Re: Improving OOM killer

[Balbir Singh]

* Lubos Lunak <l.lunak@xxxxxxx> [2010-02-03 13:10:27]:

> On Wednesday 03 of February 2010, Balbir Singh wrote:
> > * Lubos Lunak <l.lunak@xxxxxxx> [2010-02-01 23:02:37]:
> > >  In other words, use VmRSS for measuring memory usage instead of VmSize,
> > > and remove child accumulating.
> >
> > I am not sure of the impact of changing to RSS, although I've
> > personally believed that RSS based accounting is where we should go,
> > but we need to consider the following
> >
> > 1. Total VM provides data about potentially swapped pages,
> 
>  Yes, I've already updated my proposal in another mail to switch from VmSize 
> to VmRSS+InSwap. I don't know how to find out the second item in code, but at 
> this point of discussion that's just details.
> 

I am yet to catch up with the rest of the thread. Thanks for heads up.

> > overcommit, 
> 
>  I don't understand how this matters. Overcommit is memory for which address 
> space has been allocated but not actual memory, right? Then that's exactly 
> what I'm claiming is wrong and am trying to reverse. Currently OOM killer 
> takes this into account because it uses VmSize, but IMO it shouldn't - if a 
> process does malloc(400M) but then it uses only a tiny fraction of that, in 
> the case of memory shortage killing that process does not solve anything in 
> practice.

We have a way of tracking commmitted address space, which is more
sensible than just allocating memory and is used for tracking
overcommit. I was suggesting that, that might be a better approach.

> 
> > etc.
> > 2. RSS alone is not sufficient, RSS does not account for shared pages,
> > so we ideally need something like PSS.
> 
>  Just to make sure I understand what you mean with "RSS does not account for 
> shared pages" - you say that if a page is shared by 4 processes, then when 
> calculating badness for them, only 1/4 of the page should be counted for 
> each? Yes, I suppose so, that makes sense.

Yes, that is what I am speaking of

> That's more like fine-tunning at 
> this point though, as long as there's no agreement that moving away from 
> VmSize is an improvement.
>

There is no easy way to calculate the Pss today without walking the
page tables, but some simplification there will make it a better and a
more accurate metric.
 
> > I suspect the correct answer would depend on our answers to 1 and 2
> > and a lot of testing with any changes made.
> 
>  Testing - are there actually any tests for it, or do people just test random 
> scenarios when they do changes? Also, I'm curious, what areas is the OOM 
> killer actually generally known to work well in? I somehow get the feeling 
> from the discussion here that people just tweak oom_adj until it works for 
> them.
>

I've mostly found OOM killer to work well for me, but looking at the
design and our discussions I know there need to be certain improvements. 

-- 
	Balbir
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/