PHP hits 100% CPU and eats RAM at the same time Monday to Friday
- by Daniel Samuels
We run a learning platform for primary schools here in the UK and it's all been running extremely well. However at around 4PM Monday to Friday we see the same issue arise -- 1-2 PHP threads will spike to 100% CPU and gradually start eating up RAM until the server(s) fall over.
98%+ of our requests are HTTPS, these come into our Layer 7 load balancer which then decrypts the SSL data, adds the X-HTTP-Forwarded-For header and forwards the data onto an application server (we have 2 of those at the moment) on port 80. Our application servers have Varnish on port 80 which takes in the request from the load balancer and passes the request through to Nginx on port 81. Nginx then works out which 'vhost' it needs to use and passes any PHP processing through to PHP-CGI which is listening on a socket (managed through spawn-fcgi). There's an instance of Memcached running too, MySQL runs on a separate server / slave setup.
Throughout the day the load will typically go no higher than 0.8 on either of the application servers, however at around 4PM our problem arises. I've managed to run strace on a few of the actual threads when they cause the problem and I always see the same thing:
stat("/usr/share/zoneinfo/Europe/London", {st_mode=S_IFREG|0644,st_size=3661, ...}) = 0
stat("/usr/share/zoneinfo/Europe/London", {st_mode=S_IFREG|0644,st_size=3661, ...}) = 0
This is repeated infinitely and never stops until you SEGKILL the process or oomkiller kills it. There are no cron jobs scheduled to run at that time and I don't have any way of seeing exactly what Nginx request is associated with the PHP process which is running.
We are running PHP 5.3.14 which we upgraded to from 5.3.8 last week to rule out the older version being the problem. This issue has been going on a few months now and we have no idea what is causing it. We deploy our software very frequently, so it's difficult to track down a specific release which may have started the problem - especially as we do not know the date of the first occurrence of this issue. Varnish is version 3.0.1, Nginx is 1.0.6 (which I understand is about a year old now), our servers are running CentOS release 5.7 (Final) they have Intel i3 540s at 3.07Ghz and 8GB of RAM.
There's a discussion on the Debian mailing list about something very similar, you can find that here.
Has anyone seen anything like this in the past, does anyone have any ideas or suggestions? Are there a way of linking an Nginx request directly to a PHP thread? Is there a better way of seeing what the PHP process is doing? (I've seen GDB mentioned, though I'll have to recompile PHP)
Thanks!
