Re: [Ntp] Antw: [EXT] Re: Temperature Compensation for NTP?

[FUSTE Emmanuel <[e-mail]>]

Le 09/12/2020 à 08:05, Ulrich Windl a écrit :
>
> As said before, I think describing the protocol without the algorithms and clock model behind does not make sense, and when describing the clock model, one has to explain where clock wander and error estimate mainly comes from.
> Also one of the design goals of NTP has been to provide "reasonalbe" time and error estimate even when there are no servers available and to lengthen the polling interval. I think considering the temperature (optionally), can help improving those capabilities, so why ignore?
>
Because improving capabilities of the oscillator is not the NTP job. NTP 
algorithm is a consumer of those capabilities, we should have way to 
provide these capabilities to the NTP algorithm but it is not the job of 
NTP to interfere with them.
Chrony implement poor man temperature compensation, great. But if I use 
TCXO or OCXO considering the temperature (and which one ?) is non sense.
And I could even implement this basic crystal TC in a kernel driver for 
better integration with the timekeeping machinery and interface for the 
same result, NTP is completely out of the picture.
If I know the aging curve I could implement it too in this same driver 
or in chrony to better free running precision.
An oscillator could characterized by stability temp stability in ppm/c, 
Alan variance, phase noise, aging curve and more.
For sure, only a metrology lab or the oscillator maker could give you 
the characteristics of an oscillator and on most hardware and computers, 
the oscillator spec are not provided or know.
NTP appliance provider job is to use such advanced oscillators and 
compensation technique to provide better global characteristics and be 
better clock source for serving NTP.
The job of NTP is to steer the system clock and so possibly the 
underlying oscillator in one way and to provide measurements in the 
other way.
So the only question is : what property/characteristics/metric of the 
underlying oscillator/clock the NTP servo/clock model need to do its job 
better rather than default values ?
With the modern granularity and precision of measurements thanks to HW 
time stamping availability and oscillators quality, the spectrum of 
possibility is too wide for hard coded good for all default values.

Regards,
Emmanuel.