Sam Trenholme's webpage
Support this website or listen to my music

Tuning nginx + php-fastcgi

 

July 2 2011

This article talks about the dark art of making a webpage with a lot of visitors use a minimum of server resources.

Nginx is a web server that is much more efficient than Apache serving static pages.

However, one blogger has pointed out that once we add dynamic PHP content, nginx + php-fastcgi is about as fast and uses as much memory as Apache.

That is only true with the default settings for php-fastcgi.

In truth, nginx can be much more lightweight than the equivalent Apache setup. There are two important variables when tuning php-fastcgi: PHP_FCGI_CHILDREN and PHP_FCGI_MAX_REQUESTS

PHP_FCGI_CHILDREN

PHP_FCGI_CHILDREN determines the number of processes php-fastcgi spawns. This affects the number of PHP pages that can be processed at the same time. If this has a low value like 4, then the server will use less memory but will slow down if the server gets a lot of PHP requests.

To use a rough analogy, PHP_FCGI_CHILDREN is akin to the number of tellers in a bank. If there is no one using the bank, then a customer will immediately get service even if there is only a single teller. On the other hand, if there are a lot of people in the bank, having more tellers lowers the amount of time one has to wait before getting service at the bank. Just as it costs the bank more money to have more tellers, it costs our server more memory to have more PHP_FCGI_CHILDREN.

Here's another way of looking at it: If a given PHP script takes 100 milliseconds (1/10 of a second) to process a request, we can process 10 PHP queries a second if PHP_FCGI_CHILDREN has a value of 1, 40 PHP queries a second if PHP_FCGI_CHILDREN has a value of 4, and 320 PHP queries a second if PHP_FCGI_CHILDREN has a value of 32.

Note that this only affects dynamic PHP performance. Static HTML pages will always be served quickly regardless of one's PHP_FCGI_CHILDREN value.

PHP_FCGI_MAX_REQUESTS

PHP_FCGI_MAX_REQUESTS determines the number of times a given php-fastcgi process will process PHP requests before terminating and being replaced by another php-fastcgi process.

To go back to our bank teller analogy, PHP_FCGI_MAX_REQUESTS is how long a teller's shift is before being replaced by another teller. Just like a bank, it is less expensive to give a single teller a shorter shift (we don't need to pay them overtime); just like a bank, when we switch tellers, service at the bank slows down because it takes time to swap cash drawers and what not.

Likewise, a given PHP process' memory usage increases over time; sometimes a PHP script will leak memory. By setting PHP_FCGI_MAX_REQUESTS to a fairly low value, we lower the memory usage of php-fastcgi, at the expense of there being more slight delays when a php-fastcgi process is terminated then respawned.

Unlike our hypothetical bank teller, PHP_FCGI_MAX_REQUESTS switches out processes after the specified number of requests are made. If PHP_FCGI_MAX_REQUESTS has a value of 128, for example, after a single process processes 128 PHP page loads, the process is terminated and replaced by another php-fastcgi process.

Note that, on Windows systems, there is a bug where a PHP process may not properly respawn after serving PHP_FCGI_MAX_REQUESTS requests. This is a non-issue on low-end Linux OpenVZ servers.

Concurrency

There is one "feature" to watch out for in a nginx + fastcgi setup. A given PHP script will only run once for a given IP. So, if the same IP loads index.php for a website twice at the same time, index.php will be served sequentially: The second index.php load will not be processed until the first index.php load is complete.

This does not happen if two different IPs ask for index.php at the same time (unless we have more PHP requests than PHP_FCGI_CHILDREN), nor does it happen if the same IP asks for two different PHP pages at the same time.

This may be an issue in some NAT setups.

Conclusion

On this website, I have set PHP_FCGI_CHILDREN to have a value of 4 and PHP_FCGI_MAX_REQUESTS to have a value of 128. Since this site is mostly a static HTML site, this does not significantly impact performance, while greatly reducing the amount of memory used to serve web pages.

While low-end OpenVZ virtual servers are inexpensive, they require one to understand how to tune a web server to get good performance while minimizing memory usage. I hope this article has given some helpful pointers.

To post a comment about an entry, send me an email and I may or may not post your comment (with or without editing)

Previous entry Next entry Blog index