Dual Core: clock runs too fast (solved) (was: Re: [CentOS] Cool 'n Quiet)

Tue Aug 22 14:31:24 UTC 2006
Kai Schaetzl <maillists at conactive.com>

Kai Schaetzl wrote on Mon, 21 Aug 2006 18:44:16 +0200:

> I tried 
> to repro those network transfer issues and couldn't. But the keyboard repeat 
> stuff reported later sounds "promising".

Actually, I was later able to repro the network transfer issue as well. You 
just need to transfer a lot before it is visible. As someone said in the 
original report it takes at least 600 MB, for whatever reason. But then it 
jumps really quick (3 to 5times as normal) and doesn't stop until you reboot.

A better bug report/discussion is here:
http://bugzilla.kernel.org/show_bug.cgi?id=5105
There's even a (yes!) Microsoft KB article at 
http://support.microsoft.com/kb/918461/en-us which describes the problem very 
well, just for a different scenario.
It also explains why I don't get lost ticks logged: there aren't any. However, 
the kernel thinks because of the unsynced TSCs he lost some and adjusts for 
that. As I understand clock=pit switches that adjustment algorithm off and so 
the problem is gone.
The solution is to use the clock=pit kernel boot parameter. All my symptoms are 
gone with that. My clock is not rock-stable now, it leaped four seconds from 
yesterday to today, but this might be caused by something else and is something 
ntp can deal with.
The obvious solution for the kernel would be to use only one core for TSC and 
not let it balance to other cores. Don't know if the patch from that bug report 
does that or something else or if it really works.

Kai

-- 
Kai Schätzl, Berlin, Germany
Get your web at Conactive Internet Services: http://www.conactive.com