[ntp:questions] Re: NTP sync problems
David L. Mills
mills at udel.edu
Thu Dec 16 20:12:45 UTC 2004
The prefer peer designation does not affect the mitgation algorithm,
which means it could be booted off the intersection as well as any
other. However, it is never booted off the clustering algorithm, but in
fact if it would otherwise be selected for boot, the algorithm
terminates right there.
An empty intersection does not ordinarily occur if the servers are
operating correctly, as the dispersion is inherited along the path and
the confidence interval inreased accordingly. What has been a problem
with very good reference clocks, nearby servers and fast networks is
that the confidence intervals can get really tiny, like a couple hundred
microseconds or so. This tends to amplify even tiny network wiggles due
to Ethernet collisions, switch latencies and ARP cache refreshes.
My Blade 1500 with PPS routinely shows PPS offset and jitter of a
microsecond or two and dispersion 250 microseconds. It is also watching
a CDMA-based Ethernet server over a 100-Mbps Ethernet. It shows the
corresponding numbers of 45, 64 and 5077 microseconds. So, if the PPS
and CDMA offsets differed by more than 5 ms, a Mexican standoff would
have to be declared.
Strangely enough, I was watching just now, and found a curiousity. The
Blade and CDMA server are on different 100-Mbps Ethernet stubs connected
by two switches and a 100-Mbps fiber link. The roundtrip delay is
typically 0.9+-.05 ms, but every once in a while the delay becomes
3.5+-0.5 ms with offset -1.5 ms. It looks like the delay is only on one
leg due probably to the CDMA.
Terje Mathisen wrote:
> David L. Mills wrote:
>> There might be some misconception here. The mitigation algorithm
>> establishes a bound on the offset of any server with respect to
>> roundtrip delay and dispersion. This is a matter of physics and
>> theory. When multiple servers are involved, the algorithm computes an
>> intersection interval within which the server clocks can be considered
>> valid. A possibly correct time must be contained in this intersection.
>> If the intersection is empty, a correct time cannot be established.
>> This is independent of the stratum.
> Sure, this is obviously right.
> My suggestion was based on the idea that if the secondary server would
> always use the primary as its own preferred source, then it would have
> to have more or less the same time, right?
> Actually finding two servers with no intersection interval pretty much
> proves that either one or both of them are misconfigured, and without a
> proper UTC reference.
More information about the questions