On Sat 24 Oct 2009, Rolf Schaufelberger wrote: > Starting some weeks ago the server sometimes hangs with an out of > memory problem.
Assuming you are running on linux the following sysctls may help to find the culprit. vm.overcommit_memory=2 vm.overcommit_ratio=90 By default (overcommit_memory=0) when a process needs more memory linux simply says "okay here you are" no matter if the memory is currently available or not. This is based on the assumption that most processes do not use all memory they allocate at all. Later on, when the process really accesses the memory a page fault is generated and linux only then allocates the memory for the process. But now it is too late to signal the process out-of-memory. So, linux somehow has to obtain memory no matter what. So, in short-of-memory conditions linux starts the OOM killer. It implements some heuristic that chooses some "best fitting" processes to be killed. These "best fitting" processes may be totally unrelated to the original problem. I had once a case where a perl program processed mailbox files (using Mail::Box) on a box where a postgres database ran. Unfortunately Mail::Box reads in the whole mailbox file. Normally our mailbox files were about 1-10mb and the program had worked for years. But suddenly we had one of >1Gb. Instead of signaling out-of-memory to the perl process linux killed the postgres database. Make sure you have enough swap space (at least the RAM size) before experimenting with those sysctls. Torsten -- Need professional mod_perl support? Just hire me: torsten.foert...@gmx.net
