[ntp:questions] Re: Drift handling....

David J Taylor david-taylor at blueyonder.co.not-this-bit.nor-this-part.uk.invalid
Tue Jan 3 07:58:16 UTC 2006


Harlan Stenn wrote:
> OK, listen to me now and believe me later.
>
> Dave said it better when he said the drift value is updated at every
> clock update.  We only write the value once an hour.
>
> If you really want to play with this stuff, code it up in C in the
> actual codebase and run it thru the simulator (which is part of the
> distribution).
>
> A surprising number of folks have thought they had a good idea on how
> to "make it better" and then got a rude awakening from the simulator.
>
> H


Harlan,

You seem to come across aggressively, although I'm sure that's not your 
intention.

With respect, the /last/ stage would be coding up something.  First, I 
would rather discuss the principle of what might be done, and learn from 
those with a better understanding than I.  Dave has said there might be an 
obscure error, and others have observed that the drift value gained over a 
period of time is too easily discarded, so I'd like to wait until Dave has 
had a chance to study the many reports he has of what happened over the 
leap second before jumping into coding.

My own small contribution has been updated a little:

  http://www.david-taylor.myby.co.uk/ntp/ntp-events.htm

I would still like to understand whether having both a steady value plus a 
diurnal component for the drift is required for best performance, or 
whether a 24-hour average value might be enough.  It does not seem to be 
the optimum behaviour to discard a value gained over many days within an 
hour when writing to the file.  This does mess up ntpd when it restarts 
itself because of the offset being too great.

David 





More information about the questions mailing list