[e2e] Accuracy achievable with NTP

[David P. Reed]

Saad's point is well worth thinking about.   The question is ill 
formed.  The accuracy of a clock stabilized and synchronized by NTP is in 
principle relatively independent of the range of delays in the network.

NTP is, in itself, quite inaccurate.  But when you have a local clock, use 
of NTP can get extremely precise synchronization if you understand what you 
are doing.

If you think about it, when the roundtrip times are short, the accuracy is 
very high, because the cause of variance is merely congestion, and a 
short-round-trip by definition has minimal congestion in its noise 
sources.   Weighting short round-trips heavily, and averaging over a long 
baseline, allows you to synchronize a local clock's rate and deviation 
quite accurately to the master clock.   Very much closer than the RTT/2.

At 11:53 AM 4/10/2003 -0500, Saad Biaz wrote:
>Woojune:
>
>It is not the delay RANGE that hinders precise clock synchronization,
>it is the UNCERTAINTY on the delay that matters.
>
>Given a network of hosts, you must MEASURE the UNCERTAINTY on the delay
>between all hosts. Then find the diameter D of your network with edges
>weighted with the UNCERTAINTY (highest delay - lowest delay).
>
>
>Your worst case imprecision on the clocks will be  1/2 D (Diameter).
>Your average imprecision depends on the distribution of the uncertainty.
>
>For a formal proof, check:
>
>S. Biaz and Jennifer L. Welch, "Closed form Bounds for Clock
>Synchronization under Simple Uncertainty Assumptions".
>
>
>
>
>On Thu, 10 Apr 2003, Woojune Kim wrote:
>
> > Hi,
> >
> > I've been looking for info on how accurate NTP can be. I've looked in 
> various NTP related sites and still not quite sure.
> >
> > Say I have a LAN like environment with delays in the range of 10s of 
> ms. How accurate a time can I get from NTP, assuming I have the 
> appropriate HW clocks on my host ?
> >
> > Say I have a corporate network like environment, relatively segregated 
> from the Internet at large, with delays in the range of 100ms. How 
> accurate a time can I get ?
> >
> > Say the delay range goes upto 500ms, how accurate can the time be ?
> >
> > Any pointers to specific experiments or data would be appreciated.
> >
> > thanks
> >
> > Woojune Kim
> >