[ntp:questions] RE: NTP precision

Leandro Pfleger de Aguiar leandro at bry.com.br
Fri Oct 21 15:18:11 UTC 2005


> > > Leandro Pfleger de Aguiar wrote:
> > > > Hi
> > > >
> > > >     I4m using NTP to timestamp digital events in a specifit
> > > application. On my application i need to know what is the exact
> > > estimated error that my timestamp marks should consider. Some
> > > peaple frequently use offset to mean what is the real estimated
> > > difference between local and ref clock. Should i consider other
> > > parameters like jitter,  precision and accuracy manually ?
> > > Remember: i need to say how mutch can my timestamp be wrong.
> > > >
> > > > Tanks Again !
> > >
> > > Interesting question. I think the data needed, taking into account
> > > all of the servers, is in the results of "ntpq -c rl 0":
> > > ^^^^^^^^^^^^                    ^^^^^^^^^^^^
> >
> > Hello, everyone:
> >
> > Talking about NTP accuracy, I'd like to say something because I have
done
> > experiments on this.
> >
> > This is my test environment:
> > Take one machine as NTP server and use its local clock. There are other
> two
> > machines to be synchronized to the server. All machines can be
interrupted
> > by external digital events. As soon as the external event comes, the
> > interrupt routine will record its own time. Just compare the time stamps
> of
> > server and client, I can find out NTP performance.
> >
> > If a low-level oscillator(such as 100PPM) is installed on NTP client,
use
> > less shorter poll interval, such as 500us accuracy with poll interval
> > 16-32s. If the default poll interval 64-1024s, I've got the accuracy
2ms.
> >
> > Network delay and network jitter does influence NTP accuracy very much.
I
> > know this by an experiment on LAN.
> >
> > All machines that you are using are capable of running ntpd in case that
> you
> > donnot have 486 or 586 PCs. CPU speed is not a problem. But donnot make
> your
> > machine run with very very heavy burden.
> >
>

 Tanks Eric, but that´s not the point.

     With a lots of measures in your environment maybe you could say that
 your "medium error" stay around 2ms.
 But what i was intend to know is how wrong (in statistical terms) values
 taken from "ntpq -c rl 0" can be wrong.
 In other words, i need to know what is the "maximum" error that can i do
 when i say that the time taken from
 a ntpd disciplined machine is "some value".

 Tanks anyway !






More information about the questions mailing list