[ntp:questions] stupid, simple question about precision

Simple Simon ssimon at domain.invalid
Mon Nov 20 18:53:30 UTC 2006


rgilbert88 at comcast.net wrote...
> Simple Simon wrote:
> > rgilbert88 at comcast.net wrote...
> > 
> >>Simple Simon wrote:
> >>
> >>
> >>>If 'ntpq -c rl' shows the precision of my clock as -20 (roughly, 
> >>>9ns), does that translate to timestamp confidence interval of +/- of 
> >>>9ns, or of +/- of 4.5ns) ?
> >>>
> >>>Feel free to point to me to the section of the NTP FAQ I've missed, 
> >>>or any other resource that I haven't seen in trying to answer this 
> >>>question on my own.
> >>>
> >>>Any guidance greatly appreciated.
> >>
> >>I think you are mistaking precision for accuracy.  Precision is the 
> >>measure of the shortest interval your clock can represent.  It has 
> >>nothing (or very little) to do with accuracy.
> >>
> >>ntptime will show you the estimated error for your clock.  Unless you 
> >>are using a hardware reference clock, the estimated error may be in 
> >>milliseconds rather than microseconds.
> > 
> > 
> > After some focused reading of newbie materials, you are right:  I am 
> > mistaking precision for accuracy.  I appreciate your good manners.
> > 
> > I've seen the estimated error go from 18.3 ms to 3.5 ms since you 
> > reply a few days back.  With a vanilla NTP configuration I suppose 
> > I'm not going to see much additional improvement.
> > 
> > My understanding is the best accuracy I might get, on a generic PC 
> > clock, is about 0.050 ms--and that's with SSP and the kernel clock.
> > 
> > Thanks for steering me towards better understanding.
> 
> The internet tends to introduce enough noise to render microsecond 
> accuracy improbable!  If you have a hardware reference clock; e.g. a GPS 
> receiver, WWV receiver, WWVB receiver, etc, you may get microsecond 
> accuracy, or not, depending on the vagaries of radio propagation.

I was thinking more of the general, ideal case -- where there might 
be multiple Stratum 1 timeservers on a LAN, and a generic PC clock.

The machine I've been working with here has one ISP-provided NTP 
server (stratum 4) within a few ms distance, but the other servers 
are from the pool so...well, frankly, I'm happily surprised I've got 
a 3.5 ms est error.

Dumb usage question on the est error:  If I have a time of N.003555 
and an estimated error of 0.003555, do that mean the actual range of 
time is N.00355 +/- 0.003555, or +/- 0.001778 ?

Thank you again for your patience dealing with a newbie for what is 
undoubtedly the 10,000th time you've explained this basic stuff.

> Selection of internet servers can have a strong influence on accuracy.
> Look for low values of delay and jitter.  The servers, if operating 
> properly, have the correct time; what you are really selecting is the 
> quality of the usual network paths between your site and the server. 
> Note the plural; the routers do their best to get the packet where it's 
> going but the route used is by no means constant.
> 
> Have a look at the ~/scripts/stats directory for some tools that will 
> allow you to study the quality of your servers and network paths. 
> peer.awk is a real help; you say something like
> gawk -f peer.awk /var/ntp/ntpstats/peerstats.20061119
> and you get a statistical analysis of the stats in that file.

Thank you very much.  Will do.




More information about the questions mailing list