Re: Linux 2.4.4-ac10

Rik van Riel (riel@conectiva.com.br)
Fri, 18 May 2001 14:08:29 -0300 (BRST)


On Fri, 18 May 2001, Mike Galbraith wrote:
> On Thu, 17 May 2001, Rik van Riel wrote:
> > On Thu, 17 May 2001, Mike Galbraith wrote:
> >
> > > Only doing parallel kernel builds. Heavy load throughput is up,
> > > but it swaps too heavily. It's a little too conservative about
> > > releasing cache now imho. (keeping about double what it should be
> > > with this load.. easily [thump] tweaked;)
> >
> > "about double what it should be" ?
>
> Do you think there's 60-80mb of good cachable data? ;-) The "double"
> is based upon many hundreds of test runs. I "know" that performance
> is best with this load when the cache stays around 25-35Mb. I know
> this because I've done enough bend adjusting to get throughput to
> within one minute of single task times to have absolutely no doubt.
> I can get it to 30 seconds with much obscene tweaking, and have done
> it with zero additional overhead for make -j 30 ten times in a row.
> (that kernel was.. plain weird. perfect synchronization.. voodoo!)

Ahhh, I see. Remember that the "cached" figure you are
seeing also includes swap-cached data from the gccs, which
results from kswapd scanning the processes, clearing the
PTE and, a bit later, the process grabbing the page again.

I suspect that if the gccs _just_ fit in memory, you can
get some extra performance by mercilessly eating from the
cache and keeping the ggcs in memory. However, I also have
the sneaking suspicion that this is not the best tactic for
all workloads ;)

regards,

Rik

--
Virtual memory is like a game you can't win;
However, without VM there's truly nothing to lose...

http://www.surriel.com/ http://distro.conectiva.com/

Send all your spam to aardvark@nl.linux.org (spam digging piggy)

- 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/