Re: + mm-balance_dirty_pages-reduce-calls-to-global_page_state-to-reduce-c ache-references.patch added to -mm tree

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

On Sun, 2009-08-23 at 15:27 +0800, Wu Fengguang wrote:
> On Sun, Aug 23, 2009 at 01:31:17PM +0800, Peter Zijlstra wrote:
> > On Sun, 2009-08-23 at 09:32 +0800, Wu Fengguang wrote:
> > > > > I'd propose to remove the above 'if' and liberate the following three 'if's.
> > > > 
> > > > That might work, but it looses the total dirty_thresh constraint. The
> > > > sum of per-bdi dirties _should_ not be larger than that, but I'm not
> > > > sure it won't ever be.
> > > > 
> > > > The clip code Richard removed ensured that, and I think I wrote that out
> > > > of more than sheer paranoia, but I'm not sure anymore :/
> > > 
> > > Oh I assumed that your per-bdi throttling is not too permissive to
> > > exceed the global dirty_thresh. In theory the per-bdi throttling
> > > should be able to quickly stop the growing of (nr_reclaimable +
> > > nr_writeback).  Once dirty_thresh is reached we already los
> > 
> > Right, so:
> > 
> >  bdi_thresh_n = dirty_thresh * p(n) + eps.
> > 
> > and
> > 
> >  \Sum_n p(n) = 1
> > 
> > So:
> > 
> >  \Sum_n bdi_thresh_n = dirty_thresh + n*eps
> > 
> > Which yields an O(n) error bound.
> > 
> > I'm just not sure how large the thing is in reality, and paranoia won
> > out.
> 
> A wild question: is it possible to make the equation:
> 
>         bdi_thresh_n = dirty_thresh * p(n) - eps.

Possibly, I'd have to go over the math to get a proper equation for eps,
then see what needs to be done to get avg(eps) <= 0.