Re: [patch 0/8] mm: thrash detection-based file cache sizing v5

[Johannes Weiner <hannes@cmpxchg.org>]

On Mon, Oct 21, 2013 at 11:26:43AM +0200, Vlastimil Babka wrote:
> On 10/10/2013 11:46 PM, Johannes Weiner wrote:
> > Hi everyone,
> > 
> > here is an update to the cache sizing patches for 3.13.
> > 
> > 	Changes in this revision
> > 
> > o Drop frequency synchronization between refaulted and demoted pages
> >   and just straight up activate refaulting pages whose access
> >   frequency indicates they could stay in memory.  This was suggested
> >   by Rik van Riel a looong time ago but misinterpretation of test
> >   results during early stages of development took me a while to
> >   overcome.  It's still the same overall concept, but a little simpler
> >   and with even faster cache adaptation.  Yay!
> 
> Oh, I liked the previous approach with direct competition between the
> refaulted and demoted page :) Doesn't the new approach favor the
> refaulted page too much? No wonder it leads to faster cache adaptation,
> but could it also cause degradations for workloads that don't benefit
> from it? Were there any tests for performance regressions on workloads
> that were not the target of the patchset?

If anything, it's unfair to refaulting pages because it requires 3
references before they are activated instead of the regular 2.

We don't do the direct competition for regular in-core activation,
either, which has the same theoretical problem but was never an issue
in the real world.  Not that I know of anyway.

I ran a standard battery of mmtests (kernbench, dbench, postmark,
micro, fsmark, what have you) and did not notice any regressions.