Re: [patch] i386, numaq: enable TSCs again

[john stultz <johnstul@us.ibm.com>]

On Fri, 2007-05-25 at 11:16 -0700, Dave Hansen wrote:
> On Fri, 2007-05-25 at 10:41 +0200, Ingo Molnar wrote:
> > * William Lee Irwin III <wli@holomorphy.com> wrote:
> > > > yes, that's what i meant under 'slightly async'. Some AMD CPUs are 
> > > > like that too and sched_clock() now handles that fine. So we should 
> > > > try my patch.
> > > 
> > > Sorry, then. I took slight to mean something else. In any event I was 
> > > only quantifying things. I've no opinion whatsoever on the impact of 
> > > the code on NUMA-Q, only some recall of its operating characteristics.
> > 
> > there's no need to apologize at all! Thanks for reminding us about the 
> > time-scale and nature of the TSC drift on NUMAQ. I was worried that 
> > maybe the TSC was totally unusable for some reason - but that's 
> > fortunately not the case. So we now have one quirk less, hopefully :-)
> 
> Last I remember, it was totally useless for timekeeping, but was useful
> for cpu-local time measurements.
> 
> John, it's still useless for time, right?  Does sched_clock() really fix
> it?

Yea, on multi-node NUMAQ the TSC shouldn't be used for timekeeping.

However it should be fine for sched_clock(), or other cpu-local
measurements as the TSCs are constant (no cpufreq, no deep sleep
states).

-john