Nginx – php-fpm invokes oom-killer

Personally I would never use PAE (over 4G of RAM on a 32-bit system). You'll get much better mileage running an actual 64-bit kernel and system.

OOM should only trigger when a malloc might fail. (not when you have lots of swap available)

The 32-bit kernel is likely to be part of the cause. PAE uses different memory zones and it may be that one zone isn't allowed to malloc from another.

Have you modified your swappiness? (How readily the kernel will use swap.) cat /proc/sys/vm/swappiness ?

You might also investigate tuning vm.dirty_ratio or vm.lower_zone_protection = 100.

Have you captured the kernel panic? (a serial console is often a good way to do this)

You can also try to preempt the OOM-Killer with your own process monitoring software. (take a look at Monit)

Best of luck

I just looked at the oom log dump, and I question the accuracy of that graph. Notice the first 'Node 0 DMA32' line. It says free:3376kB, min:3448kB, and low:4308kB. Whenever the free value drops below the low value, kswapd is supposed to start swapping things until that value gets back up above the high value. Whenever free drops below min, the system basically freezes until the kernel gets it back up above the min value. That message also indicates that swap was completely used where it says Free swap = 0kB.
So basically kswapd triggered, but swap was full, so it couldn't do anything, and the pages_free value was still below the pages_min value, so the only option was to start killing things until it could get pages_free back up.
You definitely ran out of memory.

http://web.archive.org/web/20080419012851/http://people.redhat.com/dduval/kernel/min_free_kbytes.html has a really good explanation of how that works. See the 'Implementation' section at the bottom.