MaxClients in Apache with Preforking MPM

The advice for figuring out what your MaxClients setting should be in Apache with the Preforking MPM is pretty straight forward. First figure out the average amount of RAM each Apache process uses. Then divide the amount of RAM your system has, minus a little for other processes, by the average amount of RAM the average Apache process uses. In CentOS the following command should give you that number.

ps -ylC httpd |awk "{ s += \$8 } END { m = `cat /proc/meminfo|grep MemTotal|awk '{print $2}'`; printf \"MaxClients: %i\\n\", (m-(m*.15))/(s/NR) }"

Dealing with Memory Usage in Apache

Whatever amount of memory an Apache process uses to full fill a request, that process will continue to hold onto that amount of memory for the lifetime of that process. When you graceful Apache you cause all your current processes to end as soon as they are finished serving the requests they have already received, and then Apache creates a new set of processes. This typically frees up large blocks of memory used to serve old requests. Apache does this on it's own and it's controlled by the MaxRequestsPerChild directive. This is the number of requests a process should serve during it's lifetime. Making this number too low will cause Apache to tear down and create new processes quickly, increasing CPU overhead. Setting MaxRequestsPerChild too high will cause your machine to use more memory than it likely needs to, possibly causing your system to swap, or even run out of memory completely. Apache has no other means of garbage collection.

The default config for Apache on CentOS 5.3 has MaxRequestsPerChild set at 4000. I've seen people suggest that if your application is primarily serving dynamic content that a setting as low at 20 may make sense. I've opt'ed for 800 to 1000. It's a number I'm still playing with. Setting it this low has reduced the overall amount of memory my systems are using without increasing my CPU load significantly. It's something you'll need to experiement with.