[ntp:questions] Tighter regulation?

David Woolley david at ex.djwhome.demon.invalid
Sat May 25 21:01:34 UTC 2013

unruh wrote:

> Note that on maxpoll 10, the clock will freerun for about 7000 sec
> between disciplines. 5ms in 7000 sec is about 1PPM. Now if your computer

The time constant for the proportional correction is 16384 seconds, so 
only about 35% of the correction would be made over 7000 seconds.  In 
terms of correcting a random step phase change some time in that 
interval, one is only talking about losing an average about 17% of the 
total correction required as a result of the lower effective sampling rate.

I think the time for the integral component to remove a frequency step 
is even longer (I think it actually rings), except that the rapidly 
growing offset will cause the poll interval and time constant to drop. 
I suspect, if the poll interval remained constant, the reciprocal of the 
loop frequency would be so much longer than 7000 seconds that there 
would be little difference between 1024 and 7000 second effective 
sampling rates.

As I said above, the real issue is that of correctly identifying a step 
in the frequency, or its first derivative, and rapidly turning down the 
time constants.

I am also wondering whether setting a lower Allan Intercept would help. 
     One problem is that it assumes gaussian behaviour of frequency 
errors, but the situations where high offsets actually indicate a bad 
time generally involve very non-gaussian statistics.

> heats up and cools down by 20C over an hour that could change the clock
> by a few ms. And ntpd is slow to respond to changes in clock rate. 
> What happens if you do maxpoll 6? Give us details, not vague statements. 

More information about the questions mailing list