[ntp:questions] Keeping NTP Honest Redux

Unruh unruh-spam at physics.ubc.ca
Sat Nov 7 00:08:29 UTC 2009

"Richard B. Gilbert" <rgilbert88 at comcast.net> writes:

>Terje Mathisen wrote:
>> Evandro Menezes wrote:
>>> I've come to the conclusion that since NTP doesn't correct for
>>> temperature variations, it's long-term compensation is counter-
>>> productive.  Therefore, instead of letting the poll period increase to
>>> the default 1024s, perhaps it'd be best to limit it to 64s.  However,
>>> this requires that each server has the option maxpoll attached to it.
>>> It'd be more convenient to have set this parameter for all servers
>>> from a single place, such as the old tinker maxpoll option.
>>> So, why was tinker maxpoll removed?  Any chance to get it back in?
>> Maybe because the algorithms have been optimized to get the best 
>> possible time with minimum load on the servers?
>> Making it more convenient to lock polling at 64s isn't a good idea, imho.

>Locking the polling interval at ANY value probably isn't a good idea. 
>If not tampered with, NTPD will adjust the polling interval to the 
>optimum value for the conditions then obtaining.  The conditions do change.

Well, not really. ntp is designed to quite rapidly go to maxpoll of 10
and to stay there. The routine is very loath to decrease the poll and
quick to increase it, so that it rarely decreases the poll interval. Ie,
it is designed under the assumption that there is a optimal point of
operation ( where offset noise equals rate noise) and wants to sit
there. Unfortunately, that model is just wrong, precisely because of
temp fluctuations during the day. The rate noise is NOT gaussian, and is
far larger than the model assumes. To make things worse, ntp throws away
80% of the data it does collect in the clock filter, a procedure which
is violently against the claimed policy of trying not to overload the
network (It reminds me of the David-Saul story where Saul was thirsty--
they were under siege-- and some of the men fought through the enemy and
managed to bring back some water, and Saul threw it out on the ground
because he could not drink water which had be obtained at such a high
price.) While I understand the reason for the profligate misuse of data,
it is very weak except in certain exceptional cases. Ie, ntp is far from
optimally designed, either in its poll mitigation nor in its loading of
the network. 

But despite that, I agree that going to a maxpoll of 6 on public servers
is not the right answer. 

>NTPD is at its best from about 2300 local time to 0700 local time. The 
>net quiets down and NTP packets travel with minimal and highly 
>predictable delays.

Yes, I suspect that some of the OP problems have nothing to do with
temperature, but with network congestion instead. That going to a lower
poll will not help with. 

More information about the questions mailing list