[ntp:questions] Keeping NTP Honest

Unruh unruh-spam at physics.ubc.ca
Wed Jul 15 20:44:57 UTC 2009

"David J Taylor" <david-taylor at blueyonder.not-this-part.nor-this.co.uk.invalid> writes:

>Unruh wrote:
>> There is not terribly much that can at present be done to improve ntp
>> so
>> it would be hard to find ways of spending the money. While with
>> chrony,
>> you would have a totally new time keeping program for Windows at the
>> end
>> of the day, a program with a very different philosophy and technique
>> for
>> controlling the clock.

>Thanks, Bill, but I don't want a totally new program, if I can avoid it. 
>I have various monitoring and management tools written and in place based 
>around NTP, and I expect all those would have to change as well.  Is it 
>not possible for the chrony algorithms to be made available for NTP as a 
>user-configurable option, leaving the majority of the architecture 

Sure, that could also be an option ( an option I actually hoped would
occur a while back when chrony had not refclock support). I tried hard
to point out the advantages of the chrony approach (linear regression
and remembering the last 10-50 measurements to use in estimating the
error and the rate) but there is some strongly entrenched feeling that
the ntp way ( Markovian, almost memoryless error and rate estimation)
was the way it was going to be. The ntp algorithm has had 20 years of
testing and has come through that testing in an incredibly wide variety
of situations reasonably well. Would the chrony algorithm do as well? I
do not know. On the little bit of testing I have done, it has strongly
out-performed ntp, both in speed of convergence and in its error of
clock parameter estimation ( about 2-3 times better than ntp). That
however is a set of tests in a very small number of situations. Would it
do as well as ntp in delivering time to a Mars base? I do not know. I
cannot see why not, but then again my not being able to see problems and
the existence of problems are not necessarily synonymous.

Note that there is also the issue of who is going to do it. Chrony is a
huge program, as is ntp. It would need someone really conversant with
both to integrate chrony into ntp. That is an extremely small set of
people (0?) And they would have to want to do the job. 


More information about the questions mailing list