Re: [gpsd-dev] Mysteriously vanishing bugs don't make me happy

[Gary E. Miller]

Yo Greg!

On Mon, 04 Nov 2013 21:05:42 -0500
Greg Troxel <address@hidden> wrote:

> > Yes.  I now see +/- 400 nSec jitter on chrony socket.  I could do
> > better on a dedicated time server.  Best I can do with SHM is
> > currently about +/- 2 uSec.
> 
> I don't follow why this is different.  Isn't it all about
> sample-to-sample repeatability of obtaining the system clock
> (timecounter) at the edge?  WHy does it matter how it's conveyed to
> chronyd?

The SOCK interface is in nSec, the SHM interface is in uSec.
> 
> > I have a SiRF III, and KPPS on an eight core Xeon server ruinning
> > irqbalance.
> 
> What baud rate, and is that just PPS on DCD, with timestamping in the
> receive interrupt?

The speed is 38400, but the DCD is detected asynchronously.

>   I guess the secret is that your machine is
> unloaded so one of those cores is always ready to take the status
> change interrupt without being stuck in a critical section.

Not unloaded, but one core dedicated to that interrupt, using irqbalance.

>   Still,
> I'm surprised it's that low, and you're only 3dB better than where
> you'd be with 1 us resolution.  That's enough to make ns resolution
> worthwhile, but it also sems most people won't quite need it (for
> now, with DCD-on-serial).

Not only will almost no one need it, I'm not even wure what you could
actually use it for, other than bragging rights.

> Then, there's the questions of the unmodeled latecny beyond the
> jitter, and how the time gets used to timestamp other events.  Which
> I suspect blows past 1 us.  Still, not enough to cavalierly give away
> 0.5 us.

Plus long term wander of unknown characteristics.

But the good news is the more stable you local clock is the better NTP
can measure remote chimers.  Putting all together mst be "good".

> I tested a GR601-W on a modern 4-core box with NetBSD 6.  I think I am
> seeing delays of a few hundred us to a bit over 1 ms, relative to a s1
> one Ethernet hop away.   This is what I'd expect; some base delay when
> the clocks line up plus a delay in [0us, 1000 us), more or less.
> So I'd recommend 700 us for the default fudge for USB PPS, absent
> better measurements.

Seems fair

RGDS
GARY
---------------------------------------------------------------------------
Gary E. Miller Rellim 109 NW Wilmington Ave., Suite E, Bend, OR 97701
        address@hidden  Tel:+1(541)382-8588

signature.asc
Description: PGP signature