[ntp:questions] performance testing sure gps board via usb serialconverter
David J Taylor
david-taylor at blueyonder.co.uk.invalid
Tue Mar 27 09:42:04 UTC 2012
> All these parameters and permutations are confusing, particularly if
> experimenting with Windows and Linux.
> Thanks to David Taylor for encouraging me to experiment with
> interpolation, using the environment variable
> NTPD_USE_INTERP_DANGEROUS=1 to turn on and NTPD_USE_SYSTEM_CLOCK=1 to
> turn off.
> Does that apply to Linux by the way?
> In the past, I thought realtime priority for the NTPD process was
> causing problems. So, I've been experimenting with both priority and
> interpolation. Realtime is the default priority.
> The most accurate time source I have is my GPS. Internet, in my case,
> doesn't even come close. So, I'm testing min and max offsets of my
> computer's clock to the GPS polling every 8 seconds. The tests weren't
> too scientific nor too long, but I still saw some interesting results.
> 1) Interpolation ON , Above Normal Priority, + 1.00 / - 0.75 ms, Total
> Range 1.75 ms
> 2) Interpolation ON , Realtime Priority, + 0.99 / - 0.67 ms, Total
> Range 1.66 ms
> 3) Interpolation OFF, Above Normal Priority, + 1.21 / - 1.19 ms, Total
> Range 2.40 ms
> 4) Interpolation OFF, Realtime Priority, + 1.13 / - 1.02 ms, Total
> Range 2.15 ms
> Comparing lines 1 and 2, going from Above Normal to Realtime priority
> with interpolation on reduces range by .09 ms.
> Comparing lines 3 and 4, going from Above Normal to Realtime priority
> with interpolation off reduces range by .25 ms.
> Comparing lines 3 and 1, in that order, turning interpolation on at
> Above Normal priority, reduces range by .65 ms.
> Comparing lines 4 and 2, in that order, turning interpolation on at
> Realtime priority, reduces range by .49 ms.
> Conclusion, I'm leaving interpolation on, and I'm leaving the process at
> Realtime priority.
> Hopefully, someone will find this useful.
Yes, it's interesting to see the result, Ron. With the interpolation
enabled, the errors in timestamping the pseudo-serial data should be very
much reduced, but there is still the USB ~1 millisecond sampling interval
left, so interpolation may only reduced the offset by a factor of around
two at best. Good to see that you are now getting better results with
real-time than simply above normal priority - it suggests that things are
behaving as expected.
More information about the questions