On 09/16/2012 05:47 PM, Poul-Henning Kamp wrote:
In message<5C52FBDBA5084AD4A36300FBA73BEF5E@pc52>, "Tom Van Baak" writes:
Yes, timing accuracy has been my main focus and in general I have been
using integration times on the low side of 10000 seconds for that,
but it depends a lot on the OCXO/Rb and environment.
The PLL in NTPns is a (by now) old attempt to make a self-tuning PLL
for optimal time stability, and it does a surprisingly good job at it.
Are there papers that talk about how to optimize for best timing or best
frequency or (no free lunch) some compromise combination of the two?
The only writings I am aware of, is what Dave Mills has written and
the PLL code in NTPns, but I havn't followed this closely in the last
10 years, so do check for newer writings.
Dave Mills coined the term "allan intercept" as the cross over of
the two sources allan variances and it's a good google search for
his relevant papers.
I'm not entirely sure his rule of thumb for regulating to that point
is mathematically sound& precise, but the concept itself is certainly
valid, even if you have to compensate for the timeconstant of the
PLL you use to regulate to that point.
Well, what is being used is phase-noise intercept. Conceptually a
similar intercept point will be available in Allan variance. However, as
you shift between noise-variants, the Allan (and Modified Allan)
variance has different scaling factor to the underlying phase noise
amplitudes. The danger of using the Allan variance variant is that you
get a bias in position compared to the phase-noise plots cross-overs.
However, the concept is essentially the same, and the relative slopes is
the same. You get in the right neighbourhood thought.
The concept has been in use in the phasenoise world of things, so you
would need to search the phase-noise articles to find the real source.
It's been used to generate stable high-frequency signals.
The analysis of PLL based splicing of ADEV curves is tricky, and I have
not seen any good comprehensive analysis even if the general concept is
roughly understood. The equivalent on phase-noise is however well
understood and leaves no magic too it.
I spent a lot of time with the code in NTPns, to try to get that PLL
to converge on the optimum, and while generally good, it's not perfect.
The basic problem is that the data you have available for autotuning,
is the allan variance between your input and your steered source.
You need to treat the data as loose and tight PLL measure, depending on
what you look for. There is loads of calibration issues, covered in
literature.
If you also have the allan variance between the steered source and
a 3rd, better, source, the task is pretty trivial: Minimize the
area below that curve.
But if you do that on the curve you have, you don't optimize, you
pessimize, since the lowest area, is with a timeconstant of zero.
Going the other direction and maximizing the area is no good either
and trying to balance the area around some pivot related to the
present PLL timeconstant does not converge in my experience.
What I did instead was to (badly) reinvent Shewarts ideas for testing
if the phase residual is under "statistical process control":
I increase the timeconstant if the phase residual has too frequent
zero-crossings and loosen it if they happen too seldom.
Having read a lot more about statistical process control, since I
built those NTP servers for the Air Traffic Control 10 years ago,
I would leverage more of the theory and heuristics developed in
process control. (3sigma violations, length of monotonic direction
etc. etc.)
It's a complex field, and things like temperature dependencies helps to
confuse you.
Cheers,
Magnus
_______________________________________________
time-nuts mailing list -- time-nuts@febo.com
To unsubscribe, go to https://www.febo.com/cgi-bin/mailman/listinfo/time-nuts
and follow the instructions there.
