[ntp:questions] Startup transient with missing drift file

[Richard B. gilbert]

Hal Murray wrote:
> What should it look like?  How long should it take to get stable?
> 
> I was expecting ntpd to think for a while, then jump to close to
> the right value, then converge to the final value.  I'd expect
> that to take an hour or so.
> 
> What I'm seeing (recent ntp-dev) is that the jump doesn't get
> very close and the converge part takes multiple hours, depending
> upon how far off the jump was.
> 
> 
> Has anybody else watched a startup transient recently?
> 
> I've been looking at a lot of them while chasing a Linux
> glitch that doesn't compute the CPU frequency repeatably.
> 
> 
> What should happen if I start with an explicit drift value
> that is way off?
> 

I'd expect to see ntpd take a good deal more time pulling in to tight 
synchronization.  Startup is pretty raggedy anyway thanks to a 
non-linearity introduced into (I think, adjtimex()) to make it correct 
errors more quickly.  The nonlinearity messed up Dave's calculations so 
you get an overshoot