On Wed, 11 Jun 2008 12:45:14 +0900 (JST)
yamamoto@valinux.co.jp (YAMAMOTO Takashi) wrote:

> > > having said that, if you decide to put too large tasks into
> > > a cgroup with too small limit, i don't think that there are
> > > many choices besides OOM-kill and allowing "exceed".
> > > 
> > IMHO, allowing exceed is harmfull without changing the definition of "limit".
> > "limit" is hard-limit, now, not soft-limit. Changing the defintion just for
> > this is not acceptable for me. 
> 
> even with the current code, the "exceed" condition can be created
> by simply lowering the limit.
> (well, i know that some of your patches floating around change it.)
> 
Yes, I write it now ;) Handling exceed contains some troubles

  - when resizing limit, to what extent exceed is allowed ?
  - Once exceed, no new page allocation can success and
    _some random process_ will die because of OOM.


> > Maybe "move" under limit itself is crazy ops....Hmm...
> > 
> > Should we allow task move when the destination cgroup is unlimited ?
> > Isn't it useful ?
> 
> i think it makes some sense.
> 
> > > actually, i think that #3 and #5 are somewhat similar.
> > > a big difference is that, while #5 shrinks the cgroup immediately,
> > > #3 does it later.  in case we need to do OOM-kill, i prefer to do it
> > > sooner than later.
> > > 
> > #3 will not cause OOM-killer, I hope...A user can notice memory shortage.
> 
> we are talking about the case where a cgroup's working set is getting
> hopelessly larger than its limit.  i don't see why #3 will not
> cause OOM-kill.  can you explain?
> 
just because #3 doesn't move resource, just drop. 

Thanks,
-Kame


_______________________________________________
Containers mailing list
Containers@lists.linux-foundation.org
https://lists.linux-foundation.org/mailman/listinfo/containers