[time-nuts] refclock -> NTP server settings/tuning?

[Anders Wallin]

Hi all,

If I have an NTP server with a refclock that is assumed very accurate (~few
nanoseconds/day drift at most, with << 1 ns jitter), how well should NTP be
able to keep the system clock "on time"?

I'm using a White Rabbit (WR) [1] SPEC PCI-card that receives accurate time
over WR, and I wrote a shared-memory refclock driver [2] for it. I added it
to ntp.conf with
server 127.127.28.0 minpoll 4 maxpoll 4

and I get the following graph. The error is system time compared to
WR-time, where we can assume that WR-time is very accurate.
http://www.anderswallin.net/wp-content/uploads/2013/09/ntp_graph.png

Are the minpoll and maxpoll parameters the only way to tune how NTP locks
onto a refclock? To me the trace (blue in the graph) looks quite noisy and
I would guess it could be improved by "PID-tuning" inside NTP?
As a comparison the red trace shows another computer using standard NTP
without a refclock and the trace is much smoother.
If it matters I measured the free running system clock on this computer to
have an error of about 40 ppm [3].


thanks,
Anders

[1] http://en.wikipedia.org/wiki/White_Rabbit
[2] https://github.com/aewallin/ptp2ntpd
[3]
http://www.anderswallin.net/wp-content/uploads/2013/07/freerunning_vs_ntp_2013jul26.png