Subject: Re: [vserver] oom-killer killapps on hosts not not guest
From: Herbert Poetzl <herbert@13thfloor.at>
Date: Thu, 20 Aug 2009 17:16:11 +0200

On Thu, Aug 20, 2009 at 03:55:29PM +1200, Michael wrote:
> > in any case, check with the oom_score ...
> On host scrore 0-200000
> 
> on guest with badness=1000000
> 50015
> 251769
> 1000638
> 
> and oom_adj on both=0
> 
> Why on guest it is lower then 1000000? ( ignore badness?)

no, but after the three steps I listed in the previous
replies, some 'global' adjustment is done, i.e.
the badness is weighted up/down in certain cases ...

maybe we should simply move the vx_badness to the end
of the calculations so that it will always have the
minimal bias value ...

will wrap up a patch for that to test

thanks,
Herbert

> > okay, so the first thing would be to figure out what
> > makes your host processes 'win', then we can think
> > about mechanisms to adjust the likelyness
> 
> with badness over 1000000
> it kills the guest process:
> g 20 15:40:34 deb64lenny kernel: [165160.869590] Out of memory: kill
> process 32337:#1003 (apache2) score 509684 or a child
> Aug 20 15:40:34 deb64lenny kernel: [165160.869611] Killed process
> 32337:#1003 (apache2)
> Aug 20 15:40:34 deb64lenny kernel: [165161.483101] Out of memory: kill
> process 32339:#1003 (qmgr) score 505755 or a child
> Aug 20 15:40:34 deb64lenny kernel: [165161.483126] Killed process
> 32339:#1003 (qmgr)
> Aug 20 15:40:34 deb64lenny kernel: [165161.490567] Out of memory: kill
> process 830:#1003 (stress) score 271805 or a child
> 
> but as you can see the scrore still below 1000000
> and offending process ( stress ) geot fro some reason lower score than
> apache or mysql, which got killed first.
> 
> 
> -- 
> --
> Michael