On Thu, 26 Apr 2001, Marcelo Tosatti wrote:
> On Thu, 26 Apr 2001, Mike Galbraith wrote:
>
> > > Comments?
> >
> > More of a question. Neither Ingo's nor your patch makes any difference
> > on my UP box (128mb PIII/500) doing make -j30.
>
> Well, my patch incorporates Ingo's patch.
>
> It is now integrated into pre7, btw.
>
> > It is taking me 11 1/2
> > minutes to do this test (that's horrible). Any idea why?~
>
> Not really.
(darn)
> If you have concurrent swapping activity, pre7 should improve the
> performance since all swap IO is asynchronous now. Only paths which really
> need to stop and wait for the swap data are doing it. (eg do_swap_page)
I'll grab virgin pre7 in a few.
> > (I can get it to under 9 with MUCH extremely ugly tinkering. I've done
> > this enough to know that I _should_ be able to do 8 1/2 minutes ~easily)
>
> Which kind of changes you're doing to get better performance on this test?
Prevent cache collapse at all cost is #one. Matching deactivation rate
to launder/reclaim.. et al. Trying HARD to give PG_referenced a chance
to happen between aging scans [1].
-Mike
1. pagecache is becoming swapcache and must be aged before anything is
done. Meanwhile we're calling refill_inactive_scan() so fast that noone
has a chance to touch a page. Age becomes a simple counter.. I think.
When you hit a big surge, swap pages are at the back of all lists, so all
of your valuable cache gets reclaimed before we write even one swap page.
-
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/
This archive was generated by hypermail 2b29 : Mon Apr 30 2001 - 21:00:15 EST