Re: [RFC] kernel facilities for cache prefetching

[Wu Fengguang <wfg@mail.ustc.edu.cn>]

Diego,

On Tue, May 02, 2006 at 02:46:41PM +0200, Diego Calleja wrote:
> El Tue, 2 May 2006 15:50:49 +0800,
> Wu Fengguang <wfg@mail.ustc.edu.cn> escribió:
> 
> > 	2) kernel module to query the file cache
> 
> Can't mincore() + /proc/$PID/* stuff be used to replace that ?

Nope. mincore() only provides info about files that are currently
opened, by the process itself. The majority in the file cache are
closed files.

> Improving boot time is nice and querying the file cache would work
> for that, but improving the boot time of some programs once the system
> is running (ie: running openoffice 6 hours after booting) is something
> that other preloaders do in other OSes aswell, querying the full file
> cache wouldn't be that useful for such cases.

Yes, it can still be useful after booting :) One can get the cache
footprint of any task started at any time by taking snapshots of the
cache before and after the task, and do a set-subtract on them.

> The main reason why I believe that the pure userspace (preload.sf.net)
> solution slows down in some cases is becauses it uses bayesian heuristics
> (!) as a magic ball to guess the future, which is a flawed idea IMHO.
> I started (but didn't finish) a preloader which uses the process event
> connector to get notifications of what processes are being launched,
> then it profiles it (using mincore(), /proc/$PID/* stuff, etc) and
> preloads things optimally the next time it gets a notification of the
> same app.

My thought is that any prefetcher that ignores the thousands of small
data files is incomplete. Only kernel can provide that info.

> Mac OS X has a program that implements your idea, available (the sources)
> at http://darwinsource.opendarwin.org/projects/apsl/BootCache-25/

Ah, thanks for mentioning it.

It seems to do the job mostly in kernel, comprising of 2400+ LOC.
Whereas my plan is to write a module of ~300 LOC to _only_ provide the
necessary info, and leave other jobs to smart user-land tools.

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