Re: [RFC] [-mm PATCH] Memory controller fix swap charging context in unuse_pte()

[Balbir Singh <balbir-23VcF4HTsmIX0ybBhKVfKdBPR1lH4CV8@public.gmane.org>]

Hugh Dickins wrote:
> On Tue, 30 Oct 2007, Balbir Singh wrote:
>> At this momemnt, I suspect one of two things
>>
>> 1. Our mods to swap_state.c are different
> 
> I believe they're the same (just take swap_state.c back to how it
> was without mem_cgroup mods) - or would be, if after finding this
> effect I hadn't added a "swap_in_cg" switch to move between the
> two behaviours to study it better (though I do need to remember
> to swapoff and swapon between the two: sometimes I do forget).
> 
>> 2. Our configuration is different, main-memory to swap-size ratio
> 
> I doubt the swapsize is relevant: just so long as there's some (a
> little more than 200M I guess); I've got 1GB-2GB on different boxes.
> 

I agree, just wanted to make sure that there is enough swap

> There may well be something about our configs that's significantly
> different.  I'd failed to mention SMP (4 cpu), and that I happen
> to have /proc/sys/vm/swappiness 100; but find it happens on UP
> also, and when I go back to default swappiness 60.
> 

OK.. so those are out of the equation

> I've reordered your mail for more dramatic effect...
>> On a real box - a powerpc machine that I have access to
> 
> I've tried on 3 Intel and 1 PowerPC now: the Intels show the OOMs
> and the PowerPC does not.  I rather doubt it's an Intel versus
> PowerPC issue as such, but interesting that we see the same.
> 

Very surprising, I am surprised that it's architecture dependent.
Let me try and grab an Intel box and try.

>> 1. I don't see the OOM with the mods removed (I have swap
>>    space at-least twice of RAM - with mem=512M, I have at-least
>>    1G of swap).
> 
> mem=512M with 1G of swap, yes, I'm the same.
> 
>> 2. Running under the container is much much faster than running
>>    swapout in the root container. The machine is almost unusable
>>    if swapout is run under the root container
> 
> That's rather interesting, isn't it?  Probably irrelevant to the
> OOM issue we're investigating, but worthy of investigation in itself.
> 

Yes, it irrelevant, but I find it to be a good use case for using the
memory controller :-) I found that kswapd running at prio -5, seemed
to hog quite a bit of the CPU. But it needs more independent
investigation, like you've suggested.

> Maybe I saw the same on the PowerPC: I simply forgot to set up the
> cgroup one time, and my sequence of three swapouts (sometimes only
> two out of three OOM, on those boxes that do OOM) seemed to take a
> very long time (but I wasn't trying to do anything else on it at
> the same time, so didn't notice if it was "unusable").
> 
> I'll probe on.
> 

Me too.. I'll try and acquire a good x86_64 box and test on it.

> Hugh
> 
> --
> To unsubscribe, send a message with 'unsubscribe linux-mm' in
> the body to majordomo-Bw31MaZKKs0EbZ0PF+XxCw@public.gmane.org  For more info on Linux MM,
> see: http://www.linux-mm.org/ .
> Don't email: <a href=mailto:"dont-Bw31MaZKKs3YtjvyW6yDsg@public.gmane.org"> email-Bw31MaZKKs3YtjvyW6yDsg@public.gmane.org </a>


-- 
	Warm Regards,
	Balbir Singh
	Linux Technology Center
	IBM, ISTL