[time-nuts] Comparing the BeagleBone Black & Raspberry Pi asNTP servers
attila at kinali.ch
Wed Mar 25 16:51:44 EDT 2015
On Wed, 25 Mar 2015 09:46:05 -0700
Chris Albertson <albertson.chris at gmail.com> wrote:
> So if NTP always compensates for network delay why do you get improved
> performance with less delay? That is because what messes up NTP is
> uncertainly in the delay and likely it's the case that reducing the delay
> also reduces the standard deviation of the delay. The other thing the
> messes up NTP is its assumption that the delay is symmetric (that the
> one-way delay = one half the round trip delay) I think reducing the round
> trip time also reduces error in this assumption.
There is another assumption, that most people (especially engineers)
do (and NTP is forced to do), which does not always hold true: that
measurement noise is mean-free and uncorrelated. Especially in this
case, USB delay is definitly not mean-free if you queue up packages
(i would go so far as to say, that USB induced delay is never mean-free)
and has a non-zero auto- and cross-correlation, mediated trough
the USB clock, which is synced to the system clock's source (they
are derived from the same crystal).
AFAIK, and without checking, I think that NTP also makes the assumption,
that the noise is ergodic and time-invariant during measurement.
Again, in the case of USB (and to some extend all network based
communication systems), this does not hold true.
< _av500_> phd is easy
< _av500_> getting dsl is hard
More information about the time-nuts