[ntp:questions] Question on how to Slew time in NTP (David Woolley)

unruh unruh at invalid.ca
Thu Dec 15 18:29:16 UTC 2011


On 2011-12-15, Richard B. Gilbert <rgilbert88 at comcast.net> wrote:
> On 12/15/2011 10:32 AM, Steve Kostecke wrote:
>> On 2011-12-15, Charles Elliott<elliott.ch at verizon.net>  wrote:
>>
>>>> has the authentication in place to allow it to be added on the fly,
>>>> the biggest hit to the time will be that from stopping and starting
>>>> ntpd, to change the configuration.
>>>
>>> So true! Are there simple instructions somewhere to allow one to
>>> setup NTPD so that ntpq.exe or ntpdc.exe can be used to change the
>>> configuration on the fly, or can someone just post a few lines here to
>>> tell how to do it?
>>
>> There is no "hit" involved in a warm restart of a stable ntpd.
>>
>> Anyone claiming otherwise is spreading FUD.
>>
>
> Is this something relatively new?  ISTR that it still took about thirty 
> minutes to get a reasonable facsimile of the correct time even when 
> restarting with a good drift file.
>
> I wish I could draw a graph here.  Since I can't, I'll try to describe 
> it.  Start with an error of twenty or thirty seconds, mad dash to 
> correct the error, followed by overshoot, and ringing.  Ringing decays 
> slowly. If I were at the controls, I'd be tapping the brakes as we 
> approach zero error!
>

Well, it should not really ring. But ntpd has a very very simple
algorithm. You can see that there is a problem, but that is because you
look at the history, not simply the current data point. ntpd onlylooks
at the current data point. It does not try to estimate from a sequence
of current data where things are going. That is how chrony operates. In
the early days of ntp apparently there was a large battle as to what the
best algorithm was -- a simple higly local in time feedback algorithm or
one which uses many data points and uses regression to make a best
estimate of the offset and rate of the clock. Mills won out with the
simple version. Curnoe implimented the regression model in chrony. Test
has shown that the latter is far faster and disciplines the clocks more
tightly in general than does the former. The former is extremely well
studied (it is so simple that people have been looking at it since the
18th century). The latter is not as well studied, and knowledge as to
how it would react in extreme situations is not as good. 



More information about the questions mailing list