[ntp:questions] NTP constant delay ?

David Woolley david at ex.djwhome.demon.invalid
Thu Dec 27 10:21:49 UTC 2012


Ran Shalit wrote:

> I intend to use SNTP protocol with 2 machines connected by LAN.
> According to decumentatio about SNTP it sys that "SNTP provides
> accuracy typically within 100 ms".

That's pessimistic and is based, in part, on SNTP not defining how one 
gets from the time measurements to the local clock settings. or how 
often one makes them.

> I would like to ask please what can we expect in our configuration
> will the delay be constant within 100msec, meaning that we cn measure
> it constant delay, or will it drift such that in one meanute is will
> be delay X, and in other meanutes totally other delay within 100msec.

On a LAN, the measurement error will vary, but not by anything like 
100ms.  Over a WAN, it is possible for it to vary by up to a second.

> If the delay is constant, I assume that we can always delete this
> delay constant X and get a time synchronization around zero. Is that a
> right assumptuion ?

Even SNTP corrects for round trip delay, so the remaining delay errors 
will be due to asymmetry.  The systematic part of errors will affect a 
full NTP system as well, and people don't normally worry about them. The 
biggest problem in correcting them is measuring them.  If you have some 
means to measure them, you would be better off using that for time 
synchronisation.

In any case, if you want better handling of random variations, there is 
really no reason not to use a full NTP implementation.

Errors from typical SNTP systems arise from:

- stepping the time to match even a bad measurement;
- not correcting for clock drift between measurments;
- not being able to detect and ignore a source with a faulty clock.

If you correct for all of these, you have something that is pretty close 
to being a full NTP.

One other thing.  If you are looking for hard limits on the error, the 
only absolute guarantees that can be given are much larger than errors 
than the typical ones.



More information about the questions mailing list