[time-nuts] Divide by five

[Magnus Danielson]

Poul-Henning,

On 11/10/2014 09:16 AM, Poul-Henning Kamp wrote:
> --------
> In message <546070B3.1010209 at rubidium.dyndns.org>, Magnus Danielson writes:
>
>> The NTP scatter-plot wedge is another way to present it, and finding the
>> "tip" of that wedge is really about finding the min-value of delay in
>> each direction prior to doing the two-way time-transfer equations.
>
> The important part about the wedge diagram is that the wedge is not
> filled out.  With a very large probability only one of the two packets
> is impacted by extra delays.
>
> The current NTP filter totally fails to exploit this fact and
> performs quite a bit worse because of it:  A median filter is
> really not a suitable way to handle that situation.

Completely agree. Also, NTP uses far too few packets to get any 
meaningful statistics on the edges of the wedge, which is to say the 
minimum value in either direction.

Also, long-term averaging makes the TE part draw towards zero, so if 
there is a sufficiently long run with high asymmetric delay, the node 
tracks in the precieved offset error and thus asymmetric delay will 
affect the time of the node.

> But steering back to the original topic again:  You really should
> not look at NTP for clock-ensemble or clock-steering examples, the
> challenges NTP copes with, are nothing like the challenges you have
> with a bunch of local high-quality frequencies.

Fully agree. In fact, missing this point is part of the problem that NTP 
has.

> There used to be an academic paper on timing.com's home-page about
> their clock-ensemble algorithm called something like "Advances in
> Time-Scale Algorithms".

PTTI 24, Sam Steins work:
http://tycho.usno.navy.mil/ptti/1992papers/Vol%2024_28.pdf

> It's not super detailed but it was certainly a much better place
> to start than NTP source or documentation

I think the simplest case is that of Dave Allan and friends put down in 
the NBS Monograph 140, chapter 9:
http://tf.boulder.nist.gov/general/pdf/194.pdf

They have even thrown in the Fortran code that they used. I think it 
originally ran on a PDP-7 or something.

You do want to read the 9.3 part, as it is the runner-up to the 9.4 part 
which gets dirty into how to do it.

I think that will give a first concrete starting-point, and there is 
more development in both NBS/NIST, USNO etc. (see NIST and USNO PTTI 
archives).

I've long toyed with the idea of building ensemble clock, but never got 
around to it.

Cheers,
Magnus