[ntp:questions] Re: NTP set-up when no external time source becomesunavailable?

David L. Mills mills at udel.edu
Sat Nov 5 02:14:54 UTC 2005


Maarten,

Setups involving more than two cohabitating local clocks are 
fundamentally unstable under some failure scenarios. The recently 
implemented orphan mode is designed for graceful fallback under failure 
conditions. It is in the latest ntp-dev that leaves here, but is still 
pretty new.

You are welcome to try it by using the tos orphan <stratum> in all 
machines in addition to your normal outside servers. For the most 
survivable configuration configure a fully connected mesh network among 
the servers and clients, either using broadcast or symmetric modes. If 
anything fails, the subnet will automatically reconfigure elect a subnet 
root and form a connected tree. If any server or client finds an outside 
source, the tree reconfigures as usual.

Dave

Maarten Wiltink wrote:

> "Donat-Pierre Luigi" <luigi at ict.usc.edu> wrote in message
> news:395EDF817986AB47B7EB5C398ADC817F026667D6 at MAIL.ict.usc.edu...
> [...]
> 
>>The stratum on A and C is set to 14 and it is set to 13 on B.  That
>>is to say, I used each local clock as fall back reference clock and
>>used fudge to set the stratum with bias towards B.
> 
> 
>>After several attempts at debugging and reading NTP documentations
>>and feeling really confused, I can't prevent A and C to drift away
>>without converging or stabilizing (i.e. with a stable offset) with
>>the clock of the Windows XP PC B.
> 
> 
> The bias may be too small. A and C now have a choice between B's
> stratum 14 (it's one step away from its own clock) and their local
> clocks' stratum 14. The local clock is likely to win. Try fudging
> B's clock to stratum 12.
> 
> 
> [...]
> 
>>I would still like PC B to use up to four external servers (as
>>recommended) whenever it has a network connection.  What is important
>>to me is that the Cluster PCs have a small time difference between
>>themselves - i.e. synch together.  That is to say, the accuracy of the
>>absolute time is not as essential, it just need to be reasonable within
>>seconds.
> 
> 
> If they have connectivity _sometimes_, simply configure a few NTP pool
> servers. If they're not reachable, so what? No harm is done.
> 
> 
> 
>>Also, what I would really like is that the two Linux PC (A and C)
>>display the minimum time offset with respect to each other.
>>Shall I configure A and C to peer with each other so that each clocks
>>minimize the relative error with the reference clock from B?
> 
> 
> No, that won't help if B is reachable.
> 
> Groetjes,
> Maarten Wiltink
> 
> 




More information about the questions mailing list