[ntp:questions] Re: How to get combined clock offset
David L. Mills
mills at udel.edu
Sat May 7 19:37:13 UTC 2005
By design the two error metrics are root synchroniztion distance
(rootdelay / 2 + rootdispersion + jitter) and combined jitter (weighted
by synchronization distance). The former is the maximum error, the
latter the estimated error. There are tricky little details in the error
budget that have changed over the versions, but the original intent remains.
Please do not use ntpdc as error indicator; it is seriously flawed. Use
ntpq and the rv billboard for the rootdelay, rootdispersion and jitter
displays. Note the jitter display, which includes both peer jitter and
selection jitter, is probably the best indicator of expected time
quality. Read this as follows: the best estimate of the server time is
the offset in the rv display, with jitter as the uncertainty about that
That's all that can be said about the statistics. A more thorough
analysis would have to use the loopstats and compute the distribution
function, such as displayed in the briefings on the NTP project page.
David Woolley wrote:
> In article <mailman.78.1115371918.573.questions at lists.ntp.isc.org>,
> David Lyons <David_Lyons at Raytheon.com> wrote:
>>We need to get a measurement of "Time Quality" for our node. We are using
>>NTP Version 3.
> Why are you interested in quality but using an obsolete version?
>>It seems that the NTP system state variable "theta", the combined clock
>>offset, would be the variable we need.
> Overall theta is offset in loopinfo (ntpdc) and the individual thetas
> are the offsets in the peers displays. It's not a quality indicator.
> If ntpd where confident of the value of theta, it would have corrected
> the local clock by that amount, so theta would be reduced until it was
> no longer confident in the value. In fact, I think that there isn't
> really an overall theta, but rather theta for the most recently selected
> reference clock, at the time it was polled.
> What would be a good measure depends on what you mean by quality.
> ntpd's main quality measure is root dispersion, which is the worst case
> error, assuming that the ultimate reference is a real reference clock and
> not someone's local clock (basically it assumes that all the propagation
> delay could be in one direction and that the clock has drifted at the
> maximum reasonable rate since the last measurement).
More information about the questions