[DDL-ML] Re: Question: Low Latency Patch (fwd)

[Volker Wiegand <wiegand@suse.de>]

Liebe Mit-Modellbahner,

hier ist mal eine kurze Info zum Thema "Low-Latency-Patch und SuSE." Im
Ergebnis heißt es, dass man den Patch nicht "einfach so" auf eine SuSE
loslassen sollte. Aber wer weiß, was er/sie tut, kann es schon umbauen.

-- 
 Freundschaftlich / With kind regards
 Volker

---------- Forwarded message ----------
Date: Sat, 23 Jun 2001 18:06:55 +0200
From: Andrea Arcangeli <andrea@suse.de>
To: Volker Wiegand <wiegand@suse.de>
Subject: Re: Question: Low Latency Patch

On Sat, Jun 23, 2001 at 09:05:53AM +0200, Volker Wiegand wrote:
> Hello Andrea,
>
> can you give me some hint whether the so called "Low Latency Patch" that
> used to help with time critical applications is still applicable? And if

you can find some old patch here:

	ftp://ftp.kernel.org/pub/linux/kernel/people/mingo/lowlatency-patches

and others for 2.4 here:

	http://www.uow.edu.au/~andrewm/linux/schedlat.html

> yes, where do I find a version for 2.2.18 and 2.4.x? Or is it incorporated
> in recent SuSE releases? And if not, why not?

Part of it is just incorporated in both 2.2 and 2.4 SuSE kernels.  We
have the reschedule points in the copy_user paths which is on some side
more aggressive than the 2.4 patches from Andrew. Andrew's patch also
covers some place not covered by the copy_user copies themself, and he
also adds some configuration stuff I don't like because it's not needed.
(see the #ifdefs in low-latency.h)

So in short we could just incorporate the other most important
scheduling points from his patches. I guess the most important are the
zap_page_range (during exit(2) or munmap(2)) and then the
filemap_fdatasync/wait (during msync(2) or while flushing a dirty inode
to disk) because they're quite heavily run during production and those
lists may be long.

However also after merging those scheduling points you're not going to
have any "maximal" latency guarantee for example for RT tasks, but it
could decrease the max latency under some workload.

Another approch is to make the kernel preemptive but that's quite
invasive change.

> Questions, questions, ... but I hope you (and Monica) are well !?!

Yes, we're fine thanks, btw Monica got the degree in economics with '110
e lode' (maximum) a few months ago and she's now living in Edinburgh for
some month in order to study english. (I will meet her next month after
the UKUUG conference :)

I hope you and your family are well too!

Andrea