Test farm news
Gary E. Miller
gem at rellim.com
Tue Jun 14 21:02:54 UTC 2016
On Tue, 14 Jun 2016 16:01:34 -0400
"Eric S. Raymond" <esr at thyrsus.com> wrote:
> > I'm still waiting for the good reason. You can't ask the user to
> > pick good offsets when he does not have the data for it.
> I say this because I've watched enough Pi 3 output to know that at
> steady state we're going to get well under 10ms offset and jitter
> without fudging (I just picked one Pi at random and got 0.008 0.004).
I'm not talking about the offset to the PPS, I'm talking about the offset
to the NMEA. If that is wrong it can cause major startup hiccups.
And them time to gather data for the phase 2 how to is during phase 1.
You can't start the phase 2 tutorial with: Turn on logging and come back in
> That's good enough that we can *and should* ignore all the technical
> complexities of fudging and log interpretation. We are *not addressing
> time nuts* here. We could do half an order of magnitude worse than
> this and still be good enough for a beginner audience - in fact we'd
> still be good enough to drive production WAN time service to RFC 5905
> expectations or better.
Not without PPS and only NMEA, unless the NMEA offset is right.
And sometimes even with PPS. I have seen ntpd stop using the PPS and
follow the NMEA off into the weeds.
> You are very close to this stuff. That is good because the domain
> knowledge is extremely valuable. But it can be bad because you get a
> bit obsessive about problems that are challenging to *you* (fudging,
> poor convergence time)
Not just me. chronyd market share keeps growing because people
keep recommending it over ntpd's. Convergence is a huge deal on
You need to step back and think of this as more than just an NTP
chimer. This is critical drones and other embedded devices.
We don't need to explain it, we need to fix it.
> Please, would you at least *try* to have a bit more mercy on the
Then humor me. Can you just try to look at the logs for once? How do
you know it is not valuable when you refuse to look?
> > I am seeing what I assume to be sideral effects on a 24 hour cycle.
> > Overlaid on that is a day or two of startup settling, and obvious
> > room temperature effects. I have trouble telling the apart with 4
> > days or less data.
> There, that's exactly what I mean by valuable domain knowledge. One
> of my to-dos is a white paper on NTP performance evaluation. This
> needs to be in it.
Well, I hope to fix some of the bugs first, so less needs to be explained.
If I explain how it works now it just looks ntpd look bad.
Gary E. Miller Rellim 109 NW Wilmington Ave., Suite E, Bend, OR 97703
gem at rellim.com Tel:+1 541 382 8588
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Size: 473 bytes
Desc: OpenPGP digital signature
More information about the devel