[ntp:questions] NMEA ref.clock better than my ISP's timeserver?

Unruh unruh-spam at physics.ubc.ca
Fri Jun 12 08:35:47 UTC 2009


"David J Taylor" <david-taylor at blueyonder.not-this-part.nor-this.co.uk.invalid> writes:

>David Lord wrote:
>[]
>> I only have a pair of servers as peer and that is maxpoll 8 and
>> both sat at 256 sec. Offsets are 203us and 593us.

>I see more like 1-3ms for the Internet servers (compared to the GPS), with 
>delays in the order of 30ms.  This is with Windows, though, not a UNIX 
>system.

>> I have problems with the larger maxpoll value as temperature
>> changes are sometimes at a higher rate than ntpd can compensate
>> for but this affects the pcs differently and having peers with
>> lower maxpoll might help. Otherwise the higher maxpoll does
>> seem to tend to give lower offset variation and jitter. So far
>> this year I've not had temperature in back room shoot from 15C
>> to near 30C as happened on a couple of days last year.
>>
>> David

>Yes, it's the classic trade-off - high precision requires a long 
>time-constant, but that means a poorer response to temperature changes. 
>You have to decide when you have it "good enough", otherwise you will be 
>tinkering for life!

I disagree, A long time constant is good for getting an accurate rate
for the clocks, but that is useless if the rate changes. A long time
scale is not of any particular help in getting an accurate offset in
ntp. NTP basically uses a fixed number of data points to estimate the
offset, no matter what the poll is. Ie the exponential time constant is
a multiple of the poll interval. This means that the statistical noise
is reduced by the square root of that number, which, being a constant
independent of poll interval, the offset variance is the same. no matter
what the poll interval. The rate variance is dominated by temperature
fluctuations in computers, and it does no good to have the poll
interval anywhere near the time scale of the temperature variation,
because of ntp's terrible response to rate changes.  

Now you could get far better response by using the temperature of the
computer to also correct for rate responses. 

>Cheers,
>David 




More information about the questions mailing list