[ntp:questions] NTP makes a time jump

Brian Utterback brian.utterback at oracle.com
Thu Jul 11 13:32:06 UTC 2013


On 7/11/2013 7:37 AM, Igor wrote:
> yes David, true. was kind of stressed and forgot to update legend and values.
> these are seconds on both axis.
> X is elapsed time and Y is delta between a "real" NTP time and time of server and two clients.
> I'll re-upload gaph.
>
>

The graph shows pretty much what I would expect. Yours is a pretty 
common scenario, but it is also an impossible one.  You want system 
clocks that are undisciplined and unhardened to stay very close in time 
to each other when not connected to the Internet, but to be close to the 
real time as quickly as possible when connected and never jump the 
clock. You need to break one of those requirements for the sake of the 
others. Either get a refclock for your highest stratum servers so they 
are never disconnected from the real time, or harden the oscillators on 
you high stratum servers so that they don't drift when not connected or 
let them step so that they quickly get back to the real time when 
connected.

In fact, it looks to me that you clients are pretty much doing what you 
want. They all stick pretty close to one another as they slew to follow 
the higher stratum servers. I bet they would stick even closer if they 
peered with one another.

Realistically, are your systems really likely to drift to 15 seconds 
offset between connections? How long do they get disconnected for?

Brian Utterback


More information about the questions mailing list