Subject: Re: Do some disk accesses miss the UVM?
To: Perry E. Metzger <perry@wasabisystems.com>
From: Chuck Silvers <chuq@chuq.com>
List: tech-kern
Date: 01/26/2002 00:31:00
On Fri, Jan 25, 2002 at 12:18:34PM -0500, Perry E. Metzger wrote:
> Jason R Thorpe <thorpej@wasabisystems.com> writes:
> > So, while I agree this would be nice, stop to consider that if you
> > also plop file metadata into the page cache:
> > 
> > 	(1) You now have 3 policy knobs to futz with when picking
> > 	    a page to evict: anon, file data, filesystem metadata.
> 
> Yup. Ultimately, we need to get rid of the knobs -- see my
> "generational page replacement" notion or something similar.

actually the generational idea isn't a replacement for the knobs.
the generational bit still depends on the assumption that past behaviour
is a good predictor of future behaviour, so it falls down just like the
pre-knob code when that assumption isn't true.  the example that shows
this is the application that michael graff reported having problems
a while back.  in his case, the application had one usage pattern for
a long time, and then it switched to a very different usage pattern and
the performance became terrible.  a generational algorithm wouldn't be able
to handle that any better than what we would have right now if the usage-type
balancing code were removed.

so as far as I can tell, we can never remove the usage-type balancing code,
and as long as we have that we need to have the knobs.  so far the default
settings seem to be good, no one has complained about them yet.

-Chuck