Re: [patch -mm 1/2] oom: badness heuristic rewrite

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

On Mon, 2 Aug 2010 18:50:11 -0700 (PDT)
David Rientjes <rientjes@google.com> wrote:

> On Tue, 3 Aug 2010, KAMEZAWA Hiroyuki wrote:
> 
> > Hmm, then, oom_score shows the values for all limitations in array ?
> > 
> 
> /proc/pid/oom_score will change if a task's cpuset, memcg, or mempolicy 
> attachment changes or its mems, nodes, or limit changes because it's a 
> proportion of available memory.  /proc/pid/oom_score_adj stays constant 
> such that the oom killing priority of that task relative to other tasks 
> sharing the same constraints and competing for the same memory is the 
> same.  The point is that it doesn't matter how much memory a task has 
> available, but rather what it's priority is with the tasks that compete 
> with it for memory.
> 
> You could, of course, do some simple arithmetic to write a memory quantity 
> to oom_score_adj if you really wanted to, but that would force the user to 
> recalculate the value anytime the task's cpuset, mempolicy, or memcg 
> changes.
> 

My concern is this may corrupt LXC and google's oom_adj system,
expecially when cpuset is used.


> > > > Usual disto alreay enables it.
> > > > 
> > > 
> > > Yes, I'm well aware of my 40MB of lost memory on my laptop :)
> > > 
> > Very sorry ;)
> > But it's required to track memory usage from init...
> > 
> 
> Memcg comes with a cost of ~1% of system memory on x86 since
> struct page_cgroup is ~1% of a 4K page.  That means if we were to deploy 
> memcg on all of our servers and the number of jobs we can run is 
> constrained only by memory, it's equivalent to losing ~1% of our servers.  
> That, for us, is very large.
> 
Google can disable it ;)


> This is a different topic entirely, but it's a very significant 
> disadvantage and enough that most people who care about oom killing 
> prioritization aren't going to wany to incur such an overhead by enabling 
> memcg or setting up individual memcg for each and every job, because that 
> requires specific knowledge of all those jobs.
> 
> I'm not by any means proposing oom_score_adj as being very popular for the 
> usual desktop environments :)
> 
libcgroup will help.

Bye.
-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>