[ntp:questions] Re: How often should server hop happen?

David L. Mills mills at udel.edu
Thu Feb 3 02:25:17 UTC 2005


The clock select algorithm has a hysteresis variable for each 
association. It is charged up when the association is first selected and 
dissapates slowly after that. The clustering algorithm uses that as a 
bias to discourage, but not prevent, clockhopping. However, if there are 
multiplie sources all with similar jitter and dispersion, it makes 
little difference which one is selected, even if a clockhop does occur.

When NTP was in its teens, expected jitter was in the milliseconds, but 
now is in the microseconds. The clock mitigation algorithms have become 
really picky in order to work well in the fastest computers and 
networks. So, it doesn't take much flutter, if that word applies, to 
cause clockhopping when trimming the microseconds on the fastest 
networks. But, in those cases, the hop does not degrade accuracy, so why 
bother about it? And, in a very real sense, it is not the system peer 
that determines the offset, but a weighted average of all cluster 
survivors. Even with a hop the average doesn't change much.

For that matter, why report a clockhop to the log in the first place? If 
you are worried about the health of the server ensemble, what you should 
report is when a server drifts in and out of the correctness interval, 
since that is usually the cause of a clockhop.


David J Taylor wrote:
> I know it's like asking how long is a piece of string, but how often would 
> you expect server change to happen?
> I.e. an NTP client with say 5 servers in its list.  Would you expect it to 
> change server serveral times an hour, several times a day or what?  I'm 
> seeing several times an hour and wonder if I should put in a "prefer" 
> qualifier to reduce this.  I appreciate it does no harm, but it does 
> rather clutter the Windows event log!
> Thanks,
> David 

More information about the questions mailing list