HEADSUP: HZ=1000 by default on i386

[John Baldwin]

On Thursday 04 November 2004 11:36 am, Scott Long wrote:
> Poul-Henning Kamp wrote:
> > In message <xzp7jp1wpli.fsf at dwp.des.no>, 
=?iso-8859-1?q?Dag-Erling_Sm=F8rgrav?= writes:
> >>Poul-Henning Kamp <phk at phk.freebsd.dk> writes:
> >>>So pending any really good arguments to the contrary I plan to increase
> >>>HZ to 1000 on i386 this weekend.
> >>
> >>two good arguments:
> >>
> >> 1) I'm already working on this, and you know it, since I asked you
> >>    about it in Karlsruhe.
> >
> > Ahh, sorry, I got the impression that you were not going to do it on
> > your own.
> >
> >> 2) 1000 is not a good choice, because we can't approximate it well
> >>    with the 8254.  1268 is better, 1381 is even better, 1903 is the
> >>    best we can do between 1000 and 2000, 2299 is the best we can do
> >>    between 1000 and 5000.
> >
> > I played with it here and found that 1000 actually works better than 941.
> > (1193182 / 941 ~= 1268) because the 941 gives a slow beat against 1Hz.
> >
> > It is actually preferable to have a fast beat (jitter) than a slow
> > beat (wander), particularly for people doing benchmarks.
> >
> > Poul-Henning
>
> What timing hardware is used on amd64?  Would it suffer there too?

Identical to x86 for now.  Note that it would be really nice at some point to 
drive hardclock and statclock via the local APIC timers for SMP on x86 and 
amd64 so we can stop sending IPIs for each clock interrupt.  Alpha uses the 
per-CPU timers this way already.

-- 
John Baldwin <jhb at FreeBSD.org>  <><  http://www.FreeBSD.org/~jhb/
"Power Users Use the Power to Serve"  =  http://www.FreeBSD.org