Re: [PATCH 0/3] readahead drop behind and size adjustment

[Andrew Morton <akpm@linux-foundation.org>]

On Mon, 23 Jul 2007 22:24:57 +0800
Fengguang Wu <fengguang.wu@gmail.com> wrote:

> On Mon, Jul 23, 2007 at 07:00:59PM +1000, Nick Piggin wrote:
> > Rusty Russell wrote:
> > >On Sun, 2007-07-22 at 16:10 +0800, Fengguang Wu wrote:
> > 
> > >>So I opt for it being made tunable, safe, and turned off by default.
> > 
> > I hate tunables :) Unless we have workload A that gets a reasonable
> > benefit from something and workload B that gets a significant regression,
> > and no clear way to reconcile them...
> 
> Me too ;)
> 
> But sometimes we really want to avoid flushing the cache.
> Andrew's user space LD_PRELOAD+fadvise based tool fit nicely here.

It's the only way to go in some situations.  Sometimes the kernel just
cannot predict the future sufficiently well, and the costs of making a
mistake are terribly high.  We need human help.  And it should be
administration-time help, not programming-time help.

> > >I'd like to see it turned on by default in -mm, and try to come up with
> > >some server-like workload to measure the effect.  Should be easy to
> > >simulate something (eg. apache server, where clients grab some files in
> > >preference, and apache server where clients grab different files).
> > 
> > I don't like this kind of conditional information going from something
> > like readahead into page reclaim. Unless it is for readahead _specific_
> > data such as "I got these all wrong, so you can reclaim them" (which
> > this isn't).
> > 
> > Possibly it makes sense to realise that the given pages are cheaper
> > to read back in as they are apparently being read-ahead very nicely.
> 
> In fact I have talked to Jens about it in last year's kernel summit.
> The patch below explains itself.
> ---
> Subject: cost based page reclaim
> 
> Cost based page reclaim - a minimalist implementation.
> 
> Suppose we cached 32 small files each with 1 page, and one 32-page chunk from a
> large file.  Should we first drop the 32-pages which are read in one I/O, or
> drop the 32 distinct pages, each costs one I/O? (Given that the files are of
> equal hotness.)
> 
> Page replacement algorithms should be designed to minimize the number of I/Os,
> instead of the number of page faults. Dividing the cost of I/O by the number of
> pages it bring in, we get the cost of the page. The bigger page cost, the more
> 'lives/bloods' the page should have.
> 
> This patch adds life to costly pages by pretending that they are
> referenced more times. Possible downsides:
> - burdens the pressure of vmscan
> - active pages are no longer that 'active'
> 

This is all fun stuff, but how do we find out that changes like this are
good ones, apart from shipping it and seeing who gets hurt 12 months later?

> +#define log2(n) fls(n)

<look at include/linux/log2.h>