[ntp:hackers] Bug in frequency calculations?

Heiko Gerstung heiko.gerstung at meinberg.de
Mon Feb 4 08:02:07 UTC 2008


Hal Murray schrieb:
>> Unfortunately I have to support more than one system with this kernel
>> and even different CPU types ... It would be a workaround for the
>> development phase and I could hope for the kernel guys to work on this
>>  in the meantime. Do you have any information if someone already has
>> reported a bug for this (on the LKML I suppose) or point me to
>> discussions or newsgroup threads about it?
>>     
>
> I tacked a report on to 7963 since that seemed to cover TSC calibration even 
> though it doesn't quite match this sysmtom.
>   http://bugzilla.kernel.org/show_bug.cgi?id=7963
>
> After a while with no comment, I emailed the people listed.  I got a 
> suggestion to try the latest kernel.  (That code is still being worked on.)  
> It didn't help.  I haven't heard anything since then.
>
>   
I checked that bug report (thanks for the pointer) and applied the patch 
that John Stulz provided, will keep you posted regarding the results.

>> I checked the kernel calibration loop issue and when the kernel
>> calculated a frequency of 498.60, ntpd calculated a drift file value
>> of  0.802 ... power cycling the unit brought up a kernel cpu frequency
>>  calculation of 498.54 and ntpd went off to correct the initial 0.802
>> to  a 19.xxx (after a few polling intervals).
>>     
> Did it stabilize there?  What happened after a few more hours?
>   
It was rock solid over the whole weekend with only slight variations 
(+/- 1 us) over most of the time and a few (2-3) glitches (up to +/-10 
us), which is a good performance I'd say.

> I'd consider a jump of 18 ppm to be a serious bug.
>   
Me too.
>> Seems like the cpu freq calculation works OK ... 
>>     
>
> 498.60 to 498.54 is 60/500000 or 120 ppm.
>
> That size jump takes many hours to stabilizeon my system.
>   
It requires 2-3 hours here, but since we are selling hi-fi ntp equipment 
even this time frame is not acceptable for me. On our current products 
this is a question of minutes and not hours.


Best regards,
Heiko



More information about the hackers mailing list