[ previous ] [ next ] [ threads ]
 
 From:  Fred Wright <fw at well dot com>
 To:  m0n0wall at lists dot m0n0 dot ch
 Subject:  Re: [m0n0wall] New clock update, img & diff available
 Date:  Thu, 26 Jun 2003 20:53:52 -0700 (PDT)
On Wed, 25 Jun 2003, Michael Sierchio wrote:
> Fred Wright wrote:
> 
> > So he shows that with the 4501 running from an atomic clock synced to the
> > PPS signal, the PPS samples are good.  Wow!  What a surprise! :-)
> > Actually, I think the 4501 *can* do pretty well with this, but I wouldn't
> > use *that* experiment to prove anything.
> 
> It proves you can have a 1us NTP stratum 1 server with a net4501.

No, it proves you can have a 120ns NTP stratum 1 server with a net4501
that's been modified to replace it's main clock oscillator with a PLL
synchronized to an atomic clock.  With that kind of hardware, you hardly
even need the GPS except once at boot time. :-)  Also, you can buy a
complete Accutime system from Trimble for about half what his setup cost,
when you include the $1500 for the rubidium frequency standard.

His experiment proves nothing that couldn't have been predicted with some
basic knowledge of hardware and some purely software timing experiments.  
The latter are only necessary due to the inadequacy of the AMD
documentation.

> > He is also depending on controlling one of the GP timers with the PPS
> > signal, but the inputs with that possibility aren't available on the user
> > I/O connector.  
> 
> This requires a low-profile RS422 card (which fits in the Soekris
> case) for a long-run cable to the GPS.

That's not sufficient for what *he* did.  As I said above, his setup
involves *directly controlling* one of the SC520's GP timers with the PPS
signal.  The pins that have that capability are not available on any
connector in the 4501.  Another mod to the 4501, in addition to the
crystal replacement.

I'm not saying you *have* to do that to get adequate performance, only
that that's what *he* did.

On Wed, 25 Jun 2003, Michael Sierchio wrote:
> 
> This means that the computer's clock will be selected when no
> other sources are available.  This will typically drift a bit,
> but the net45xx have very good oscillators.

I'm not sure what you eman by "very good oscillators".  AFAICT they're
run-of-the mill crystals, probably rated at +/-100ppm.  I haven't measured
the 33.33MHz one on mine yet, but since you say "oscillators" plural, it's
worth mentioning that my RTC runs 62ppm slow, which is the worst I've ever
seen on a computer battery clock.  It's within spec for a "general use"
crystal, though usually one finds better 32768Hz crystals for timekeeping.  
By comparison, most digital watches are specced at +/-6ppm, unadjusted.

The main thing you have going for you is that the temperature in the 4501
is probably fairly stable, so once you compensate for the fixed offset in
the frequency, you can probably do considerably better than 1ppm.

On Wed, 25 Jun 2003, Michael Mee wrote:

> > whenever I used AboutTime [...] and have it communicate to
> > m0n0walls msntp server, it times out.
> 
> Thanks Joe. I didn't enable msntp as a *server*, just as a client (i.e.
> to set the time on the Soekris box) so that's why it hangs with AboutTime.
> Server mode is v2, though its simple to do right now. I *think* you do the
> following, though doubtless others will have more refined ways:

I don't know where you get "v2" from, unless this is a really ancient
version of msntp.  The oldest msntp I have is V1.4 (Sep-97), which uses
NTP V3.  But note that in server mode, the version number is an echo of
the client's version number.

					Fred Wright