[LinuxPPS] Experience with current linuxPPS kernel patch.

[Miroslav Lichvar]

On Thu, Mar 04, 2010 at 11:58:58AM -0700, clemens at dwf.com wrote:
> One thing it brings to mind tho.
> Your plot is symmetric on the + and - sides.  
> I do an offset vs time plot, and it shows the same symmetric fuzz around a 
> central
> value.  Now the central value can move up and down, especially at a reboot, but
> also due to room temperature (I assume).  I guess the question with your plot 
> is,
> are we seeing fuzz around a stable central value, or are we seeing the central 
> value moving back and forth?

Good question. I think it depends on how good is NTP (or chrony in my
case) in tracking the temperature changes.

Here is another plot from samples collected over the last weekend,
temperature was quite stable and the frequency was moving only in
0.3ppm range. It includes the samples from GPS and also the local
clock offsets as corrected in 16 seconds intervals (poll 4), which is
basically a mean from last 128-1024 samples. 

http://fedorapeople.org/~mlichvar/tmp/18xlvc_errdist2.png

The local clock dispersion is mostly below 200 ns, so I'd say
it didn't expand the PPS distribution much.

> The interesting thing here, is that when I do the same plot on data from ntp
> on FreeBSD, the fuzz is ONE SIDED.  If you think about it, the one-sided 
> 'answer' is
> the correct one if you believe that what you are seeing is delays in answering 
> interrupts.  There is no way to answer an interrupt early, but plenty of ways 
> to
> answer it late.  So, the symmetry that you and I see on Linux puzzles me.

I see an asymmetry in offset/time plot, one side has spikes about 5-20
us. It's interesting that they seem to happen more frequently when the
machine is idle.

Beside that asymmetry, is the dispersion you see on FreeBSD smaller?

-- 
Miroslav Lichvar