[ntp:questions] 500ppm - is it too small?

David Mills mills at udel.edu
Thu May 7 20:52:07 UTC 2009


Terje,

Don't conclude the 1000-Hz Linux problem is lost interrutps. My old 
Alpha under Tru64 ran at 1024 Hz and never lost an interrupt.

 From previous posts it seems the Linx kernelmongers didn't realize 
changing the timer frequency requires changing the kernel discipline 
phase and frequency gains. The phase gain is multiplied by the inverse 
ratio of the new frequency divided by 100 Hz and the phase gain is 
multiplied by the inverse square of this ratio. Some gent some time back 
did that and Linux got well. I don't know what the Windows timer 
frequency is or whether it has a kernel discipline similar to Unix, but 
that consideration might apply there, too.

If the Windows kernel uses only the Unix adjtime() model and no kernel 
discipline, the above does not apply.

Dave

Terje Mathisen wrote:

>Hal Murray wrote:
>  
>
>>>An greater than 500 PPM suggests seriously broken hardware!
>>>      
>>>
>>Or software.
>>
>>    
>>
>Indeed.
>
>This sounds a lot like the trouble many Linux boxes got into with 
>HZ=1000, where dropped timer ticks could cause all sorts of problems in 
>the form of unstable systems clocks that seemed to lose a lot more than 
>500 ppm.
>
>Terje
>
>  
>




More information about the questions mailing list