[ntp:questions] refclock on Windows graphs

Dave Hart davehart at gmail.com
Wed Mar 18 09:28:51 UTC 2009


On Mar 18, 07:28, David J Taylor wrote:
> Not far more constant, no.  On the FreeBSD/Pentium-133 system over the
> last 72 hours the frequncy varied between about -0.6ppm to +0.4ppm from
> the mean, and on the Windows XP system from about -1.4ppm to +1.1ppm from
> the mean.  So about 1ppm variation as opposed to 2.5ppm, about 40% of the
> frequency variation.
>
> As far as I can tell, the NTP offset is more a function of the control
> loop parameters than of the actual hardware (with ~150us offsets and yet
> only 2.5us jitter).

The 150us offsets are while slewing to try to catch up with the 3ppm
frequency shift as the heat comes on.  I have to imagine a Windows PC
with a more stable oscillator (lower temperature coefficient to
frequency) would do better, though I'm not deluding myself that it
would match the performance of FreeBSD, with its advantage of knowing
the precise relationship between the counter and the system clock.

> There is another issue - version numbers.  The FreeBSD runs 4.2.0-a from
> May 2005, the Windows 4.2.4.p6 from 2009.  I don't know whether there may
> have been internal changes affecting the offset compared to the jitter.

I have no familiarity with 4.2.0 but keep in mind it's even older than
2005, that's the FreeBSD release date and they were not keeping up wth
NTPv4 builds very closely then.  From my perspective comparing its
performance to the new Windows stuff, it would be awfully nice to get
4.2.4p6 on your FreeBSD box.  There are plenty of helpful FreeBSD
wizards hanging around here who could help with any issues you have
building the reference implementation and installing it in place of
the system-provided one on that FreeBSD box.  At least I think they
would, but given your treacherous discussion of moving away from
FreeBSD in favor of Windows maybe you'd get the cold shoulder ;)

> I have checked for a temperature monitoring SNMP plug-in, but not found a
> compatible one as yet.

SNMP is ideal, but anything that could read a temperature that had a
high degree of correlation with your osciallator's crystal temperature
would be enough to build upon.  I'm afraid anything built in to your
hardware (like CPU core temp) is not well-enough correlated the
crystal temp.

Cheers,
Dave Hart




More information about the questions mailing list