[ntp:questions] NTP tunning for OWD measurements

unruh unruh at invalid.ca
Sat Oct 27 16:14:00 UTC 2012

On 2012-10-27, pret3nder at gmail.com <pret3nder at gmail.com> wrote:
> Hello all, and thank you again for your valuable insight 
> on this matter, it's been a great help to me.
> I'll now try to reply to all your comments, please excuse 
> me if I missed a particular comment.
> Indeed I was wrong about the tunnel reference, that won't
> correct the fact that a delay asymmetry will always exist
> in the path. I'm sorry for that remark.
> I already have ntpd setup, collecting statistics and I'm using
> munin to graph some of those. Currently I'm only using
> S1 <-> X1 one-way delay measurements to simplify.
> Both S1 and X1 use the same stratum 1 time server, but
> right now S1 is down because of a problem on the
> datacenter. 

As I pointed out, a one way delay to the stratum 1 server will show up
as a one delay on your lines. 

> Nevertheless, here goes a link for the stats on 
> the X1 server: http://imgur.com/a/Oer6z
> The spikes show there are related to a network problem that
> as been identified and fixed.
> I'll also follow unruh's advice to plot the delay vs offset.
> We are aware that we are sitting on 2nd class seats, and 
> perfectly conscious that GPS on both ends of the link 
> would give us the most accurate results. Nevertheless, I want
> to be able to know what's the maximum error that will
> affect my owamp measurements, using only ntpd synching
> to stratum 1 servers on our network.
> I'm also aware of the need to take the results with a grain of
> salt, because there are also errors related to the OS overhead.
> Since it's a fiber-optic network, we would like to have
> microsecond accuracy, accounting of course the errors discussed
> previously. I will try to get GPS signals at least in one of the links 
> (on both ends) so I can compare results, and draw conclusions 
> from there.

I am sorry. Even gps will have trouble giving single microsecond
accuracy, and it is totally and utterly hopeless to hope for single
microsecond accuracy using a stratum 1 server. The link to that server
will dominate the error budget. If that server is on the same physical
network as your systems, you might hope for 10s of ns. If it is on the
net. it will by 100s of ns to mseconds.

gps has the problem of getting the 10s of ns accuracy of the gps into
the computer, since the interrupt processing of the computer takes up a
variable amount of time, from 1-10us with an average variation of about
3 ns.(my measurements).  and it is all delay-- you computer will never process the
interrupt before it occurs.

> About plotting (delay+offset)/2 and (delay-offset)/2, I suspect
> the interesting data would show in the condition the offset would
> be negative on one way, and positive on the reserve way?

That is exactly the same as plotting delay vs offset except that the
axes are rotated by 45 degrees and rescaled. If you cannot rotate the
paper by 45 degrees on your table (or in your mind) you should not be
doing the project.

> David Woolley is right, the project is about making the measurements,
> although I don't like the idea of doing them blindly, without
> knowing how to interpret the results. I won't have time to
> investigate the deeps of NTP, but I do want to know the basics,
> as that's one of the objectives of my thesis.

Read through the RFC You will not understand everything, but when
questions come up, you will know where to look.

> Kind Regards,
> Pedro
> On Saturday, October 27, 2012 3:56:22 PM UTC+1, John Hasler wrote:
>> Terje Mathisen writes:
>> > The crux is this: You _MUST_ have a path to an absolute time
>> > reference, at both ends, which is totally independent of the path you
>> > are trying to measure!
>> And a tunnel is _not_ an independent path.
>> -- 
>> John Hasler 
>> jhasler at newsguy.com
>> Dancing Horse Hill
>> Elmwood, WI USA

More information about the questions mailing list