[ntp:questions] improving ntpd performance on Windows
davehart at gmail.com
Wed Mar 4 22:34:13 UTC 2009
On Mar 4, at 19:44 I wrote:
> I released another test version today onhttp://davehart.net/ntp/refclock/
> Contrary to my previous claim, my recent versions have not been using
> the observed performance counter frequency instead of the nominal one
> from QueryPerformanceFrequency. A range-gate in my code to observe
> the performance counter was preventing any changes silently. This had
> apparently been the case since I switched to the new interpolation
> strategy. Today's version re-enables the tuning and seems to have
> knocked a bit more jitter out of the interpolated time in my case.
> The nominal frequency is used for the first 36 minutes as four 512
> second samples of the observed counter frequency are collected, then
> the observed value is switched into use and updated every 512 seconds
> thereafter. This resulted in cutting short-term jitter in half
> (observed as a vertical smearing of the offset graph line).
That's a one-hour before and after snapshot of my refclock test
machine with the transition to using the observed counter frequency in
the middle at 21:54. I hope the coincident 90 usec step is a fluke.
The short-term jitter before and after is encouraging, looks like from
over 20 usec to under 10.
More information about the questions