[ntp:questions] Meinberg NTP monitor, silly question

unruh unruh at wormhole.physics.ubc.ca
Tue Dec 22 21:02:35 UTC 2009


On 2009-12-22, John Hasler <jhasler at newsguy.com> wrote:
> Richard B. Gilbert writes:
>> Lots of luck.  My understanding is that it can't be done without loss of 
>> accuracy and/or stability.
>
> Bill Unruh writes:
>> Nonsense. chrony does it, without loss of accuracy (chrony is about 3
>> times as accurate as ntp is) or stability.
>
> Unfortunately, that stability has only been demonstrated, not proven.
> Know anybody who needs a thesis subject?

It has also not been proven for ntpd with its stepping algorithm, its
clockfilters ( including huffandpuff), its peer selection algorithm,
etc. I cannot see how they could introduce instabilities, but
non-linearities are notorious for doing surprizing things.
The simple feedback loop certainly has been proven stable (
assuming the parameters are appropriately chosen, which ntpd does).

Chrony also impliments a feedback loop but with far more "memory" ( the 
algorithm remembers 5-60 of the past data points, in addition to the
rate which ntpd remembers), which is
easily provable to be stable under the linear assumptions (ie if the
number of points remembered stays constant). It is this extra memory
which allows chrony to respond far more rapidly to changes than does
ntp, while remaining stable.  But the
"breathing" of chrony ( adjustment of the number of points used in the
linear regression) is again a non-linear feature which is much harder to
prove to be unconditionally stable. My interest in chrony began because
I saw features which suggested oscillation in the chrony rate, which
,while not unstable, suggested some sort of amplification of certain
frequencies. I have not been able to track them down to anything in
chrony, but do not have a proof that they are, or are not, an artifact
of the algorithm.  

Yes, a thesis project would be interesting. 






More information about the questions mailing list