[ntp:questions] NTP - orphan mode with SAME stratum ("tos orphan 6") FAIL to sync.

David Mills mills at udel.edu
Tue Apr 7 17:50:34 UTC 2009


I've tested the time island before and it works even for long periods, 
because once the children have all agreed on a single parent, they 
dutifully follow the parent no matter what it does. My latest test was 
different; it involved two distinct and separate orphan parents, each 
drifting its own way. After a day of drift their correctness intervals 
did not intersect. As of now the intersectio algorithm is bypassed and 
the ranking distance calcualted separately. Seems to work as intended.

However, there are a couple other issues here. I understand that the 
configuration includes two Ethernet rails and each client/server has two 
Etherfaces and presumably two broadcast servers and an association for 
each other machine on both rails. For seven machines each has two 
broadcast servers and 12 broadcast clients. Each of the seven hosts 
would have to mitigate over all 12 client associations. The single 
parent would be the one with the lowest address

Now, start these puppies up with initial offsets in the minutes and see 
if the pinball machine settles down. In the original design they 
probably fall victim of the intersection problem mentioned earlier. If 
they were initially set to within a few milliseconds with ntpdata or 
whatever, they probably would continue as in my earlier tests.

Bottom line is the only real safe solution is the one I described.

Further complications. I don't know what NTP version is in use here. If 
the release version, that includes a mongrel collection of old and new 
routines and cannot be considered safe, especially when using relatively 
new features. A newer snampshot would be much more solid, but only the 
latest development version I have here has the fix needed for the real 
time island.


Steve Kostecke wrote:

>On 2009-04-06, David Mills <mills at udel.edu> wrote:
>>You scratch a sore spot as I verified over the last couple of days. You 
>>can always increase the mindist, but eventually the orphan servers will 
>>drift away from each other.
>Right. Because all of the servers are free-wheeling.
>>The problem is the orphan server has to be selected before the
>>intersection algorithm. I changed that in the development version and
>>am testing it now. The result is that only the lowest-distance orphan
>>server makes it to the intersection algoritihm and only it will be
>Is this a change from what is documented at
>http://www.eecis.udel.edu/~mills/ntp/html/assoc.html#orphan ?
>"While ordinary NTP clients use a selection metric based on delay and
>dispersion, orphan children use a metric computed from the IP address of
>each core server. Each orphan child chooses the orphan parent as the
>root server with the smallest metric."

More information about the questions mailing list