349	6,431	5,966
Template (100 rows)	22,780	16,995	10,563	9,355	7,419	5,557	5,290
Database (single row)	14,443	9,366	5,222	6,972	5,421	4,527	3,887

The chart is drawn in the browser; the table below carries the same data.

The most striking movement is Phalcon’s. Third on the empty route (15,021), third again on the template, and fifth once a database query is involved (5,222) — below Symfony. On the twenty-row list it slips further, to sixth (2,040), behind Laminas and CodeIgniter. A framework written in C gains nothing while the process is waiting on MySQL; what it gains is where PHP would be running PHP.

Candidate	Empty route	JSON	Route (47/50)	Middleware	Template	DB single row	DB 20 rows
Slim	25,975	25,609	24,774	24,020	22,780	14,443	4,551
Yii2	19,834	18,731	18,775	18,877	16,995	9,366	3,245
Phalcon	15,021	15,161	15,061	14,906	10,563	5,222	2,040
Symfony	13,067	12,714	12,187	12,964	9,355	6,972	3,060
Laminas	8,349	8,114	8,127	8,036	7,419	5,421	2,874
CodeIgniter	6,431	6,233	5,960	6,037	5,557	4,527	2,494
Laravel	5,966	5,848	5,603	5,704	5,290	3,887	1,293

The closing of the gap fits in one line: the ratio of fastest to slowest is 4.4× on an empty route, 3.7× on a single-row database request and 3.5× on a twenty-row one. Phalcon’s lead over Laravel falls from 2.5× to 1.3×.

Latency

Data table

Taken from the repetition that produced the best throughput, 64 connections

Series	Slim	Yii2	Phalcon	Symfony	Laminas	CodeIgniter	Laravel
p50	0.25 ms	0.35 ms	0.46 ms	0.58 ms	0.88 ms	1.16 ms	0.75 ms
p95	0.43 ms	0.59 ms	0.73 ms	0.89 ms	1.52 ms	1.84 ms	1.97 ms
p99	0.53 ms	0.73 ms	0.94 ms	0.96 ms	2.23 ms	1.98 ms	2.59 ms

The chart is drawn in the browser; the table below carries the same data.

Laravel’s p50 (0.75 ms) is better than Laminas’s and CodeIgniter’s, while its p99 (2.59 ms) is the worst in the field. Laravel finishes the typical request quickly and handles the tail badly.

The arguments are cheap, the defaults are not

Resolving a parameterised route in a fifty-route table costs between 0% and 7% against an empty route. A five-layer middleware chain costs 1% to 8%. The two things developers argue about most when comparing frameworks turn out to be single-digit percentages.

Then you look at what Laravel ships by default:

Laravel route	req/s	p50	p99
No middleware group	5,858	1.32 ms	2.88 ms
Default web group session + cookie encryption + CSRF	2,176	1.73 ms	71.89 ms

This was measured with the file session driver. Laravel’s shipped default is database, which adds a read and a write per request on top. The real figure is worse than the one above.

Saturation: the ceiling is eight workers

Raising the connection count from 8 to 256 left throughput and p99 flat for all seven.

Connections	Slim	Yii2	Phalcon	Symfony	Laminas	CodeIgniter	Laravel
8	25,812	19,242	14,894	12,817	8,149	6,193	5,909
32	25,785	19,018	15,053	12,811	7,828	6,259	4,774
64	26,234	18,964	14,802	9,324	8,114	6,276	5,890
128	25,842	19,251	14,783	11,795	8,044	6,067	6,016
256	25,802	19,808	15,109	12,129	8,075	6,140	5,901

How it was measured — and why it was restarted twice

This measurement was run three times. The first two were abandoned, both found a real defect, and both are kept in the repository with the reasoning.

First abort — the machine was not quiet. Slim’s /json gave 15,260 req/s in one repetition and 2,107 in the next, with p50 at 0.23 ms in both. Identical per-request cost, seven times fewer requests: the application had not slowed down, the run had not been given CPU. The host is a laptop running its window server, a terminal and an editor; cpusets isolate a container inside the VM, they cannot isolate the VM from macOS.

The fix for that is not to take the average. Interference can only ever subtract throughput — a run can be slowed by a noisy neighbour, but it cannot be faster than the hardware allows. So the reported figure is the best of the repetitions, with the median and the spread published beside it. A fixed nginx reference that never reaches PHP is also sampled before and after every block: across 150 samples the worst was 84% of the best.

Second abort — port exhaustion. Half the requests in the database scenarios were returning 500: SQLSTATE[HY000] [2002] Cannot assign requested address. php-fpm opens a fresh TCP connection to MySQL on every request and the container runs out of ephemeral ports at a few thousand requests a second. The single pre-flight request could never catch that. All seven moved to persistent connections (PDO::ATTR_PERSISTENT, pConnect for CodeIgniter) — the documented production setting for php-fpm. The effect is large: Slim’s /db went from 3,594 to 14,074 req/s, which means the earlier measurement was measuring port churn rather than the data layer.

Findings that did not fit the table

1. There is exactly one PHP version all seven agree on: 8.4. Symfony 8.1 requires at least 8.4.1; laminas/laminas-mvc 3.8 accepts no more than ~8.4.0. The window is one minor version wide.
2. Phalcon and Symfony cannot share a PHP process. Phalcon requires ext-psr, and with that extension loaded Symfony is fatal on PHP 8.4: Declaration of Symfony\Component\Cache\CacheItem::expiresAt(?DateTimeInterface $expiration): static must be compatible with PsrExt\Cache\CacheItemInterface::expiresAt($expiration). The C extension declares the PSR interfaces without the parameter types the userland packages use. The “one image, one extension set” plan therefore forks in exactly one place: Phalcon runs on its own runtime image, the other six share one.
3. CodeIgniter 4 has no PDO driver — its MySQL support is MySQLi only. ext-mysqli is installed for all seven so the extension set stays identical; the other six never use it.
4. Laminas’s official skeleton has not been released for PHP 8.4. laminas/laminas-mvc-skeleton 2.4.0 still declares ~8.1.0 || ~8.2.0 || ~8.3.0 and refuses to install without --ignore-platform-req=php. The component it boots, laminas/laminas-mvc 3.8.0, does accept ~8.4.0.
5. Yii2 is still 2.0.x. Yii3 has no stable release, so the latest stable Yii is a line whose architecture predates the others’ current versions.
6. Middleware is not the same concept in all seven. CodeIgniter filters and Yii2 action filters are scoped to a route and cost the other scenarios nothing. Symfony, Laminas and Phalcon have no per-route pipeline; the documented equivalent is a global event listener, so in those three all five layers are consulted on every request and return early. That is their architecture, not a handicap invented here.
7. Slim does less work in two scenarios. It has no data layer (/db runs on raw PDO) and no template engine of its own (slim/php-view). Its first place in those two rows should be read with that in mind.
8. One cell contradicted itself and was topped up. Slim’s /route came out at 6,778 req/s over nine repetitions — with the lowest p50 in the field (0.23 ms), at 26% of its own empty-route throughput. None of the nine had landed in an unclipped window. Two rules fixed in advance (two or fewer clean repetitions, or throughput below half of 8/p50) flagged six cells, all of them Slim or Yii2 — the fastest two, which is the direction the bias was predicted to run. Six repetitions were added to each; Slim’s /route settled at 24,774 and no cell remained flagged.
9. Zero non-200 responses. After the persistent-connection fix, all 546 load measurements returned HTTP 200 for every request.
10. JIT was off. None of the seven documents opcache JIT as a production step, and enabling it would have added a second variable. Results with JIT on may differ; that is a separate measurement.
11. Worker mode was not measured. Persistent-process models such as FrankenPHP or Octane are not officially supported by all seven; classic php-fpm is the one floor they share. Worker mode deserves its own record.