Re: [Linuxptp-devel] Optimal P, I constants

[Richard C. <ric...@gm...>]

On Thu, Apr 25, 2013 at 01:08:13PM +0200, Miroslav Lichvar wrote:
> On Wed, Apr 24, 2013 at 08:51:36PM +0200, Richard Cochran wrote:
> 
> > Your simulation results are interesting. Do you propose to find PI
> > values for different sync rates in this way?
> 
> I think the simulations show that the constants can be scaled
> proportionally with the update rate.

I didn't expect that the constants may be scaled, and so a wrote a
simulator of my own (not to be compared with clknetsim, of course).

  http://linuxptp.sourceforge.net/pi_const/

This program only simulates the PI controller with an ideal,
noise-free error signal. By starting out with a time offset of 100
microseconds, the program plots the step response of the servo.

First let us look at what linuxptp does today when the actual sync
rate is faster than the default of 1 Hz.

  http://linuxptp.sourceforge.net/pi_const/plot1.png

The x-axis is the simulation time, and the y-axis is the output of the
servo in PPM. The red 2^0 line shows how the default constants cause
the servo to settle within the expected 15 second interval. If we just
use these same constants with an increased sync rate (2^ -1, -2, and
-3), the the servo still converges in the same interval, but with
over-steering. The faster the sync rate, the greater the overshooting.

If the actual sync rate is slower than the default of 1 Hz, then the
servo becomes unstable:

  http://linuxptp.sourceforge.net/pi_const/plot2.png

So the current linuxptp behavior of applying constants tuned for a 1
Hz sync to other rates does not work very well, or not at all, unless
the user provides constants tuned to the expected sync rate.

Now let us consider scaling the constants to the sync rate.

  http://linuxptp.sourceforge.net/pi_const/plot3.png

With rates over 1 Hz, the shape of convergence is improved, but you
can still see the over-steering effect. For example, the 2^-3 case
produces an initial 800 PPM adjustment for the 100 microsecond offset.

When the sync rate is slower than 1 Hz,

  http://linuxptp.sourceforge.net/pi_const/plot4.png

the result appears at first glance acceptable, but any noise in the
error signal would probably spoil things.

So my conclusion is that, although automatically scaling the constants
is an improvement over blindly applying the defaults, still scaling is
not the right solution. What we really want is a set of tuned PI
constants, one for each combination of supported sync interval and
HW/SW time stamping.

As an example, compare the using the scaled defaults with using hand
tuned constants for the 2^-3 sync interval.

  http://linuxptp.sourceforge.net/pi_const/plot5.png

Thoughts?

Thanks,
Richard