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

[Nai Xia]

On 2012/05/18 05:08, Johannes Weiner wrote:
> On Wed, May 16, 2012 at 08:56:54PM +0800, nai.xia wrote:
> > On 2012/05/16 14:51, Johannes Weiner wrote:
> > > 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.
>
> Damn, that sounded way harsher and arrogant than I wanted it to sound.
> And it's only based on what I gathered from the discussions on the
> list archives.  Sorry :(

No harm done, man. I just understood your words in this way. :)

But I do think that Clock-pro deserves its credit, since after all
it's that research work firstly brought the idea of "refault/reuse
distance" to the kernel community. Further more, it's also good
to let the researchers and the community to together have some
brain-storm of this problem if it's really hard to deal with in
reality.

> > 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.
>
> I'm with you on that, but I can't see an alternative in this case.  We

I trend to agree, I once tried to deal with an anti-LRU pattern(e.g. the
big loop like you said) of a app from kernel space and failed. Seems
it's hard to gather a very accurate information of a program's real memory
footprint in mixed workloads with only the help of pte bits...(but also
may due to my lack of skills in tweaking the reclaiming code...)

> can't predict future page accesses very well, so we have to take
> speculative shots and be considerate about the consequences.
>
> And BECAUSE we may get it wrong, the algorithm does not rely on the
> decisions it makes to be correct.  For example, it does not activate
> pages based on refault distance, but requires the refaulted page to
> win the race against an actual active page.  Likewise, pages are not
> evicted from the active list directly, instead they get a chance at
> re-activation when challenged.

Yes. That sounds a smart handling.
--
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/