Re: oom killer rewrite

[KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>]

On Tue, 25 May 2010 02:42:14 -0700 (PDT)
David Rientjes <rientjes@google.com> wrote:

> On Thu, 20 May 2010, KAMEZAWA Hiroyuki wrote:
> 
> > I've pointed out that "normalized" parameter doesn't seem to work well in some
> > situaion (in cluster). I hope you'll have an extra interface as
> > 
> > 	echo 3G > /proc/<pid>/oom_indemification
> > 
> > to allow users have "absolute value" setting.
> > (If the admin know usual memory usage of an application, we can only
> >  add badness to extra memory usage.)
> > 
> > To be honest, I can't fully understand why we need _normalized_ parameter. Why
> > oom_adj _which is now used_ is not enough for setting "relative importance" ?
> > 
> 
> The only sane badness heuristic will be one that effectively compares all 
> eligible tasks for oom kill in a way that are relative to one another; I'm 
> concerned that a tunable that is based on a pure memory quantity requires 
> specific knowledge of the system (or memcg, cpuset, etc) capacity before 
> it is meaningful.  In other words, I opted to use a relative proportion so 
> that when tasks are constrained to cpusets or memcgs or mempolicies they 
> become part of a "virtualized system" where the proportion is then used in 
> calculation of the total amount of system RAM, memcg limit, cpuset mems 
> capacities, etc, without knowledge of what that value actually is.  So 
> "echo 3G" may be valid in your example when not constrained to any cgroup 
> or mempolicy but becomes invalid if I attach it to a cpuset with a single 
> node of 1G capacity.  When oom_score_adj, we can specify the proportion 
> "of the resources that the application has access to" in comparison to 
> other applications that share those resources to determine oom killing 
> priority.  I think that's a very powerful interface and your suggestion 
> could easily be implemented in userspace with a simple divide, thus we 
> don't need kernel support for it.
> 
I know admins will be able to write a script. But, my point is
"please don't force admins to write such a hacky scripts."

For example, an admin uses an application which always use 3G bytes adn it's
valid and sane use for the application. When he run it on a server with
4G system and 8G system, he has to change the value for oom_score_adj.


One good point of old oom_adj is that it's not influenced by environment.
Then, X-window applications set it's oom_adj to be fixed value. 
IIUC, they're hardcoded with fixed value, now. 

Even if my customer may use only OOM_DISABLE, I think using oom_score_adj
is too difficult for usual users.


Thanks,
-Kame






--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>