really slow PLL

On 7/21/2022 10:41 AM, bitrex wrote:
>> The cables and connectors could only be described as a bit clunky!
>> They really didn't get on with metal swarf being around but were OK in clean 
>> dry electronics/physics labs - much less so in chemistry ones...
> 
> I feel like there wasn't really a good relatively inexpensive standard for 
> interfacing PC peripherals until USB. Serial and parallel were slow 
> (occasionally some devices supported ECP, I remember having to enable it in the 
> BIOS sometimes), and external SCSI wasn't really well-suited to anything but 
> external disk drives.

That's sort of like claiming there really wasn't a good, relatively
inexpensive standard for browsing distributed resources across The Internet 
(gopher, anyone?)

Hardware costs have shrunk to the point where things that were science fiction
a decade ago are commonplace, nowadays.  (who'd have thought of wirelessly
connecting a mouse/keyboard to a PC -- or PHONE! -- decades ago?)

SCSI has always supported more than "just external disk drives".  I used
5380's as "message passing coprocessors" in a design decades ago.  Skip
the cost of the cables and connectors and just run a ribbon between cards,
driven by 5380's on each.

SCSI's problem as a universal interconnect comes from those cabling costs
along with the dearth of devices that embraced SCSI.  Other than disk,
tape and scanners, SCSI was a dead-end -- and only suitable for intermachine
communication at very short ranges.  Running 8 (or 16 or 32) data conductors
is costly in terms of cabling, connector *and* silicon.  (there's a reason
it's USB and not UPB!)

Who's gonna embrace something that has a small potential market?

> I don't think you could hot-swap IEE-488 either but it seems like it was pretty 
> fast and more amenable to a wide class of devices than SCSI, but just didn't 
> seem to catch on outside the test equipment realm.

On 21/7/22 22:52, Gerhard Hoffmann wrote:
> Am 21.07.22 um 13:19 schrieb jlarkin@highlandsniptechnology.com:
> 
>> Where does the 10 MHz come from?
> 
> Choise of implementer. One local clock generator is needed.
> This clock determines short term stabiity and phase noise.

Exactly. It's weird really, since the GPS P code bit rate is 10.23Mbps, 
meaning the PRNG shift registers are being clocked with 10.23MHz.
Perhaps the phase noise of the 10.23MHz clock isn't that critial...

At that rate, a 1-bit misalignment is 30m of distance. To get better 
accuracy you need a polyphase clock, and the correlation would move a 
shift register's clock to a clock which is forwards or back in phase 
from its current clock, rather than just skipping or adding clock pulses.

The Apollo ranging system used a bit rate around 1Mbps, but could 
resolve down to about 1m, or approximately 1 degrees of clock phase. If 
GPS could do that, it would be accurate to 10cm.

So anyhow, folk love their "GPSDO" 10MHz clock outputs often without 
really questioning the actual phase noise coming from the internal 
oscillator which synthesizes that output, and which ultimately isn't 
even the clock that's being used to recover the signals.

Clifford Heath.

On 7/21/2022 8:31 AM, John Walliker wrote:
> On Thursday, 21 July 2022 at 15:42:40 UTC+1, bitrex wrote:
>> HP used to have GPIB on their power supplies, I've never used it but I
>> expect it wasn't really useful for tight synchronization.
> 
> The Group Execute Trigger command does allow quite tight synchronisation
> between different GPIB devices.

If you're rolling your own PROPRIETARY protocol, you can design the
stack so that really tight (limited by hardware layer) latencies
are possible.  E.g., if your protocol KNOWS that a "sync" will
be "coming along shortly", the code can take steps to minimize the time
required to detect and respond to that.

On 7/21/2022 7:04 AM, bitrex wrote:
>> You really need to define longterm before the problem becomes well posed. Do 
>> you mean hours, days, weeks or months of runtime?
> 
> Yeah I don't quite get it, either. My rack of synthesizers can each play one 
> voice of the Maple Leaf Rag via MIDI and they all stay synced together really 

How is "really well" defined?  In the domain of human auditory perception?

> well, at least over a timespan of several minutes...superficially at least it 
> sounds like he wants a sequencer.

Can you disconnect the controller and have them continue playing (from stored
sequence) AND retain that synchornization "indefinitely"?

> Using the nuts & bolts system clock for synchronization of "user tasks" also 
> makes me uncomfortable, if the device behavior only need to align to the 
> millisecond why not trigger them using some simple network protocol and let 
> their hardware figure out how long a millisecond is independently. Do the 
> timings of these boxes need to be tighter than the Maple Leaf Rag?

Providing (and distributing) individual "events" on a shared medium suffers
from scaling problems.  How do you tag them so that the intended clients
know which events are of significance to them?

Folks can understand how to consult a single time reference WITHIN a device.
Admittedly, there can be skew in those references due to the frequency at
which the reference is updated as well as consulted (and, the time it
takes individual clients to "process" that information).

Bolting on a mechanism that allows the time references on different
nodes to act AS IF a single reference addresses the distribution of that
reference separately from the reference's internal consumers.

How frequently you resync those references depends on how much variation
you can encounter in the individual local timebases (even OCXOs have
tolerances) and how much slip the "application" can tolerate.  You
can choose to address the absolute time reference (it is now XXX:XX)
and/or the rate of time passage (it has been X time units since event Y).

If your MIDI instrument is told to emit a particular note/sound at
"beat #564", doesn't it depend on having some notion of WHEN beat #0
occurred and how long each beat is?  How long will that "rate" be
observed by its hardware (osc) along with those of its peers?

One pass-time at KCP was to "read" "Row, Row, Row Your Boat" into a
group of machines.  Then, twiddle the pitch of each and have them
replay the text from internal memory, in a "round".  Of course, it
was damn near impossible to get the *rate* controls at the same
setting (potentiometers) so it wasn't long before the "singers"
drifted far enough apart that it sounded like an angry mob!
(rhubarb, rhubarb, rhubarb).

But, there was never a need for such synchronization (beyond novelty)
so there were no design measures put in place to ensure it would be
repeatable and trackable.

On 7/21/2022 7:12 AM, bitrex wrote:
> On 7/21/2022 4:33 AM, John Walliker wrote:
>> If you have an ethernet interface to each unit then Precision Time Protocol
>> should do exactly what you want.
>> https://en.wikipedia.org/wiki/Precision_Time_Protocol
>> John
> 
> Yeah, that sounds like the ticket to me also. Trying to use each box's system 
> clock for purposes of keeping "user-space" tasks in sync across boxes makes me 
> uncomfortable, sounds like a srs hack.

How else would you synchronize events generated (or DETECTED!) across boxes?
Run a separate wire for each event?

> If you need to tightly synchronize events between physically separate hardware 
> why not use a standard designed for the task rather than some roll-your-own shit

Because standards try to address ALL POSSIBLE users of a particular
feature/mechanism.

Done "as intended", PTP requires timestamping hardware in the NIC to note
when the packets actually hit (or come off) the wire.  With a good deal of
care, you can get sub-microsecond synchronization between nodes.  But, swap
the switch or let network traffic change significantly (e.g., unconstrained)
and all bets are off.

If all you are trying to do is synchronize some notion of time across
devices, why take on all the overhead of an NIC, network stack,
PTP protocol, etc. JUST for that?

How tightly must the time references be synchronized?  How tightly must
they REMAIN synchronized (local oscillators drift at different rates)?

OTOH, if those mechanisms are already present/needed in your product/design,
then bolting PTP on top can make sense.

But, *how* you do that is your own decision -- if you don't need to
"play nice" with other vendors' products, then why take on the cost
of doing so?  (why so many oddball "serial port" connectors and
implementations, over the years?  wasn't there ONE standard??)

[As to the "hack", you underestimate how powerful the notion of
just being able to say "do this at t=X" is in designing a box!]

The biggest downside with shared resources -- esp ones that are
as crucial as this -- is sorting out how to handle the situation
when that resource fails or is unavailable.  E.g., PTP allows
a new master to be elected.  Fine.  But, implicit in that is the
fact that the old master is now "not functional"... can you
continue to operate under those conditions?

On 7/21/2022 9:04 PM, Don Y wrote:
> On 7/21/2022 7:04 AM, bitrex wrote:
>>> You really need to define longterm before the problem becomes well 
>>> posed. Do you mean hours, days, weeks or months of runtime?
>>
>> Yeah I don't quite get it, either. My rack of synthesizers can each 
>> play one voice of the Maple Leaf Rag via MIDI and they all stay synced 
>> together really 
> 
> How is "really well" defined?  In the domain of human auditory perception?
> 
>> well, at least over a timespan of several minutes...superficially at 
>> least it sounds like he wants a sequencer.
> 
> Can you disconnect the controller and have them continue playing (from 
> stored
> sequence) AND retain that synchornization "indefinitely"?
> 
>> Using the nuts & bolts system clock for synchronization of "user 
>> tasks" also makes me uncomfortable, if the device behavior only need 
>> to align to the millisecond why not trigger them using some simple 
>> network protocol and let their hardware figure out how long a 
>> millisecond is independently. Do the timings of these boxes need to be 
>> tighter than the Maple Leaf Rag?
> 
> Providing (and distributing) individual "events" on a shared medium suffers
> from scaling problems.  How do you tag them so that the intended clients
> know which events are of significance to them?


His machine's sequences are all programmed individual to a given box, 
there's no provision for controlling one from the others directly other 
than telling one to start its program and providing some kind of master 
sync, whatever it is.

The way this would be handled in the music-world analogy is something 
like MIDI timecode, it's not unusual for individual machines to have 
their own sequencers and want to gang events together in that realm, either.

<https://en.wikipedia.org/wiki/MIDI_timecode>

The resolution of the timecode doesn't automatically imply the timing 
resolution, given a "start" command and the clock any good-quality 
device will be able to interpolate between quarter-frames to know when 
it should fire an event, if its internal sequencer has finer granularity 
than that.

I mentioned elsewhere that there's a more general-purpose sync standard 
that's been around a long time:

<https://en.wikipedia.org/wiki/IRIG_timecode>

On 7/21/2022 6:19 PM, bitrex wrote:
> On 7/21/2022 9:04 PM, Don Y wrote:
>> On 7/21/2022 7:04 AM, bitrex wrote:
>>>> You really need to define longterm before the problem becomes well posed. 
>>>> Do you mean hours, days, weeks or months of runtime?
>>>
>>> Yeah I don't quite get it, either. My rack of synthesizers can each play one 
>>> voice of the Maple Leaf Rag via MIDI and they all stay synced together really 
>>
>> How is "really well" defined?  In the domain of human auditory perception?
>>
>>> well, at least over a timespan of several minutes...superficially at least 
>>> it sounds like he wants a sequencer.
>>
>> Can you disconnect the controller and have them continue playing (from stored
>> sequence) AND retain that synchornization "indefinitely"?
>>
>>> Using the nuts & bolts system clock for synchronization of "user tasks" also 
>>> makes me uncomfortable, if the device behavior only need to align to the 
>>> millisecond why not trigger them using some simple network protocol and let 
>>> their hardware figure out how long a millisecond is independently. Do the 
>>> timings of these boxes need to be tighter than the Maple Leaf Rag?
>>
>> Providing (and distributing) individual "events" on a shared medium suffers
>> from scaling problems.  How do you tag them so that the intended clients
>> know which events are of significance to them?
> 
> His machine's sequences are all programmed individual to a given box, there's 
> no provision for controlling one from the others directly other than telling 
> one to start its program and providing some kind of master sync, whatever it is.

But there is still a need to deploy that "program" (sequence).

Is there an inherent limit to the time spanned by such a (future)
definition?  Why not sync them on the assembly line and then
forget about it?  (Ans:  because there would be intolerable
drift in references)

> The way this would be handled in the music-world analogy is something like MIDI 
> timecode, it's not unusual for individual machines to have their own sequencers 
> and want to gang events together in that realm, either.
> 
> <https://en.wikipedia.org/wiki/MIDI_timecode>
> 
> The resolution of the timecode doesn't automatically imply the timing 
> resolution, given a "start" command and the clock any good-quality device will 
> be able to interpolate between quarter-frames to know when it should fire an 
> event, if its internal sequencer has finer granularity than that.

Of course.  I've been designing "clocks" (timepieces) for decades that
only display time to nothing better than 100Hz -- yet rely on the mains
frequency to discipline the local oscillator over the long run (days).

But, (decorative) timepieces only have to deal with synchronization
on the scale of human perception.  And, while mains power is available
(as a reference frequency), there is no slippage between timepieces.
The problem of spanning outages is where the real engineering comes
into play; how much will the current timepiece's XTAL drift wrt
other timepieces in the absence of that line frequency reference?
(timepieces can't talk to each other so any skew that occurs during
an outage is "baked in" thereafter)

> I mentioned elsewhere that there's a more general-purpose sync standard that's 
> been around a long time:

IMO, there's no need to transfer time information.  You need a reference
point (in time) and then a reference *rate*.

The rate needs to be broadcast ("shared") often enough that slip can
be noticed and corrected -- before it becomes "significant" (for whatever
value of "significant" is pertinent).

The reference point can be indicated in-band by a double-pulse
(i.e., first pulse occurs at the intended "rate" point, in time;
presence of another pulse within some delta thereafter means
"that rate pulse was also a reference point").

Of course, if you are a stickler for detail, you'd need some way of
addressing difference in transport delays (likely not significant
within a single rack but of interest when devices are hundreds of
meters apart!)

On 22/7/22 03:10, Phil Hobbs wrote:
> Gerhard Hoffmann wrote:
>> Am 21.07.22 um 16:15 schrieb Phil Hobbs:
>>
>>> I wonder if there's an advantage to using the closure phase for an 
>>> array that large.  With 17 oscillators you've got 136 independent 
>>> phase differences, so maybe there's a way to get 22 dB instead of 12 
>>> dB improvement.
>>
>> -v ?
>>
>> what do you mean with closure phase? Where do the 22 dB come
>> from?
>>
>> The idea was simply to have all 16 regulated to the be
>> synchronous and then feed them into a 16-to--1 Wilkinson
>> combiner. The phase noise should average out among the
>> 16 units. Just as proof of concept. The MTI-260 are quite ok,
>> but with bleeding edge oscillators that could be interesting.
>> In the region where you just cannot improve an oscillator.
>>
>>> Cheers
>> Gerhard
> 
> Sure.  Thing is, that wastes a lot of information that you could maybe 
> use to get 10*log(136) = 21.3 dB improvement instead of 10*log(17) = 
> 12.3 dB.  [136 = N(N-1)/2 when N = 17.]
> 
> Closure is a really cute idea, which I first came across in the context 
> of very long baseline interferometry (VLBI) radio telescopes.  See the 
> discussion from BEOS 3e here:
> 
> <https://electrooptical.net/www/sed/closure.png>.

Interesting, thanks.

Some frequency synthesiser chips employ proprietary majick to reduce the 
phase noise associated with integer divide/multiply ratios. Polyphase 
oscillator and slipping by partial cycles I think. Perhaps they're doing 
something like closure against the different clock phases?

Clifford Heath

On Thursday, July 21, 2022 at 10:30:47 AM UTC-7, lang...@fonz.dk wrote:
> torsdag den 21. juli 2022 kl. 01.21.08 UTC+2 skrev John Larkin: 
> > Suppose I have several rackmount boxes...
> > Each box has a 40 MHz VCXO and I want to phase-lock them, or at least 
> > time-align them to always be the same within a few microseconds, 
> > longterm. 
> > 
> > I could call one the leader (not "master") and make the others 
> > followers (not "slaves") and have the leader make an active low pulse 
> > maybe once a second.

> why so slow?
> > A follower would use her (not "his") clock to 
> > measure the incoming period and tweak its local VCXO in the right 
> > direction. That should work.

> it'll only make the run at at the same speed, not time aligned

If you ever apply a command time setting  once, and they have the same
speed, they will stay aligned.   It'd be an improvement over letting the
clocks drift for a second at a time, then correcting.   I'm not sure what
'measure the incoming period' has to do with it, though; just use an up/down
counter to compare leader to follower, and D/A the count to trim the VCO.

All it takes to sense a frequency difference is a counter.   Negative
feedback nulls the difference.

On Fri, 22 Jul 2022 00:08:56 -0000 (UTC), Cydrome Leader
<presence@MUNGEpanix.com> wrote:

>jlarkin@highlandsniptechnology.com wrote:
>> On Thu, 21 Jul 2022 07:43:18 +0200, Gerhard Hoffmann <dk4xp@arcor.de>
>> wrote:
>> 
>>>Am 21.07.22 um 01:20 schrieb John Larkin:
>>>> 
>>>> 
>>>> Suppose I have several rackmount boxes and each has a BNC connector on
>>>> the back. Each of them has an open-drain mosfet, a weak pullup, and a
>>>> lowpass filtered schmitt gate back into our FPGA.
>>>> 
>>>> I can daisy-chain several boxes with BNC cables and tees.
>>>> 
>>>> Each box has a 40 MHz VCXO and I want to phase-lock them, or at least
>>>> time-align them to always be the same within a few microseconds,
>>>> longterm.
>>>
>>>I have a backburner project of locking 16 MTI-260 oscillators
>>>slooowy to another one, and when they are in sync, combine
>>>them  with an array of Wilkinsons. That should have a nice
>>>effect on phase noise by averaging over 16.
>>>The CPLD has enough resources to implement that as a delay
>>>locked loop with 1 pps, but low hanging fruit first.
>>>
>>>> 
>>>> I could call one the leader (not "master") and make the others
>>>> followers (not "slaves") and have the leader make an active low pulse
>>>> maybe once a second. A follower would use her (not "his") clock to
>>>> measure the incoming period and tweak its local VCXO in the right
>>>> direction. That should work.
>>>> 
>>>> Don't GPS receivers lock their 10 MHz oscillators to a 1 PPS pulse
>>>> from the satellites?
>>>
>>>No. There is no 1PPS pulse from the sat nor the need for exactly 10 MHz. 
>>>The sats transmit a pseudo noise sequence that is
>>>aligned to the second of their local clock source.
>>>The GPS receiver knows the polynomial and runs a local copy of
>>>the polynomial. It knows by cross correlation if the local
>>>pseudo noise is the same as that of the sat and therefore knows
>>>the start of the second. Usually that won't be the case.
>>>Then the receiver delays its own polynomial by omitting a
>>>clock to the shift register that generates it and tries again.
>>>Sooner or later it will fit.
>> 
>> Where does the 10 MHz come from?
>
>https://en.wikipedia.org/wiki/GPS_disciplined_oscillator

"GPSDOs typically phase-align the internal flywheel oscillator to the
GPS signal by using dividers to generate a 1PPS signal from the
reference oscillator, then phase comparing this 1PPS signal to the
GPS-generated 1PPS signal and using the phase differences to control
the local oscillator frequency in small adjustments via the tracking
loop."

That's what I meant: the 10M xo is locked to the 1 PPS GPS output.

The GPS 1 PPS is perfect (by definition) long-term but terrible
short-term, so the XO or rubidium has to be very good itself, and the
loop has to be very slow. Big flywheel.

I'll be doing something similar, locking my 40 MHz clock to some 1 PPS
input, the difference being that I don't mind a few us of jitter, so I
can lock quick and crude.