[ntp:questions] Re: Sufficient # servers to sync to

David L. Mills mills at udel.edu
Sat Mar 19 18:38:36 UTC 2005


Guys,

I am considered by most to be somewhat a carmudgeon and I apologize in 
advance for my shortcomings. However, this discussion has gone on too 
long without reference to the facts. There are really important 
principles behind the design of this monkey. These are presented in a 
carefully thought out briefing on the architecture, protocol and 
algorithms at the NTP project page. Careful inspection of the 22 slides 
in that presentation reveals most of the secrets involved.

As a necessary precursor to futher dialog, may I strongly advise we 
start from those principles. That is not in any way intended to 
proscribe further argument, just to plead that we all start from the 
engineering model and assumptions. I welcome dialog on the wisdom or 
lack of it about the various algorithms.

Dave

Richard B. Gilbert wrote:

> David Schwartz wrote:
> 
>> "Brad Knowles" <brad at stop.mail-abuse.org> wrote in message 
>> news:mailman.29.1111192970.588.questions at lists.ntp.isc.org...
>>
>>  
>>
>>> At 3:16 PM -0800 2005-03-18, David Schwartz wrote:
>>>   
>>
>>
>>  
>>
>>>>     The assumption is that you have three servers, two of which are 
>>>> working
>>>> and one of which is not. If two of them are working, then their 
>>>> intervals
>>>> *must* overlap.
>>>>     
>>
>>
>>  
>>
>>> They can't overlap, since it takes two machines to *define* an 
>>> interval. If you have three machines, you may have as many as two 
>>> different intervals.  If one of those machines is insane, then you're 
>>> back to the well-known two-machine problem.
>>>   
>>
>>
>>    A machine reports a time of X and it reports its own accuracy as Y 
>> and the round trip time is Z. Why does this one machine not define an 
>> interval X+Y+Z to X-Y-Z?
>>
>>  
>>
>>>>                  Obviously if you have three servers and no two of 
>>>> them are
>>>> reliable, you are screwed.
>>>>     
>>>
>>> You're pretty much screwed if you have just three servers and any one 
>>> of them is unreliable.
>>>   
>>
>>
>>    I don't see why. Common sense say that if you have three servers 
>> and they report:
>>
>> 10:11
>> 10:11
>> 12:03
>>
>>    The time is probably 10:11.
>>
>>    DS
>>
>>
>>  
>>
> Ummm....   When is the last time you saw two servers reporting exactly 
> the same time?   The difference is usually in the range 0.5 to 40 
> milliseconds but it's there.  If they do happen to agree, is it random 
> chance or do they both actually have the correct time?
> 



More information about the questions mailing list