HI

NTP running on the same sort of network will deliver roughly the same performance.
There is very little asymmetry and the “source” is properly done. The net result
is roughly microsecond timing on a simple LAN.

With proper time stamping throughout and good hardware and a high enough packet
rate you can get to the nanoseconds level with 1588.

Bob

So much to absorb and learn from what people have responded with.

Thanks all!

On 01/12/2016 12:01 PM, Chris Albertson wrote:

That's the take I have on it.

That's good to know about PPS, 'cause the computer's load while polling
hosts over the internet is getting me horrible variance in offsets.

Looks like fun, but way better than I need. More than I can afford too.

O.K.. I'm pretty sure I don't understand that, but the issue I found was
not that different hosts offsets varied from one another, but that the
offset reported by "a" host would jump around. And when the load on the
computer was changing, low-to-high or high-to-low, several hosts',
sometimes all hosts, offsets jump around, with that multi-host variance
continuing for some minutes after the computer was running at the new
load. This settled down a little when I turned core parking off, but
only a little. I've attached a sample of the offsets I'm getting to show
this variance. Oddly, a sustained higher load would often settle the
variance and give the most consistent results: one such period is
between poll 6 & 10 on the "test run N24" attachment, and the graph
shows the offset slope and hosts' offset variances as the load moves
from heavy to medium and then light.

After giving up on SMAs to tame individual host's offsets, to get a
usable offset from the reported offsets I implemented a cascade of
filters: applying factors on standard deviation to implement truncation
means to remove outliers, then winsorizing means, with independent bias
factors applied to selection and winsorization. This all worked rather
well once I tuned the filters' parameters. Then the variance got worse,
so I added some increasing attenuation on increasing corrected offset
values and that made my corrected offset usable within the tolerance I
needed, until from a certain date the reported offsets went all over the
map.

But we shouldn't go down this road, other than curiosity (and I wish I
had the time to explore the why), as going to a separate machine as an
NTP host removes all of those types of issues. And I don't have to grow
my own code...

Exactly.
Any module that can get a usable GPS signal can discipline time and be
delivered over my local Ethernet to better than I need.

Thanks for those. Good to know.

I believe the location issues narrows it down to the MAX-M8Q or the
NEO-M8T.

Both have great sensitivity, but their firmware varies to address
intent. The M8Q can be explicitly set to Dynamic-Mode-Stationary (as it
should go to automatically with an unmoving antenna) and the M8T will
set there as it moves to focusing on a better time solution after
establishing a location fix. In comparing their product descriptions,
the M8T seems the better choice while sitting still for obtaining usable
results in questionable locations, but - speculating - that wording may
be marketing wording in response to prior issues with earlier T series
modules. And so far, I've not found any accounts of first hand
experience with a M8T.

The other issue is what breakout board the modules are available on.

• With the M8Q, there's hats or boards that can connect direct to a Pi
  or such, but lack protection with supply voltage or outputs if I want to
  feed them to another computer.
• And the M8T is available on a board that provides power regulation and
  some protection, so that should be able to feed NEMA & PPS to any
  suitable computer without risk.
• And I found a board that accepts GPS modules-on-breakout, protects
  them, and can feed any computer, but the breakout boards with M8Q or M8T
  have pins don't match the header. A small custom cable would fix that.

But without trying them, I can't Know which will be best by my location.
Or if my location means one will work and the other won't...

(waiting to do that GPS sat test at my location...)

Hi

For an “antenna challenged” location, the T is the better choice. It is simply an update (as the M8Q) of the earlier uBlox parts. The function is very similar to the earlier parts. You nail down the antenna location (like with duct tape) and put the module in survey mode. Eventually it completes the survey and you save that location. That location is then used as part of the fixed location setup. This eliminates any need to ever do a survey again. The reason you want the fixed location operating mode is that it will work with a single satellite. You don’t need the accuracy, but you do want to eliminate dropouts. The module should be set to only put out a PPS when it has a valid timing solution. You very much do not want it to simply put one out that is based on the internal oscillator on the module.

Bob

Hi Lewis,

Here is a sample data-point related to processor load, on the RPI 2.
Stepping from Idle to full load on 4 cores resulted in a temp rise near the
XO of approximately 14 degC, and correspondingly the XO shifted 3.6 PPM.

On Sat, Dec 3, 2016 at 10:29 AM, MLewis mlewis000@rogers.com wrote:

On 03/12/2016 12:33 PM, Bob Camp wrote:

Thank you!

That is exactly what I "thought" I was getting from reading various
sources, but I wasn't confident I was getting it right.

I think:

• There's GPS Time and with the offset (currently 17 seconds?) in the
  GPS location message we get UTC Time.
• The GPS module receives the satellite's GPST, offset, and hence UTC
  Time, and its location fix allows for correcting for the message delay.
• A better fix, and that correction is more accurate.
• With a number of satellites reporting, that correction is more accurate.

If the module has a good fix and a good solution from several satellites
and is providing me with a local UTC Time that is close to true UTC, but
then conditions degrade to a single satellite:

• will the module be able to maintain a quality correction to true UTC
  Time?
• or perhaps I should be asking: what is the precision from a GPS
  module's best time solution vs. that from a single satellite.
• or is that irrelevant given my rather trivial goal of 1 ms precision?

And here I was thinking that internal oscillator was going to save me by
continuing to provide PPS from a recently disciplined TCXO.

If I have a good sat solution and am getting PPS for some hours, then it
loses all satellites,
- is the internal TCXO sufficient to trust for a number of minutes?
- if so, how would I determine how much error for how many minutes?
- or do I just have to set an alarm, let NTP fall back to the
internet hosts, and monitor my data clusters for drift?

I'd guess this is related to the Best Practices I saw that said that for
time critical applications, as a minimum: use a local GPS time source,
but have NTP polling four or more NTP hosts (non GPS based) through the
internet as backup if the GPS system goes out.
Or in my case, simply if I can't track a sat.

Thanks,

Michael

Hi

Not quite correct. Until the module has a proper fix it does not have useful time. The delay from the sats is long enough that for most purposes there is no value
in putting out raw time. You could re-write the firmware from scratch to do it a different way, but that’s the way it works now on all these modules.

If it has a properly surveyed location (and thus can calculate the delay in getting the signal from the sat) yes.

With a proper survey, it’s in the 10’s of ns. It degrades roughly 1.5 ns / meter for typical survey errors. The survey induced error tends to
show up as a long term (hours) drift in the output. It is slow enough that NTP will try to track it. The 10’s of ns basic error is related to things
like ionosphere corrections and that also can be a long term drift. Both errors may go up a bit with a single sat. How much depends a lot on
the geometry involved both with the single sat and the group you replace it with. It’s possible (but unlikely) to have a single sat in a “perfect” location give
you a better solution than a group of sats in a really crummy location ….. This is yet another reason for wanting a full sky view. Seeing several
sats at a crummy angle may still be a less than ideal thing.

On a LAN, you should be aiming for < 10 us in terms of the time input to the main server. Your “delivered time” budget needs to allow for a lot of
different errors that all add up. The idea is to make the time input error insignificant compared to the rest. If your time input error is 1 ms, and you
are after 1 ms, the rest of the errors would have to be zero. Depending on your setup, they may well be a significant chunk of 1 ms.

The TCXO in the module is not disciplined. A GPSDO is a device that disciplines an oscillator. These modules are not GPSDO’s.
If you have a GPSDO, then “holdover” is what takes care of the PPS output when you loose GPS.

It may slew at a couple microseconds per second. That’s much worse than just ignoring it.

Measure it under your conditions. It likely will be a bit different each time.

That is why the PPS drops out. NTP then knows to ignore it and go do it’s own thing. That is a much better solution than
tracking a slewing source.

For a variety of reasons, you want multiple sources with NTP. Unless it has at least three reasonable time sources, it gets into all sorts of strange behavior.
Setting up with four to six sources is a pretty good idea. With 4, if one drops out you are at the limit of three. Six gives you a bit more margin.

The biggest issue with NTP is offsets. This is why they created PTP.  If you run NTP for a long time with a good local reference and have a stable connection
to a fixed set of remote servers … NTP can work out what the offsets are. That magic depends on the local reference being there all the time along with a
few other things.

Bob

On Sun, 4 Dec 2016 09:46:45 -0500
Bob Camp kb8tq@n1k.org wrote:

Frühauf proposed using satellite dishes on WAAS satellites for
improved timing accuracy/stability[1]. Unfortunately, I have not
seen any good analysis of the achieved performance.

		Attila Kinali

[1] "WAAS for Telecom applications", Hugo Frühauf, 2003, updated 2011
http://hugofruehauf.com/pdf/24-WAAS_for_Telecom_2003-upd_2011.pdf

--
It is upon moral qualities that a society is ultimately founded. All
the prosperity and technological sophistication in the world is of no
use without that foundation.
-- Miss Matheson, The Diamond Age, Neil Stephenson

Hi

Simple example:

Single sat straight overhead, well compensated by broadcast ionosphere. All survey error in the X direction.

Multiple sats all clustered at the horizon. Cluster located on the +X axis. Ionosphere not well modeled over the
much longer path.

Yes this is a bit contrived. It’s also what can happen in an “overhanging balcony” sort of antenna location.  I can
also make it happen pretty easily from a USB stick indoors and locating “just right” relative to the windows.

Bob