Re: [Ntp] The bump, or why NTP v5 must specify impulse response

[Kurt Roeckx <[e-mail]>]

On Thu, Apr 23, 2020 at 01:38:59PM -0600, Philip Prindeville wrote:
> Not sure I agree…
> 
> 
> > On Apr 23, 2020, at 12:37 PM, Doug Arnold <[e-mail]> wrote:
> > 
> > User case A:
> > A Stratum 2 server get it's time from a pool of stratum 1 servers through the public internet and distributes it to clients in computers, servers and routers in a large enterprise network.  The clients need time to about 1 second for log file entries and security time out intervals.  Accuracy at the stratum 2 server and clients is almost completely determined by queuing noise from so many hops that the noise is almost Gaussian.  After removal of outliers the stratum 2 server and clients average clock corrections in a PLL based on a low pass filter.  Because accuracy requirements are loose, clients request updates only every few minutes.
> 
> 
> As you get into 10GBE networking, events can happen in extremely small intervals.
> 
> Half a second of variance either way can completely obscure if two events are related, which happened first, what’s cause and what’s effect, etc.  I’d say that millisecond accuracy is the minimum useful precision in Enterprise and data center scenarios.

I think your requirements seem to fall under use case B, not A.

If you want millisecond accuracy, the stratum 1 servers should
probably be in the order of 10 millisecond away, and always
reachable, never congested, and you probably need an update rate
that's smaller than 1 minute. And I'm not sure that 10 millisecond
is actually good enough.


Kurt