Re: [PATCH][RFC] dirty balancing for cgroups

[Peter Zijlstra <a.p.zijlstra@chello.nl>]

On Fri, 2008-07-11 at 16:13 +0900, KAMEZAWA Hiroyuki wrote:
> On Fri, 11 Jul 2008 14:59:26 +0900 (JST)
> yamamoto@valinux.co.jp (YAMAMOTO Takashi) wrote:
> 
> > > > >  - This looks simple but, could you merge this into memory resource controller ?
> > > > 
> > > > why?
> > > > 
> > > 3 points.
> > >  1. Is this useful if used alone ?
> > 
> > it can be.  why not?
> > 
> > >  2. memcg requires this kind of feature, basically.
> > > 
> > >  3. I wonder I need more work to make this work well under memcg.
> > 
> > i'm not sure if i understand these points.  can you explain a bit?
> > 
> In my understanding, dirty_ratio is for helping memory (reclaim) subsystem.
> 
> See comments in fs/page-writeback.c:: determin_dirtyable_memory()
> ==
> /*
>  * Work out the current dirty-memory clamping and background writeout
>  * thresholds.
>  *
>  * The main aim here is to lower them aggressively if there is a lot of mapped
>  * memory around.  To avoid stressing page reclaim with lots of unreclaimable
>  * pages.  It is better to clamp down on writers than to start swapping, and
>  * performing lots of scanning.
>  *
>  * We only allow 1/2 of the currently-unmapped memory to be dirtied.
>  *
>  * We don't permit the clamping level to fall below 5% - that is getting rather
>  * excessive.
>  *
>  * We make sure that the background writeout level is below the adjusted
>  * clamping level.
> ==
> 
> "To avoid stressing page reclaim with lots of unreclaimable pages"
> 
> Then, I think memcg should support this for helping relcaim under memcg.

That comment is unclear at best.

The dirty page limit avoids deadlocks under certain situations, the per
BDI dirty limit avoids even mode deadlocks by providing isolation
between BDIs.


The fundamental deadlock solved by the dirty page limit is the typical
reclaim deadlock - needing memory to free memory. It does this by
ensuring only some part of the total memory used for the page-cache can
be dirty, thus we always have clean pages around that can be reclaimed
so we can launder the dirty pages.

This on its own generates a new deadlock for stacked devices, imagine
device A on top of B. When A generates loads of dirty pages it will
eventually hit the dirty limit and we'd start to launder them. However
in order to launder A's dirty pages we'd need to dirty pages for B, but
we can't since we're at the global limit.

This problem is solved by introducing a per BDI dirty limit, by
assigning each BDI an individual dirty limit (whoes sum is the total
dirty limit) we avoid that deadlock. Take the previous example; A would
start laundering its pages when it hits its own limit, B's operation
isn't hampered by that.

[ even when B's limit is 0 we're able to make progress, since we'll only
  wait for B's dirty page count to decrease - effectively reducing to
  sync writes. ]

Of course this raises the question how to assign the various dirty
limits - any fixed distribution is hard to maintain and suboptimial for
most workloads.

We solve this by assigning each BDI a fraction proportional to its
current launder speed. That is to say, if A launders pages twice as fast
as B does, then A will get 2/3-rd of the total dirty page limit, versus
1/3-rd for B.


Then there is the task dirty stuff - this is basically a 'fix' for the
problem where a slow writer gets starved by a fast reader. Imagine two
tasks competing for bandwidth, 1 the fast writer and 2 the slow writer.

1 will dirty loads of pages but all things being equal 2 will have to
wait for 1's dirty pages.

So what we do is lower the dirty limit for fast writers - so these get
to wait sooner and slow writers have a little room to make progress
before they too have to wait.

To properly solve this we'd need to track p_{bdi,task}. However that's
intracktable. Therefore we approximate that with p_bdi * p_task. This
approximation looses detail.

Imagine two tasks: 1 and 2, and two BDIs A and B (independent this
time). If 1 is a (fast) writer to A and 2 is a (slow) writer to B, we
need not throttle 2 sooner as there is no actual competition.

The full proportional tensor p_{bdi,task} can express this, but the
simple approximation p_bdi * p_task can not.

The approximation will reduce 1's bandwidth a little even though there
is no actual competition.


Now the problem this patch tries to address...

As you can see you'd need p_{bdi,cgroup,task} for it to work, and the
obvious approximation p_bdi * p_cgroup * p_task will get even more
coarse.

You could possibly attempt to do p_{bdi,cgroup} * p_task since the bdi
and cgroup set are pretty static, but still that would be painful.

So, could you please give some more justification for this work, I'm not
seeing the value in complicating all this just yet.


Thanks for reading this far,

Peter

--
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>