[ntp:questions] Re: Effects of mindist setting
David L. Mills
mills at udel.edu
Tue Sep 26 21:20:38 UTC 2006
Guys,
The tinker mindist given on the cited ISC page is misleading. Each case
with each refclock, driver and kernel is different. The mindist defaults
to the minimum dispersion MINDISP (.005 s). This implies a rather close
agreement between the serial timecode signal and PPS signal. I have
observed wide variations between various refclocks with these signals,
due mostly to the FIFO in late model UART chips. The FIFO needs to be
disabled and, in some systems, the kernel driver software FIFO needs to
be disabled as well. Means for doing this vary widely between operating
systems and motherboards.
Those drivers that use the median filter in the refclock interface
usually find residual jitter well beneath the noise floor established by
the measured system precision. Peek at rackety.udel.edu and note the
jitter is usually only a few microseconds. One reason for the low PPS
jitter is that the atom driver uses the median filter and kernel PLL
discipline, but not the kernel PPS discipline. Note also that the
variation between the timecode and PPS signals is usually only a few
microseconds. Not too shabby for a very old Intel machine with FreeBSD,
GPS, parallel-port PPS and many hundreds of clients.
The experience with rackety suggests an interesting conclusion. The
residual errors with the timecode signal are almost as low as with the
PPS signal. So, why use the PPS at all? You need it to calibrate the
timecode signal with respect to the PPS signal, normally the most
accurate. The trick is to get the systematic offset (fudge time1) with
the serial driver) whittled down as far as possible and tinker the
mindist to something greater than the residual jitter. The calibrate
command in the configuration file can be used to do this automatically.
The Spectracom GPS seen by rackety is rather like a rock, but our
Spectracom WWVB receivers are very much worse due to local RFI generated
by UPS systems. The only one remaining that works at all is at my
suburban home, but even there RFI pollution has degraded accuracy and
consistency between timecode and PPS signals to several milliseconds and
does require tinker mindist. Your mileage may vary.
Dave
Harlan Stenn wrote:
>>>>In article <12hgom22phb2d2e at corp.supernews.com>, "Jeff W. Boote" <boote at internet2.edu> writes:
>
> Jeff> First - once we set the time1 value the magnitude of the dispersions
> Jeff> should be decreased. (Was the 0.010 recommendation in
> Jeff> http://ntp.isc.org/bin/view/Support/ConfiguringHPZ3801ARefclocks based
> Jeff> on any tests? Or perhaps assumptions based on Unix scheduling?)
>
> No idea - please add your experience and ideas there.
>
> H
More information about the questions
mailing list