[ntp:questions] Jitter versus polling interval

Unruh unruh-spam at physics.ubc.ca
Sun Mar 1 19:08:44 UTC 2009

"David J Taylor" <david-taylor at blueyonder.neither-this-bit.nor-this.co.uk> writes:

>David Woolley wrote:
>> David J Taylor wrote:
>>> Neither, directly.  I mean the peak-to-peak variation in the reported
>>> offset, as "measured" by eye on a graph like the "Daily" graph here:
>> For a Millsian network, the peak to peak value would be unbounded, so
>> one needs to consider RMS, or assume a non-Millsian constraint, such
>> as bounded peak to peak measurement error.
>> Thinking more about it, I think there will be no nett effect from the
>> measurement error, so the only contribution will be from oscillator
>> frequency.  It looks like that is assumed to vary as the square root
>> of poll interval.
>> So, I would say, for small poll intervals, asymptotic to constant and
>> for large poll intervals, asymptotic to the 1.5th power) of the
>> interval.  ntpd attempts not to choose poll intervals that take it
>> beyond the transition between these two, so maybe sub-linear for all
>> of the normal operation range.  If you get something significantly
>> different, you may not have a Millsian network, and ntpd may not be
>> the right tool for you.
>> If you want the real answer, you will need to ask Dave Mills, but he
>> will probably just point you to his mathematical analysis, and I
>> suspect you are asking here because you don't feel competent to use
>> those, or rich enough to buy the book.


>A figure of the 1.5th power is fine for me.  As I only have two "by eye" 
>measurements, I can neither say that's right or wrong.  It's certainly not 

???? From two measurements you can conclude nothing whatever.

Mill's assumptions about the behaviour of clocks is almost certainly wrong.
Not least because computer clocks are in a highly variable thermal
environment. This means that the drift rate varies with time of day and day
of week and is NOT 1/f or any other time invariant spectrum. Tests have
been run where the temperature is measured and an additional correction due
to temperature is factored into the ntp algorithm, and vastly improved
results are obtained. Chrony gets better results, primarily I believe
because its use results in a varying Allan intercept, and the clock rate
adjustment benefits from that in the noisy thermal environment of computers
which are used for things other than timekeeping. 

>The book is on my Amazon wanted list, but I'm not sure how much I will 
>appreciate if it's mainly maths.  One of my criteria for judging image 
>processing papers was "doe it contain images?"!

And a book on time should contain sounds of a clock ticking? Mathematical
analysis is crucial to proper operation of the system.
But it does have lots of graphs if that helps.


More information about the questions mailing list