[ntp:questions] Polling time backoff

Rob nomail at example.com
Thu Dec 3 10:08:56 UTC 2009


Bill Unruh <unruh at physics.ubc.ca> wrote:
> On Sun, 29 Nov 2009, Richard B. Gilbert wrote:
>>> 
>> A much oversimplified explanation is that short poll intervals are used to 
>> correct large errors quickly and long intervals are to correct small errors 
>> very accurately.
>
> That is true if you are refering to rate errors, it is not for offset
> errors. For offset errors you are probably better off always with a
> short poll interval-- it catches all errors more rapidly. For rate
> errors there is the balance between white noise ( the independent errors
> from measuements of the offset due to delays, etc,) which become less as
> the baseline is increased, and fluctuations in the rate itself. But most
> people want their clock to minimize offset not to minimize rate errors,
> unless you expect there to be long periods of time in which the clock is
> offline.

And of course it is always "better" to have a short poll interval, for
any kind of error that you want to correct.  You may need to add some
filtering to the samples before feeding the result in the loop, but 16
samples in 1024 seconds will always be providing better information than
a single sample in that interval.

I understand that it does not work like this in the ntp reference
implementation, but that is a characteristic of the implementation, not
of the sampling.

Also, a convenient side-effect of the increase of the poll interval is
that systems which have stabilized poll less often, so the load on the
servers is decreased.




More information about the questions mailing list