[ntp:questions] Re: Does NTP model clock skew?

David Woolley david at djwhome.demon.co.uk
Sun Feb 13 08:37:36 UTC 2005

In article <1108246334.523244.237070 at g14g2000cwa.googlegroups.com>,
affanahmed at gmail.com wrote:

> I am trying to currently understand exactly how does NTP cater for
> clock skewing at individual nodes in the network. From my current
> understanding, it does so based on just statistical measurments of
> many,many offset calculations 

It's origins are in control engineering, not in statistics.  Look
in the literature for proportional integral controllers.  These (or
the more general proportional integral differential form) are commodity
items in industrial process control.

>                               made through a two way exchange between
> the NTP server (another question.. are the server and client supposed

The innformation flow is generally one way (peer mode complicates this, but
even then the primary information flow is in one direction for extended
periods - indefinitely if things are all working and stable).  I suspect
you are thinking that NTP generates a consensus time; it doesn't, it 
distributes time from a single source, UTC time.  The two way interaction,
which is used to estimate propagation delay, to tell the server who wants
time and for some of the authentication protocols.

> to be on the same network i.e. one hop away?)- and it changes something
> in oscillator PLL. But it doesnt do anything like what e.g. RBS does -
> meaining that it does not form a line which can be used to say.. yes
> this is a value of skew that i have (based on some line drawn to best
> fit the datapoints).

The processing in NTP has its history in analogue hardware designs which
have very limited memory (however these techniques are also intrinsically
fast to compute as well).  There are people (e.g. Nick McClaren of 
Cambridge University, who has a statistics background) who argue that it
would be better to use statistical methods that look at a number of recent
data point each time and try to make the best estimate of the skew and
offset based on those (taking into account the effects of the corrections
already applied.  My statistics isn't all that strong, but I do have
a gut feeling that there may be some value in that.  However, part of
the error term is the oscillator itself, and there is a definite limit
of how far back you can usefully go (characterised by the Allen Intercept

> I hope this isnt a wrong place to ask the question, and I hope I have

You are expected to make reasonable endeavours to find the right place
before posting.  As such, you should only apologise if you believe
you are in the wrong place, but there is no right place.  This is the
right place for this question.

More information about the questions mailing list