[ntp:questions] Ignore one server except in extreme conditions?

Dave Hart hart at ntp.org
Thu Nov 17 00:09:48 UTC 2011

On Wed, Nov 16, 2011 at 23:09, A C <agcarver+ntp at acarver.net> wrote:
> Yes, if I disable the NMEA source with a "noselect" and leave the Internet
> servers and the PPS clock (22) running, my overall system offset drops and
> holds at a few tens of microseconds.  If I leave it in, the system offset
> wanders around.  The magnitude of the wander appears to correlate with the
> magnitude of the NMEA offset.  For very large NMEA offsets (sometimes
> exceeding +/- 50ms) the system itself starts to drift away to large ms
> offsets.  Overall I seem to get better performance without the NMEA driver
> contributing than with it included hence the desire to make it accessible
> only when all the other Internet sources fail (but GPS is still working).

Which version of ntpd are you seeing this with?  Based on


I would expect the PPS to control the clock on its own, when it's
active.  The combine algorithm is run, but the averaged offset is then
replaced with the PPS offset which actually drives the clock.  I
suspect you're running 4.2.6 or older and it may not be following the
current mitigation rules.

> Both PPS and NMEA are coming from the same physical GPS just using two
> serial ports, one for the PPS and one for the serial data (this split is
> required due to serial port driver limitations).

You don't have to use two separate drivers for NMEA and PPS, even with
the signals coming in on different serial ports.  At least in 4.2.7,
the NMEA driver tries /dev/gpsppsX for PPSAPI before falling back to
using the same port as for NMEA, /dev/gpsX.  In this configuration,
you don't need to mark any peers as prefer, which I much prefer
compared to using ATOM and marking all your network peers and NMEA
prefer, because prefer has profound effects on mitigation which are
tricky to wrap your head around.

I encourage you to try a recent 4.2.7 ntpd with only the NMEA driver
with its PPS handling enabled

Dave Hart

More information about the questions mailing list