[time-nuts] Holdover recovery

[jimlux]

On 4/3/16 8:39 PM, Hal Murray wrote:
>
> Has anybody studied what happens when a GPSDO comes out of holdover?  Has
> anybody seen any specs?  I don't think I have.
>
> I think you have a choice of quick recovery for time or frequency, but you
> can't get both.
>
> Suppose your setup has been in holdover for a while.  The frequency is
> slightly off.  The time offset of the PPS pulse will be the integral of the
> frequency offset.
>
> What happens when you come out of holdover?  If you fix the frequency, the
> PPS will stay off.
>
> Suppose the PPS has drifted by 1 ns.  If you correct that in 1 second, the
> frequency will need to be off by 1E9 during that second.
>

This is something we've been developing some algorithms for. (and I 
think NTP does this already).. depending on whether you were high or low 
in frequency, you need to smoothly bring the frequency back, with some 
overshoot (to "catch up").

I think it really boils down so some simple requirements:

Time (aka phase) must always be monotonically increasing
Time and frequency (which is the derivative of time) and some number of 
higher order derivatives must be continuous.
The maximum error is bounded to some number.



As you've noted, going for the minimum error in time might lead to a 
discontinuity in frequency.

We have an interesting problem in that we want to discipline a clock 
which is inherently of high accuracy to a reference that is poorer in 
accuracy, but which is defined as the "master" so we need to follow it. 
  The usual NTP and similar time discipline algorithms are really 
predicated on a "flow down" from more accurate to less accurate.