Re: [patch 0/5] refault distance-based file cache sizing

[nai.xia]

Hi,

On 2012/05/16 14:51, Johannes Weiner wrote:
> Hi Nai,
>
> On Wed, May 16, 2012 at 01:25:34PM +0800, nai.xia wrote:
> > Hi Johannes,
> >
> > Just out of curiosity(since I didn't study deep into the
> > reclaiming algorithms), I can recall from here that around 2005,
> > there was an(or some?) implementation of the "Clock-pro" algorithm
> > which also have the idea of "reuse distance", but it seems that algo
> > did not work well enough to get merged? Does this patch series finally
> > solve the problem(s) with "Clock-pro" or totally doesn't have to worry
> > about the similar problems?
>
> As far as I understood, clock-pro set out to solve more problems than
> my patch set and it failed to satisfy everybody.
>
> The main error case was that it could not partially cache data of a
> set that was bigger than memory.  Instead, looping over the file
> repeatedly always has to read every single page because the most
> recent page allocations would push out the pages needed in the nearest
> future.  I never promised to solve this problem in the first place.
> But giving more memory to the big looping load is not useful in our
> current situation, and at least my code protects smaller sets of
> active cache from these loops.  So it's not optimal, but it sucks only
> half as much :)

Yep, I see ;)

> There may have been improvements from clock-pro, but it's hard to get
> code merged that does not behave as expected in theory with nobody
> understanding what's going on.
>
> My code is fairly simple, works for the tests I've done and the
> behaviour observed so far is understood (at least by me).

OK, I assume that you do aware that the system you constructed with
this simple and understandable idea looks like a so called "feedback
system"? Or in other words, I think theoretically the refault-distance
of a page before and after your algorithm is applied is not the same.
And this changed refault-distance pattern is then feed as input into
your algorithm. A feedback system may be hard(and may be simple) to
analyze but may also work well magically.

Well, again I confess I've not done enough course in this area. Just hope
that my words can help you think more comprehensively. :)


Thanks,

Nai
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/