[ntp:questions] Re: NTP precision

David Woolley david at djwhome.demon.co.uk
Mon Sep 26 06:39:23 UTC 2005


In article <43373166.90201 at cag.zko.hp.com>,
Tom Smith <smith at cag.zko.hp.com> wrote:

>     "true" time = systemtime + offset +- (rootdispersion + some other error)

You also need to add half the root delay.  I believe that the actual
figure is calculated by ntpd and used to reject times with too large an
error band.

Also note that, if things are working properly, the error term will be 
larger than the offset.

> Where "some other error" may or may not be the "jitter" of the root
> server (the stratum 0 server/reference clock), and where "true time"
> would have to be interpreted as meaning the "true time as perceived

I need to delve into the code, in default of an NTP4 specification to
be sure, but I think that that is accounted for in the reported 
root dispersion.  Generally, this question needs to be answered from
the code, rather than black box guesses.

> by the root server" - probably very small compared to the other
> error components. The "offset" is at any time the remaining difference
> to be worked off to agree with the consensus time among all the peers,

I don't think that is quite true, but I don't have time to check
properly.

> whereas the "rootdispersion" includes the aggregate contribution of delays

rootdispersion doesn't include delay.  The primary component is the assumed
worst case error due to clock frequency error integrated over the time
between when the ultimate reference clock was read and the time was
measured at the leaf node.  On NTP 4, I seem to rememher it includes the
jitter and clock resolution errors, but, again one really ought to look
at the code.

> only through the currently selected peer chain back to its root. Not quite

There isn't a peer chain.  There is a client server chain.

Also note that all this assumes that every node in the chain is well 
behaved.  Lost interrupts will almost certainly invalidate things, for
example.  Also, delay, in particular, will often be known to be more
symmetric than is assumed by using half of it as an error term.




More information about the questions mailing list