kocsismate/php-version-benchmarks

Official PHP benchmark suite

dev-main 2026-08-01 09:48 UTC

MIT

• Máté Kocsis

• php
• benchmarks

Published on Packagist.org on 2026-08-01 09:48 UTC

README

Table of Contents

• Introduction
• Results
• Install
• Usage
• Contributing
• Support
• Credits
• License

Introduction

This is a fully automated benchmark suite for comparing the performance of any PHP releases or branches. The framework is heavily inspired by Intel's work from quite a few years ago: https://01.org/node/3774.

Results

You can find the benchmark results here.

Install

You can simply download or clone this repository. You can also install it via Composer by running the command below:

$ composer require kocsismate/php-version-benchmarks:dev-main

Usage

Configuration

All the configuration of the benchmark is done via .ini files in the config directory. By default, a few templates are bundled with the benchmark (having .ini.dist extension in their names). Feel free to rename them by removing the .dist suffix in order to take advantage of them.

Infrastructure

Option	Description	Supported values
INFRA_ID	The ID of the infrastructure configuration to be used in the benchmark code	String
INFRA_NAME	The name of the infrastructure configuration which is displayed in the benchmark results	String
INFRA_INSTANCE_TYPE	The AWS EC2 instance type	A valid AWS EC2 instance type (e.g. t2.micro)
INFRA_ARCHITECTURE	Any supported CPU architecture code	One of the following: "x86_64", "amd64", "arm64"
INFRA_DEDICATED_INSTANCE	Whether the dedicated instance feature is enabled	Boolean, either "0" or "1"
INFRA_DISABLE_DEEPER_C_STATES	Whether deeper CPU C-states are disabled	Boolean, either "0" or "1"
INFRA_DISABLE_TURBO_BOOST	Whether turbo boost is disabled	Boolean, either "0" or "1"
INFRA_DISABLE_HYPER_THREADING	Whether hyper threading is disabled	Boolean, either "0" or "1"
INFRA_LOCK_CPU_FREQUENCY	Whether CPU frequency is locked to the base frequency	Boolean, either "0" or "1"
INFRA_CPU_NUMA_NODE	Which CPU NUMA node to use for running the benchmark	Empty string for single NUMA systems, integer for multi-NUMA systems
INFRA_MAX_ALLOWED_CPU_TEMP	Maximum allowed CPU temperature in °C before running a test	Integer value (e.g. "50"). When set to "0", no temperature is enforced.
INFRA_WORKSPACE	A unique identifier to use when the same AWS account is reused across different benchmarking environments.	String
INFRA_RUNNER	Deprecated option. Only host runner is supported	One of the following: "host"
INFRA_COLLECT_EXTENDED_PERF_STATS	Whether to collect extended perf stat events (LLC-loads, LLC-load-misses, LLC-stores, LLC-store-misses, iTLB-load-misses, dTLB-load-misses)	Boolean, either "0" or "1"
INFRA_DEBUG_ENVIRONMENT	Whether to collect the environment state for debugging purposes	Boolean, either "0" or "1"

PHP

Option	Description	Supported values
PHP_ID	The ID of the PHP version to be used in the benchmark code	String
PHP_NAME	The name of the PHP version which is displayed in the benchmark results	String
PHP_BASE_ID	Refers to a PHP_ID	An existing PHP_ID value
PHP_REPO	The name of the git repository where the PHP version is available	An existing git URL containing the PHP source code
PHP_BRANCH	The branch within the PHP_REPO to be benchmarked	An existing branch name
PHP_COMMIT	The commit within the PHP_BRANCH to be benchmarked	An existing long commit SHA, or empty string to use the HEAD
PHP_JIT	Whether to enable tracing JIT	Boolean, either "0" or "1"

Test

Option	Description	Supported values
TEST_ID	The ID of the test to be used in the benchmark code	String
TEST_NAME	The name of the test which is displayed in the benchmark results	String
TEST_ITERATIONS	The number of times the tests are repeated	Integer
TEST_WARMUP	The number of warmup requests to perform within a test iteration	Integer
TEST_REQUESTS	The number of measured requests to perform within a test iteration	Integer
TEST_TYPE	The type of the benhcmark	One of "real", "micro"
TEST_FILE	The PHP file to execute during the benchmark	An existing PHP file
TEST_URL_PATH	The URL to use for running real benchmarks	An accessible URL
TEST_ENV	The APP_ENV environment variable to pass to real tests	String

Usage on AWS EC2

As a prerequisite, you need the following:

• a UNIX system (Linux, Mac)
• git
• Terraform

Then you have to create the necessary AWS-related config file by copying the aws.tfvars.dist to aws.tfvars in the build/infrastructure/config/aws.tfvars.dist directory:

cp build/infrastructure/config/aws.tfvars.dist build/infrastructure/config/aws.tfvars

Then, you need to override some values in it:

• access_key: the access key of your AWS account
• secret_key: the secret access key of your AWS account
• region: it is "eu-central-1" by default, but you should choose the closest one to your area
• state_bucket: The S3 bucket name where the state file is stored

Now, you are ready to go:

./benchmark.sh run aws

System-level Stability Measures

Getting consistent benchmark results requires careful system-level tuning. The following measures have been applied to minimize noise and ensure reproducibility:

• Disabling Hyper-Threading, Turbo Boost, and deep C-states prevents the CPU from dynamically changing its behavior during measurement. Setting the benchmarking CPU to its exact baseline frequency was particularly impactful for result stability.
• IRQ affinity pinning ensures hardware interrupts are handled only by OS-reserved cores, keeping the benchmarking core free from interrupt-related jitter. The CPU selection algorithm further prefers the core with the fewest IRQ affinities within the target NUMA node.
• Maximum CPU scheduling and I/O priority is assigned to the benchmark workload to prevent it from being preempted by other processes.
• cgroup-based CPU isolation replaced taskset to enable NUMA node-aware process placement, which is necessary on multi-NUMA instances.
• NVMe instance store is used instead of the default network-backed EBS disk, as network disk latency and variance was one of the most significant sources of measurement noise.
• Kernel parameters isolcpus, nohz_full, and rcu_nocbs are set to fully isolate the benchmarking cores from the OS scheduler and kernel housekeeping tasks.
• L3 cache isolation is also supported by the benchmark framework, but it is currently not supported on AWS.
• MySQL stability tuning was applied for WordPress benchmarks, which require a running database, to reduce query latency variance.
• CPU temperature gating prevents a benchmark run from starting if the CPU temperature exceeds a defined threshold, eliminating thermal throttling as a source of variance.

Contributing

Please see CONTRIBUTING for details.

Support

Please see SUPPORT for details.

Credits

• Máté Kocsis
• All Contributors

License

The MIT License (MIT). Please see the License File for more information.