[ntp:questions] Re: Sufficient # servers to sync to

brian.utterback at sun.com brian.utterback at sun.com
Sat Mar 19 03:10:52 UTC 2005


Yes, I understand all that. That's why I said that in the case of three

servers, you might still find yourself with what is obviously (to
humans) a
falseticker, because even if one is somewhat off of the other two, the
correctness interval might still include all three servers, allowing
all of them to proceed to the clustering algorithm.

I just thought that since it comes up so often and is counter-intuitive
and
kind of subtle, a gross example would make it clearer for those newly
initiated to the world of NTP. Last time I make that mistake, I guess.

Next time I will be sure to add a disclaimer that the example is
totally
made up and may have no relation to reality.


David L. Mills wrote:
> Brian,
>
> The selection algorithm doesn't vote; all it does is establish the
final
> correctness interval. Servers within the interval are truechimers,
while
> those outside are falsetickers. "Voting" among the truechimers is
done
> by the clustering algorithm on the basis of stratum and root
distance.
>
> When I say "vote", what I mean is that a majority vote among three of

> the four servers is possible in order to declare the fourth a
traitor.
> All this says is the three servers are truechimers and the fourt is a

> falseticker. The "vote" is not for the winner, it's for the loser.
>
> Dave
>
>
>
>
> Brian Utterback wrote:
> > Yes, to all of your points. I wasn't dealing with the dispersion
> > intervals and jitter and all that. I was just trying to illustrate
> > why three servers won't cut it. In reality the offsets shouldn't be
> > so large and all that. But if the intervals don't overlap (and they
> > often don't) then the example stands.
> >
> >
> > Brian Inglis wrote:
> >
> >> On Thu, 17 Mar 2005 12:59:29 -0500 in comp.protocols.time.ntp,
Brian
> >> Utterback <brian.utterback at sun.removeme.com> wrote:
> >>
> >>
> >>> Richard B. Gilbert wrote:
> >>>
> >>>
> >>>> Take an extreme case:
> >>>>
> >>>> Server A says it's 11:53
> >>>> Server B says it's 11:55
> >>>> Server C says it's 23:52
> >>>
> >>>
> >>> To add to what Brad said, what you really have is the following
> >>> candidates for the time:
> >>>
> >>> 1. The interval 11:53-11:55
> >>> 2. The interval 11:55-23:52
> >>> 3. The interval 11:53-23:52
> >>
> >>
> >>
> >> If the uncertainty for all servers was 2 minutes, the candidate
> >> intervals would be:
> >>
> >> peer  low   mid   high
> >> A    11:51 11:53 11:55
> >> B    11:53 11:55 11:57
> >> C    23:50 23:52 23:54
> >>
> >>
> >>> Let's see... number 1 has two servers on it, number 2 has
> >>> two servers, and number 3 has all three servers. So, with
> >>> three servers voting for number 3, I guess it is the winner.
> >>>
> >>> This means that all three servers are allowed to proceed to
> >>> the next stage in the selection, so server C may still get the
nod.
> >>
> >>
> >>
> >> The overlap interval would be 11:53-11:55 with servers A and B in
the
> >> clique and C flagged as a falseticker.
> >>
> >>> Just to clarify, the vote goes to the shortest interval that has
> >>> at least n/2 servers on it.
> >>
> >>
> >>
> >> Okay.
> >>
> >>> So, if we add one more server right
> >>> around 11:5x, we see that this would then create a new interval
> >>> with three servers, not allowing server C.
> >>
> >>
> >>
> >> Only if the new server interval overlapped the existing interval.
> >>
> >>> The original long
> >>> interval will have all four, but is trumped by the shorter one
> >>> with three, since 3 is still more than n/2 = 4/2 = 2 servers on
> >>> it.
> >>
> >>
> >>
> >> Depending on the overlap, it could join the existing clique, or
form a
> >> new clique with another server.
> > 
> > 
> >




More information about the questions mailing list