Re: Clock Skew Due to CPU Frequency Management? (was Re: Dual-core system will not create NTP peers)
On Tue, Feb 19, 2008 at 07:37:32AM -0600, Moshe Yudkowsky wrote:
> I had a very kind private communication from a list member. Let me
> reformulate my question.
>
> On the AMD64 dual-core platform, the system clock has far too much drift
> even though the hardware clock seems well behaved. I don't seem to be
> experiencing the "double-speed clock" problem that I see references to
> on the net as of a few years ago, but I am experiencing some other odd
> behavior. The most likely culprit, as far as I can tell from research,
> is some interaction between the hardware's ability to adjust CPU clock
> speed to save power/control temperature (which then influences the
> system clock's theory of what a CPU tick means in terms of actual time
> passing) with the dual-core mode.
>
> But that seems rather wacky. I find it hard to believe that a kernel
> release wouldn't work on dual core CPUs.
>
> But if the problem really is that's the case, I would like to know what
> boot-time settings I can use to test this hypothesis. For that matter
> I'd like to know why the kernel itself isn't picking up on this
> theoretical problem -- ISTR reading about an "enhanced" CPU clock mode
> that could solve this problem, but I can't find references to it or how
> to build a kernel or modules from source that would include this capability.
>
> I've spent a substantial amount of time on this problem, and I can't
> come to any conclusion. I have to wonder if it's a problem with NTP
> itself, but if so it's a problem that only manifests itself on the AMD64
> box, although I continue to fiddle with NTP settings.
Someone here at work was having trouble getting NTP working on their
system a few weeks ago. Disabling spread spectrum in the BIOS solved
the problem.
Of course as far as I know disabling spread spectrum violate emmisions
certifications in europe, but I could be wrong about that since I don't
actually live there (anymore at least).
Spread spectrum clocking really screws with the timings NTP depends on.
--
Len Sorensen
Reply to: