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

[Miroslav L. <mli...@re...>]

On Wed, Apr 24, 2013 at 08:51:36PM +0200, Richard Cochran wrote:
> On Tue, Apr 23, 2013 at 03:45:12PM +0200, Miroslav Lichvar wrote:
>  
> > This one is with 100us jitter, 10ppb/s wander and 1s interval. The
> > best I constant now becomes impractical as it will take hours
> > for the servo to converge.
> 
> What clock has 100us jitter? Are you trying to model time stamp error,
> like in SW time stamping?

Yes, 100 us was used to simulate a bad SW time stamping.
> 
> > I was wondering how useful it would be to have a new servo parameter
> > for the Allan intercept and set the P, I constants automatically
> > according the intercept and measured clock update interval. What do
> > you think?
> 
> I would like to have a table of PI constants, with one pair for each
> (reasonable) sync rate. I don't think these need to be tuned to
> particular oscillator characteristics, but rather just for "typical"
> hardware. Most people will not know what they have and will not be
> able to measure the Allan deviation of their clock.

What exactly is a typical hardware? :) I think we could create a table
with safe P, I constants for different rates. We could also check them
at runtime and print a warning or clamp them if they are in the unsafe
range.

I'd prefer if they were scaled automatically as the update rate can
change in the runtime and the user now has to pick the constants so
they are safe even with the longest expected sync interval.

> 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.

Thanks,

-- 
Miroslav Lichvar