[time-nuts] Divide by five ->Ensemble

[Bob Camp]

Hi

The main point is that NTP picks *one* source from among it’s batch of inputs and uses that. The ADEV of the output can be no better than the ADEV of the output. 

In the case of an ensemble of clocks combined with a better approach the ADEV of the output can be better than the ADEV of the best clock in the group.

No this not an idea I came up with at all. It’s what’s been done by people like NIST and BIH for many decades. It also is the idea behind David Allan’s original Smart Clock stuff. I’d just like to try in in the basement …

Two examples:

NTP has three clocks that are half way around the world with flaky paths, two clocks that are in your local area on the internet,  and  one or two GPS clocks on the local LAN. It needs to figure out that one on the LAN is the best bet. Once it does so, it does it’s best to lock on to that clock and use it. 

We have a group of (say) 16 Rb’s that all are equally good at (say) 4x10^-13 at 1,000 seconds. None of them are really any better or worse than the rest. Our math needs to figure out how to take that group and get 1x10^-13 at 1,000 seconds. It can’t do that by just picking one out of the group and locking to that. (NTP goes a bit crazy when presented with many very good / equally good clocks …)

David Allan has a number of papers on how you *can* indeed get results from a group that are better than any one member of the group. 

Bob

 
> On Nov 9, 2014, at 2:51 PM, Paul <tic-toc at bodosom.net> wrote:
> 
> On Sun, Nov 9, 2014 at 1:11 PM, Bob Camp <kb8tq at n1k.org> wrote:
> 
>> 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.
>> 
> 
> By default NTP makes a paper clock from all acceptable sources (up to some
> maximum) to discipline a/the/its system clock.  One source is special -- "The
> mitigation rules select from among the survivors a *system peer* from which
> a set of system statistics can be inherited and passed along to dependent
> clients, if any" -- but not for the discipline process.  Unfortunately
> "peer" can be a bit confusing in this context so I prefer source.
> 
> This is not to say that the NTP paper clock algorithm is particularly good
> or suitable for other purposes and the code may be simplified because the
> assumption is a relatively small (<10) number of "survivors".
> _______________________________________________
> time-nuts mailing list -- time-nuts at febo.com
> To unsubscribe, go to https://www.febo.com/cgi-bin/mailman/listinfo/time-nuts
> and follow the instructions there.