This is an example configuration php-fpm
which is the preferred way of serving PHP with FastCGI. The old
php-cgi
is deprecated. You can even use a
patch for PHP versions prior to 5.3.3, when fpm
become an official PHP project.
The way that php-fpm operates is through a master process that launches children processes (workers) that serve the requests. These workers can be run under a particular user. Making it appropriate for environments where there are several users each one running their own set of PHP apps.
php-fpm is pretty much a work in progress, although on one hand is IMHO the best thing to come out of PHP land in a lot of time, still lacks a lot of nice features like graceful restarts. Currently adding a new pool or reloading the config requires a full restart. Which causes some downtime. Work is in progress to provide graceful restarts/reloads like Nginx has.
It has the capacity to adjust the number of workers dynamically to the load, varying from a minimum to a specified maximum.
The configuration comes in two flavors:
-
unix which is the default. It uses UNIX domain sockets for communication between the FCGI responder provided by php-fpm and the server or request frontend.
-
tcp. It uses TCP sockets for communication between the FCGI responder provided by php-fpm and the server or request frontend.
Choose the one that is more appropriate for your setup. Up until PHP 5.3.8 TCP sockets, although, theoretically slower, if your setup is on the loopback, behave better than UNIX sockets for high-traffic sites.
This might have changed in 5.3.10 and 5.4.x. Try it and report back please.
There's no algorithm for determining the number of children. It depends on your application.
A thread in the highload-php-en gives some tips on how to determine the number of children.
-
If your load is CPU bound then the rule is that the number of children should be equal to number of CPUs plus 20 %.
Example: Machine with 8 CPUs. Number of children = 10.
-
If your load is I/O bound then apply the following rule:
number_of_children = 1.2 * total_memory / average_space_per_process
Example: PHP processes occupying 256 MB of average space in a machine with 2GB of RAM that can be used for running this PHP application.
number_of_children = 1.2 * 2048 / 256
giving:
10
children.Determine the medium space occupied by a PHP process and apply the above formula. The 1.2 factor is just a security factor, to use a much abused engineering term.
-
It uses
UNIX
sockets for connections from the web server to the FastCGI daemon as the default. It provides a branchtcp
that relies on TCP sockets. Choose which suits you better. -
The
php.ini
is modified from the stock one that comes with the Debian package. The modifications were made by using a tiny script that I wrote for cleaning up a PHP config and that's available here on github. -
There are three pools on this config that run under the
www-data
user. The idea is that the server should load balance and distribute the load as you wish over it. -
Support for the status and ping functionalities of
php-fpm
. See here how to enable it for Nginx.
-
Clone the git repo:
git://github.com/perusio/php-fpm-example-config
. -
Checkout the
tcp
branch if that's suits you better:git checkout -b tcp
If on the contrary the
unix
sockets approach is more suited to your site(s) then you can ignore this step and proceed to 3 directly. -
Alter the
php-fpm.conf
and thepool.d/www.conf
file to your liking. Add any pool that you might want. -
Copy the files to the destination directory:
cp php5-fpm.conf /etc/php5/fpm
cp -a pool.d /etc/php5/fpm
-
(Re)start
php5-fpm
withservice php5-fpm restart
orservice php5-fpm start
if startingphp-fpm
anew.
Remember to always do service php5-fpm restart
after adding a
new pool or modifying the configuration of an existing one.