[ntp:questions] Re: Peer question

David L. Mills mills at udel.edu
Thu Sep 11 16:32:22 UTC 2003


Actually, the protocol dance upon restarting one peer relative to
another is much more complicated than you might first imagine. Each peer
has to verify that the response from the other corresponds to a packet
it sent. When you restart one peer of the pair, the other one is
confused and it may take a couple of exchanges for the sequence numbers
(timestamps) to be believed one against the other. Also, when one peer
is restarted and the poll interval is 64 s while the other has already
ramped to 1024 s, it may take at least one poll exchange at 1024 s to
stabilize the protocol.

Consider the case when one peer as symmetric passive has synchronized to
an external source and the other peer starts the protocol as symmetric
active. The passive one mobilizes an association, but does not beleive
the time from the active one until the active one is synchronized too.
After a few rounds the active one is indeed synchronized to the passive
one, but then the passive one has to dance the steps to believe it. Only
after both peers have danced the jig has the protocol stabilized.

Now, imagine the dance when public key cryptography (Autokey) is
involved. Much to complicated to revere here; see the NTP project page
for dreary dance steps.


"B.A.Baumgart" wrote:
> I rebooted one of my servers, partly to move some cables and partly to
> watch it rejoin the NTP subnet.  It established communication with the
> machines that its conf file has as servers, but did not talk to its peer
> for over ten minutes.  The peer has a key associated with it.  Since ten
> minutes is (very) roughly half of a poll of 1024, I figure that it wouldn't
> talk until the peer tried to talk to it at its normal poll.  Is this even
> close?
> If it is right, it answers a yet unasked question.  What are the
> disadvantages of increasing maxpoll to 12 or more?
> --
> Bruce Baumgart
> bab at inel.gov

More information about the questions mailing list