PHP-8.5.0: performance discrepancy in container cold-start

[pentium10]

Description

I am currently evaluating the performance of a stock Laravel application within a containerized environment and have observed a noticeable performance difference between PHP 8.4 and PHP 8.5.

Observation

In a direct comparison of cold startup times, the container running on php:8.4-fpm is consistently faster than the one on php:8.5-fpm. My initial benchmarks show the PHP 8.4 container starting up around 300ms faster than the PHP 8.5 container. This is contrary to my expectation that a newer version would yield similar or improved performance.

Request for Input

I would like to understand what might be contributing to this performance gap. Could you provide any insights on the following:

• Default Configuration Changes: Are there any significant changes to the default php.ini or PHP-FPM configuration values between versions 8.4 and 8.5 that could impact startup performance?
• Container Image Differences: Are there any known differences in the official php-fpm Docker images for these versions that might explain the slower startup time for PHP 8.5?
• Known Performance Regressions: Are there any acknowledged performance regressions in PHP 8.5, particularly concerning startup, that the community is aware of? While benchmarks often show similar or slightly improved performance for newer PHP versions, my findings suggest a potential issue in this specific context.

My goal is to ensure we are using the most performant and up-to-date version of PHP. Any guidance or suggestions on where to investigate further would be greatly appreciated.

Thank you for your time and assistance.

Dockerfiles are different just the 8.4 vs 8.5 version

#FROM php:8.5-fpm
FROM php:8.5-fpm AS builder

# Get latest Composer
COPY --from=composer:latest /usr/bin/composer /usr/bin/composer

# Install system dependencies required by Composer
RUN apt-get update && apt-get install -y git zip unzip libzip-dev && docker-php-ext-install zip

# Set working directory
WORKDIR /var/www/html

# Copy rest of the application source code
COPY . .

# Install Composer dependencies
RUN composer install --no-dev --optimize-autoloader

# ---- Runtime Stage ----
FROM php:8.5-fpm

# Set working directory
WORKDIR /var/www/html

# Copy application files from the builder stage
COPY --from=builder /var/www/html .

# Expose port 8080 and start php-fpm server
EXPOSE 8080
CMD php artisan serve --host=0.0.0.0 --port=8080

The test was made with exact same codebase, exact same requests, in exact same time, region etc.. 100% cold starts, measured in Google Cloud Run platform, using the same official container slow start latency metrics. Results are consistent.

PHP Version

PHP 8.5.0

Operating System

No response

[iamacarpet]

It is potentially unrelated, but FYI, doing this:

“php artisan serve --host=0.0.0.0 --port=8080”

My understanding is that this is using PHP’s built in development HTTP server, i.e. “php -S”, and isn’t actually using “php-fpm”.

I’m not super sure how this would affect performance specifically, but I’m pretty sure it isn’t recommended for use in production.

You said elsewhere you are on Cloud Run: have you tried Google Cloud’s build packs? I believe PHP 8.5 is available in preview, and it should properly configure nginx/php-fpm for you on a source deploy.

[pentium10]

I went ahead and I changed the container to php:8.5-apache to have another set of metrics not just the php:8.5-fpm.

FROM php:8.5-apache

# 1. Install development packages and clean up apt cache.
RUN apt-get update && apt-get install -y \
    git \
    zip \
    unzip \
    libzip-dev \
    && docker-php-ext-install zip \
    && a2enmod rewrite

# 2. Apache configuration
# Change the document root to the public directory
ENV APACHE_DOCUMENT_ROOT /var/www/html/public

RUN sed -ri -e 's!/var/www/html!${APACHE_DOCUMENT_ROOT}!g' /etc/apache2/sites-available/*.conf
RUN sed -ri -e 's!/var/www/!${APACHE_DOCUMENT_ROOT}!g' /etc/apache2/apache2.conf /etc/apache2/conf-available/*.conf

# 3. Modifty Apache to listen on port 8080 (Cloud Run requirement)
RUN sed -i 's/80/8080/g' /etc/apache2/sites-available/000-default.conf /etc/apache2/ports.conf

# 4. Install Composer
COPY --from=composer:latest /usr/bin/composer /usr/bin/composer

# 5. Set working directory
WORKDIR /var/www/html

# 6. Copy application source
COPY . .

# 7. Install dependencies
RUN composer install --no-dev --optimize-autoloader

# 8. Set permissions
RUN chown -R www-data:www-data /var/www/html/storage /var/www/html/bootstrap/cache

# Expose port 8080
EXPOSE 8080

These show way better cold starts metrics, but that is because apache is serving the page, and thus its faster to establish the connection on the port itself, this way in a nutshell we are measuring which service gets faster to send first headers, but actualy php script processing is not yet done with it. For example I triggered a cold start from browser tab, and in the container latency I see 300ms, but in the browser I see 1000ms as a response, as measurements are having a difference.

I consider the raw php serve being the best here to compare JUST the php part, so my initial attempt which I used previously many times I consider the best aproach.

@iamacarpet I don't rely on Buildpacks because I have no control of what they add there, and I am not recommending at all using that for apps, even I do Google Cloud speaking. Its way easier to have a Dockerfile in your app and deal with your own scripts.

[iluuu1994]

My understanding is that this is using PHP’s built in development HTTP server, i.e. “php -S”, and isn’t actually using “php-fpm”.

Yes.

I’m not super sure how this would affect performance specifically, but I’m pretty sure it isn’t recommended for use in production.

And yes.

https://www.php.net/manual/en/features.commandline.webserver.php

Warning

This web server is designed to aid application development. It may also be useful for testing purposes or for application demonstrations that are run in controlled environments. It is not intended to be a full-featured web server. It should not be used on a public network.

The built-in webserver is neither built for performance nor is it secure. It is simply there to aid development. Please don't use it in production.

@pentium10 Given you don't have the same issue with the Apache image, is it ok to close this issue? Edit: Re-reading your answer: Can you clarify whether this discrepancy goes for any page, or is specific to some of them? And if it's the latter, can you investigate what exactly is causing the slowdown?

[pentium10]

We are still seeing slowness in coldstarts with PHP8.5, I put it to a few projects and I am actively trying to find some more concret stuff, and since I am struggling, and its more likely regression thats why I raised the ticket. I didn't abandoned metrics, just I wasn't posting on the ticket as it has so much in it already.

I acnowledge that php serve might not be production ready, but there the regression is reproducible for most.
While with the apache its two folds, I am mostly concerned we are measuring with apache the wrong cold start, that doesn't contain the PHP execution entirely, even with this gap measurements, it does show its worse than PHP8.4 (by a bit only but its been steadly slower like that for 10 days now).

the below chart is a cold start metric for each time mark

[pentium10]

To answer your question: I mirror the same request on these 2 containers, and its only one http hit every 30m, very basic, the laravel homepage / no database / only pure php. It does not have production traffic to it. And there are no noisy local services here, as these are cold starts pure for php. Ignore the "spikes" as those are cloud infra specific spikes, just focus on the lower section, and notice "green/php85" constantly slower than PHP8.4

[iluuu1994]

Ok, thanks for the explanation. It will be very hard to figure out why your application runs slower with PHP 8.5 without any source code. It could be due to cold-start, but it might also just be the application itself. Maybe a profiler might offer some insight? E.g. tideways or blackfire. There are also free alternatives.

Default Configuration Changes: Are there any significant changes to the default php.ini or PHP-FPM configuration values between versions 8.4 and 8.5 that could impact startup performance?

Opcache is always bundled since PHP 8.5. I don't know if the same is true for the docker image <PHP 8.5 (the Docker image for PHP is maintained by Docker). Opcache usually helps performance, but only for subsequent requests. So if it's now enabled and wasn't before, maybe the overhead comes from your application spending extra time setting up shared memory.

That's one guess, but you can also use phpinfo() or php -i to compare the settings in the two containers.

[pentium10]

php ini and opcache differences are posted here, notice the 2 set of comments for php8.4 vs php8.5: #16551 (comment)