Re: 2.6.23.8: OOM killer kills wrong jobs

[Alan Cox <alan@lxorguk.ukuu.org.uk>]

> ...but I've run into a situation in which a system on which I *have* set
> no overcommit is being blasted by the OOM killer anyway.

Looks like the kernel is eating all the resources needed.

>    Linux babyalcor 2.6.23.1 #1 SMP Fri Oct 26 15:35:18 EDT 2007 \
>     i686 Dual Core AMD Opteron(tm) Processor 280 AuthenticAMD GNU/Linux

32bit kernel, 16GB of RAM. 

No suprise I'm afraid. Handling 16GB on a 32bit kernel, which has to
manage it all through a small addressible memory window is right on the
limit of what the standard kernel will handle (8GB is probably as high as
I would go). The no overcommit code ensures that user space doesn't
overcommit, but the kernel can get itself short of low memory resources
on a big box with 32bit kernels very easily. (In 64bit mode the CPU can
address all the memory directly so the problem vanishes).

You will *probably* get stable 16GB with the vendor tuned enterprise
kernels (RHEL, CentOS etc), or run a 64bit kernel and then the kernel
isn't trying the software equivalent of managing a filing cabinet through
the keyhole.

Alan