lkml.org 
[lkml]   [2017]   [Aug]   [25]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [v6 2/4] mm, oom: cgroup-aware OOM killer
On Fri 25-08-17 11:39:51, Roman Gushchin wrote:
> On Fri, Aug 25, 2017 at 10:14:03AM +0200, Michal Hocko wrote:
> > On Thu 24-08-17 15:58:01, Roman Gushchin wrote:
> > > On Thu, Aug 24, 2017 at 04:13:37PM +0200, Michal Hocko wrote:
> > > > On Thu 24-08-17 14:58:42, Roman Gushchin wrote:
> > [...]
> > > > > Both ways are not ideal, and sum of the processes is not ideal too.
> > > > > Especially, if you take oom_score_adj into account. Will you respect it?
> > > >
> > > > Yes, and I do not see any reason why we shouldn't.
> > >
> > > It makes things even more complicated.
> > > Right now task's oom_score can be in (~ -total_memory, ~ +2*total_memory) range,
> > > and it you're starting summing it, it can be multiplied by number of tasks...
> > > Weird.
> >
> > oom_score_adj is just a normalized bias so if tasks inside oom will use
> > it the whole memcg will get accumulated bias from all such tasks so it
> > is not completely off. I agree that the more tasks use the bias the more
> > biased the whole memcg will be. This might or might not be a problem.
> > As you are trying to reimplement the existing oom killer implementation
> > I do not think we cannot simply ignore API which people are used to.
> >
> > If this was a configurable oom policy then I could see how ignoring
> > oom_score_adj is acceptable because it would be an explicit opt-in.
> >
> > > It also will be different in case of system and memcg-wide OOM.
> >
> > Why, we do honor oom_score_adj for the memcg OOM now and in fact the
> > kernel memcg OOM killer shouldn't be very much different from the global
> > one except for the tasks scope.
>
> Assume, you have two tasks (2Gb and 1Gb) in a cgroup with limit 3Gb.
> The second task has oom_score_adj +100. Total memory is 64Gb, for example.
>
> I case of memcg-wide oom first task will be selected;
> in case of system-wide OOM - the second.
>
> Personally I don't like this, but it looks like we have to respect
> oom_score_adj set to -1000, I'll alter my patch.

I cannot say I would love how oom_score_adj works but it's been like
that for a long time and people do rely on that. So we cannot simply
change it under people feets.

> > > > > I've started actually with such approach, but then found it weird.
> > > > >
> > > > > > Besides that you have
> > > > > > to check each task for over-killing anyway. So I do not see any
> > > > > > performance merits here.
> > > > >
> > > > > It's an implementation detail, and we can hopefully get rid of it at some point.
> > > >
> > > > Well, we might do some estimations and ignore oom scopes but I that
> > > > sounds really complicated and error prone. Unless we have anything like
> > > > that then I would start from tasks and build up the necessary to make a
> > > > decision at the higher level.
> > >
> > > Seriously speaking, do you have an example, when summing per-process
> > > oom_score will work better?
> >
> > The primary reason I am pushing for this is to have the common iterator
> > code path (which we have since Vladimir has unified memcg and global oom
> > paths) and only parametrize the value calculation and victim selection.
>
> I agree, but I'm not sure that we can (and have to) totally unify the way,
> how oom_score is calculated for processes and cgroups.
>
> But I'd like to see an unified oom_priority approach. This will allow
> to define an OOM killing order in a clear way, and use size-based tiebreaking
> for items of the same priority. Root-cgroup processes will be compared with
> other memory consumers by oom_priority first and oom_score afterwards.

This again changes the existing semantic so I really thing we should be
careful and this all should be opt-in.
--
Michal Hocko
SUSE Labs

\
 
 \ /
  Last update: 2017-08-25 12:59    [W:0.112 / U:29.316 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site