[ntp:questions] 64-bit serialpps.sys despite unsigned?

Dave Hart davehart at gmail.com
Thu Mar 17 14:27:38 UTC 2011


On Thu, Mar 17, 2011 at 12:31 UTC, David J Taylor wrote:
> It's certainly working, Dave, and that's great news!  Compare the plot of PC
> Alta from 07:30 UTC this morning when the 64-bit PPS was installed:
>
>  http://www.satsignal.eu/mrtg/performance_ntp.php
>
> Having the kernel-mode has provided a major improvement in the "noise" on
> the offset, frequency and jitter plots.

Yes, it looks good, and thanks for clarifying the "before" was using
the user-mode PPS hack in the Windows port of ntpd.

> However, it doesn't seem to be working anything like as well as nominally
> the same software installed in 32-bit Windows on PC Stamsund.  Compare the
> offsets of 32-bit PC Stamsund (well under 50 microseconds) with 64-bit PC
> Alta (well over 500 microseconds and somewhat periodic).
>
> Looking in more detail, the offset as reported by NTP is varying from
> approximately -0.4ms to +0.4ms with a period of about 1.62 minutes (~97
> seconds).  That particular PC does have a very tight lock to another PC
> (server with max poll = 8 seconds), so I'll drop that back to 32s like the
> other servers and see if it makes any difference.  Changed at 12:30 UTC. I'm
> not expecting to see a difference, but one never knows!

My guess:  the key difference is the one you already identified,
Stamsund is using a precision around -20 (microseconds) while Alta is
stuck with precision -10 and the resulting minimum 0.977 ms jitter.
If you manage to get Alta's system clock to tick 64 times per seconds
instead of 500, 1000, or 2000 times per second, the interpolation code
can correctly correlate the high-resolution counter and system clock
given the 1000 Hz scheduler resolution.  Things to try would include
remove -M from ntpd options, prevent Flash or Java or media player
apps from being run.

Cheers,
Dave Hart



More information about the questions mailing list