On Wed, 16 Nov 2011, Miroslav Lichvar wrote:
On Tue, Nov 15, 2011 at 02:25:09PM -0800, Bill Unruh wrote:
> > http://mlichvar.fedorapeople.org/tmp/chrony_corr2.png
>
> Are those two separate realisations of the random process which perturb
> the
> offset and rate, or are they the same? It might be an idea to use
> exactly the
> same one for both so one can see better what the difference is. (Yes, I
> recognize that the key difference is that the green curve is smoother
> than the
> red curve which is a good thing, but other details are harder to
> separate out.
Running two clients in exactly the same conditions is a bit tricky do
to in the simulator. The NTP server and NTP client run in parallel, as
the clknetsim server uses the same system random() for all clients and
they pick numbers from the sequence in a random order, it's very
unlikely to have a deterministic simulation. The sequences for the
clock frequency and the network delays in each direction have to be
pregenerated to files and clknetsim configured to read them instead of
using the random generator.
Here is a new simulation with the default corrtimeratio and the same
random sequences used for both NTP clients:
http://mlichvar.fedorapeople.org/tmp/chrony_corr3.png
Ah, OK, it maily smooths out the very short term jumps -- ie, as if a low
pass
filter had been run over the data.
I assume that if you used a higher corrtimeratio you would get further
smoothing. The question is whether with that same data you would also begin
to
get evidence of oscillations as well with that same dataset
Here is the uncorrected clock frequency if you want to see how it
correlates to the offset oscillations:
http://mlichvar.fedorapeople.org/tmp/chrony_corr3_clk.png