[time-nuts] Divide by five
magnus at rubidium.dyndns.org
Mon Nov 10 02:49:33 EST 2014
On 11/09/2014 07:11 PM, Bob Camp wrote:
>> On Nov 9, 2014, at 11:18 AM, Chris Albertson <albertson.chris at gmail.com> wrote:
>> On Sat, Nov 8, 2014 at 11:13 PM, Neil Schroeder <gigneil at gmail.com> wrote:
>>> At one point I was considering phase locking all of them together - but
>>> again that seemed less than straightforward. You can do it PLL back to
>>> back, but is there a way to have a loop that contains multiple clocks? I
>>> would think the "telephone game" would apply.
>> NTP does this but on a MUCH lower frequency and longer time scale. But I
>> think NTP's general method could apply. NTP will accept any number of
>> reference clocks. (Yes sone people run NTP using just one GPS receiver as
>> a reference but best practice is to use five references.) NTP compares the
>> set of ref. clocks with each other and first tries to find the subset of
>> clocks that track each other, assuming the outliers are "wrong". It
>> continuously checks this and maintains a set of "true tickers". From these
>> it computes a consensus time using a weighted average of the "true"
>> clocks. The weights are based on the jitter and other quality measuring
>> statistics. Using this method reference clocks can be taken on and off
>> line without need to re-start NTP.
> That may (or may not) give you the best ADEV on the output. My guess is that the filtering algorithm will need to be a bit more complex. NTP’s aim is mainly to throw out bad clocks and pick one as best. We would more likely want to combine the outputs and use all of the good clocks we have. The idea is to improve on the ADEV of the *best* source you have available.
The aim is to remove false-tickers and then build the best ensemble of
the remaining sources and weigh them according to stability.
It seems this goal is not very well met in practice, but the theory
foundation is pretty good.
More information about the time-nuts