Poul-Henning,
On 08/27/2014 07:32 PM, Poul-Henning Kamp wrote:
--------
In message <[email protected]>, Bob St
ewart writes:
So here is a pretty interesting way to optimize a GPSDO that I've
been playing with for some years. I don't have a formal mathematical
formulation of it.
It is somewhat related to what Dave Mills calls "the Allan intercept"
except this you can actually measure and not just estimate.
I followed the paper-trail back to an article by Judah Levine, that
describes it in the context of the ACTS system. It's translation over to
NTP is somewhat unfortunate, due to the difference in noise characteristics.
You run several (long!) test-series with different timeconstants
in your PLL, and you record the resultant EFC and phase offset
as a function of time.
If your timeconstant is too short, you will have a lot of
high-frequency signal in the EFC, too long and you get too
much high-frequency signal in the phase offset.
The optimal timeconstant is where you have the least sum of
spectral power where the two curves cross each other.
The "Allan intercept" is rather the intercept point between the Allan
variance of the reference signal and that of the local oscillator. In
the original paper separate measurements made the intercept point more
or less a static value for those assumptions.
The noise of a packet network and the noise of a modem line of that time
is quite different in characteristics.
My experience so far is that the curve around the optimum is
very flat, getting the timeconstant wrong by a factor of two
hardly changes the resultant performance.
I agree. As long as you are in the right neighborhood, you do fairly
well. You just need good enough models to properly model the
neighborhood you're in.
I have yet to seen a thorough ADEV/AVAR analysis of that cross-over
point for optimum performance, all I have seen is translation of the
phase-noise type rule-of-thumb analysis.
Cheers,
Magnus
_______________________________________________
time-nuts mailing list -- [email protected]
To unsubscribe, go to https://www.febo.com/cgi-bin/mailman/listinfo/time-nuts
and follow the instructions there.
