[ntp:questions] Re: [Bug 177] Clock stepping messes up frequency. (fwd)

Wolfgang S. Rupprecht wolfgang+gnus20031111T075953 at dailyplanet.dontspam.wsrcc.com
Tue Nov 11 16:41:49 UTC 2003

Ulrich Windl <Ulrich.Windl at RZ.Uni-Regensburg.DE> writes:
> Stupid as I am, I'd say that ntpd should only increase the polling
> interval if it observes the clock being stable enough. A similar thing
> is true for the PPS. IMHO the hardwired tolerance is far too big,
> making the algorithms increase the polling interval. Dave will hate me
> for saying that...

I saw some really weird dynamics when I first fired up my M12 Oncore
GPS reference.  I didn't want to artificially stack the cards in favor
of my server by adding a "prefer" to it's config line.  The startup
went this way: The oncore woke up first and was selected as the
preferred server.  Once the half-dozen outside servers became
reachable for a while they "out-voted" the more correct and much lower
delay and much lower jitter oncore.  The lapse in poor judgment was
exacerbated by the fact that the polling interval then went way up (so
all future decisions were now made in "slow motion") and the GPS clock
was deemed to have "high jitter" by virtue of it's high offset from
the assumed time.  This high jitter value would fade down to zero
after a long time, but every time the a different external server was
chosen as the one to follow, the measured jitter would bump up again.
It took a long time for the GPS to be eligible for selection again.

Yes, I could (and did) slap a "prefer" into the server config line,
but that seems like a bit of a cop-out.  I believe the underlying
problem was that "jitter" is being measured relative to the very clock
that is being adjusted by ntp.  Any time ntp cranks the time,
everybody gets a boatload of jitter added to their "how well am I
driving" column.  This effects the low jitter guys more (at least
proportionally speaking) than the high jitter guys.  After a time
movement everyone's jitter is more-or-less the same large awful
number.  The effect is that after a time movement, the low-jitter
sources don't really get any credit for being low jitter.

I guess I need to poke at the old ntp research papers a bit more.  I
can't believe I'm the first person to notice this.

Wolfgang S. Rupprecht 		     http://www.wsrcc.com/wolfgang/
           The From: address is valid.  Don't mess with it.

More information about the questions mailing list