[ntp:questions] Is dispersion > jitter in all situations

unruh unruh at wormhole.physics.ubc.ca
Tue Jan 5 19:19:37 UTC 2010


On 2010-01-05, John Hasler <jhasler at newsguy.com> wrote:
> Bill Unruh writes:
>> However, the only way you can really know is to get a GPS with PPS,
>> hook it up to your machine and measure the difference in time between
>> the Stratum 3 set clock and the GPS time.
>
> I think that what he wants is the "expected time" (he has that: the time
> Ntpd reports) and a 95% confidence interval for it.  That is, he wants
> to be able to say "The time is within +- 4.2usec of 2300UTC with 95%
> confidence".

Unfortunately he cannot get that. the random part he can get (jitter)
but the systematic part he cannot. The dispersion is a very very
conservative estimate of the systematic part, but it is in almost all
situations far far larger than the true 95% confidence interval. As I
said the only way to estimate the systematics is by having a local gps
pps clock to compare it with, but that is obviously overkill, since then you
would just use the gps to give the time. 

for example, I have a machine with a gps pps and with another source
being a stratum 1 server 2000km away. The round trip time is 45ms, which
would give a "dispersion" component of 22.5ms. But actually the offset of
that source from the local gps time is only about +-.1ms.Thus the
estimate -- 1/2 the roundtrip -- of the systematic error is out by about
a factor of 200.  Thus the estimate -- 1/2 the
roundtrip -- of the systematic error is out by about a factor of 200.


As I said, it is possible that all outgoing ntp requests go via a 1Gb
ethernet, and all return packets go via a 300baud modem. In that case
the estimate of 1/2 the round trip would be a good estimate of the
systematic error. Unfortunately there is absolutely nothing ntp can do
on its own to figure that out. 





More information about the questions mailing list