This article sums up what to expect in terms of performances and how to optimize serverless PHP applications. The benchmarks included in this page can be reproduced via the code on GitHub.
CPU power and memory size
The main factor affecting performances is memory size. Indeed, the amount of memory is proportional to the CPU power available.
In other words, more memory means a more powerful CPU. A 1024M lambda has a CPU two times more powerful than a 512M lambda.
From 64M to 1,792M, applications run with up to one CPU (1,792M gives 1 full CPU). From 1,856M to 3,008M, applications run with 2 CPU (3,008M gives 2 full CPU). Since PHP is single-threaded and one lambda handles only 1 request at a time, using 2 CPU usually does not provide any benefit.
It is recommended to use 1024M for PHP applications, or at least to start with that. This is what Serverless deploys by default, so there is nothing to do.
To customize the amount of memory, set the
memory option in
functions: foo: handler: index.php # ... memory: 512 # set to 512M instead of 1024M (the default)
In the benchmark below, we run PHP's official
bench.php script. This script is CPU-intensive.
bench.php runs in 1.3s on a 512M Digital Ocean server, in 0.8s on a 2.8Ghz i7 and in 0.6s on a 3.2Ghz i5. It is safe to say that a 1024M lambda provides a powerful CPU.
AWS Lambda bills the number of events + the execution time. The more memory configured for a lambda, the more expensive is the execution time.
It might be tempting to lower the memory to save money. However, a function might run slower on a smaller lambda, canceling the cost savings. For example, both of these scenarios cost the same thing:
- a function running in 400ms on a 512M lambda
- the same function running in 200ms (because of the faster CPU) on a 1024M lambda
In general, use smaller and slower lambdas only when speed is not important at all.
PHP runtime overhead
The Bref runtime for HTTP applications does not add overhead to response times.
Here are execution times for an empty PHP application:
Unless we use a particularly slow lambda (see the next section, 128M is not recommended), 1ms is the same execution time when PHP runs with Apache or Nginx on a classic server.
We can see the same result with a "Hello world" written in Symfony (4ms being the minimum execution time of the framework):
The Bref runtime for PHP functions (non-HTTP applications) adds a small overhead:
Since this runtime is often used in asynchronous scenarios (for example, processing queue messages), it is often negligible.
Code on AWS Lambda runs on-demand. When a new Lambda instance boots to handle a request, the initialization time is what we call a cold start. To learn more, you can read this article.
Bref's PHP runtimes have a cold start of 250ms on average.
On a website with low to medium traffic, you can expect cold starts to happen for about 0.5% of the requests.
Optimizing cold starts
On small websites, cold starts can be avoided by pinging the application regularly. This keeps the lambda instances warm. Pingdom or similar services can be used, but you can also an automatic ping via
While the memory size has no impact, the codebase size can increase the cold start duration. When deploying, remember to exclude assets, images, tests and any extra file in
package: exclude: - 'assets/**' - 'node_modules/**' - 'tests/**'
Read more about the
package configuration in the serverless.yml documentation.
A note on VPC cold starts
Running a function inside a VPC used to induce a cold start of several seconds. This is no longer the case since October 2019.