[ntp:questions] NTP and PPS calibration interval (convergence speed)
david at ex.djwhome.demon.invalid
Tue May 28 07:55:33 UTC 2013
Miguel Gonçalves wrote:
> For me, it is not clear why calibration occurs every 256 seconds when
I don't understand what you mean by this, and I don't see any evidence
of a 256 in your posting.
> I have a GPS receiver with a PPS output. I understand that a good
> frequency drift estimation would be good when free running but frankly
> I don't want my NTP servers to serve time when the GPS signal is not
> available. I can trust my Meinberg box to do that because it has a
> TXCO but not my regular NTP servers.
The way to do this is to set tinker dispersion on the server, to reflect
a realistic worst case clock frequency error, and set tos maxdist, on
the clients, to represent how large an error margin you are prepared to
> I was also surprised that Meinberg's NTP server calibration interval
> never goes above 4 seconds. They must have done what I did.
> I changed my FreeBSD 8.3 kernel to never go above 4 seconds and the
> results are in the graphics attached.
comp.protocols.time.ntp is not a binaries newsgroup. I believe the
mailing list imposes similar restrictions.
> I changed PPS_FAVGDEF in the file /sys/kern/kern_ntptime.c from the
No such file in the source distribution, at least not for 4.2.7p333. No
such symbol in the ntpd sources.
> default 8 (256 seconds) to 2 (4 seconds). The frequency is therefore
> adjusted every 4 seconds instead of 256.
I'm going to guess it has something to do with the averaging time for
the frequency. Generally a longer averaging time will give you a more
accurate frequency, assuming that phase measurement error noise
dominates, which should be true below the Allan intercept.
Unfortunately, your Allan intercept may be exceedingly low and
statistically poorly behaved.
More information about the questions