[ntp:questions] Re: ntpd going wild

Richard B. Gilbert rgilbert88 at comcast.net
Fri Feb 17 17:16:43 UTC 2006

Hermann Hastig wrote:

> Hi,
> Johan Swenker wrote:
>>Last summer something similar happened to me. but it strongly depended on
>>my configuration. So please tell us about your configuration (like
>>operating system etc.)
> It's a plain Linux 2.4 kernel on a Linux from scratch system, nothing
> extra that plays with the system's clock. The machine is located in a
> fairly large datacenter and does not have any other clock source than
> the three upstream servers from de.pool.ntp.org.
> So, I think I can say that the datacenter's network connection has
> definitly been stable as is the case with the temperature (as Richard
> guessed), and there's also no external clock that could mess things up
> somehow.
> I've watched the clock the last days. The drift file stays at 9.642
> while the offsets to the (new) servers move between 1 and 20 ms (except
> one that keeps shifting stratum between 3 and 5 and shows offsets up to
> 83 ms).

I'd consign the server "shifting stratum between 3 and 5" to the garbage 
can if I were you!

> Seems pretty fine to me (I don't care if the clock's a second wrong or
> not). Or am I wrong here?

If your clock is off by as much as one second, something is wrong 
somewhere.  I expect, and usually get, offsets less than twenty 
milliseconds using only internet servers.  I do much better than that 
with a GPS reference clock.  Some servers are not very reliable, 
especially those synchronized by an HF radio signal but if you configure 
four or five it's easy to see which servers are not delivering good time.
> About that maxpoll setting: Of course the offset may grow larger the
> longer the poll interval which in turn makes it more difficult to keep
> the local time right. But doesn't ntpd automatically keep the poll
> interval within limits that still allow to slew the clock, or does it
> simply go up until maxpoll, whatever may happen?

The MAXPOLL setting does not mean that ntpd will ever increase the 
polling interval to that value.  I do believe, however, that capping the 
poll interval at 2^10 seconds has some value in preventing a situation 
where it might take hours or days to correct an error.

More information about the questions mailing list