Re: [chrony-dev] [GIT] chrony/chrony.git branch, master, updated. 1.25-pre1-18-g20a4340

[Bill Unruh <unruh@xxxxxxxxxxxxxx>]

On Wed, 13 Apr 2011, Miroslav Lichvar wrote:

> On Wed, Apr 13, 2011 at 01:30:28PM +0100, Ed W wrote:
> > On 13/04/2011 12:10, Miroslav Lichvar wrote:
> > > It seems the problem is that in the weights calculation is used
> > > weighted variance, which can create the positive feedback. Using
> > > unweighted variance instead should fix it nicely.
> >
> > I guess I meant simply sanitise the range of values.
> >
> > Presumably there is a numerical speed advantage in using weighted
> > values? Therefore use weighting, but only up to a constrained level.
> > This will lead to suboptimal convergence in some edge cases, but faster
> > convergence elsewhere?
>
> In the case I've seen all weights except the last one would end up at
> that limit. The limit would have to be high enough to allow giving
> very small weight to samples on a wireless network where the delays
> can vary by say 3 orders of magnitude, which in the problematic case
> would mean that the regression line would go through the last sample.

Hm, whether one used normalised or unnormalised weights should not make a
difference to the fit, or the variance in the slope. 
I looked at that code a few years ago, but have forgotten it by now. Will have
to look again.

> Anyway, using the unweighted variance fixes the problem, but it
> changes the clknetsim results. I'll see if I can get close to the
> previous behavior by tweaking the SD_TO_DIST_RATIO constant.

--
William G. Unruh   |  Canadian Institute for|     Tel: +1(604)822-3273
Physics&Astronomy  |     Advanced Research  |     Fax: +1(604)822-5324
UBC, Vancouver,BC  |   Program in Cosmology |     unruh@xxxxxxxxxxxxxx
Canada V6T 1Z1     |      and Gravity       |  www.theory.physics.ubc.ca/

---
To unsubscribe email chrony-dev-request@xxxxxxxxxxxxxxxxxxxx with "unsubscribe" in the subject.
For help email chrony-dev-request@xxxxxxxxxxxxxxxxxxxx with "help" in the subject.
Trouble?  Email listmaster@xxxxxxxxxxxxxxxxxxxx.