[time-nuts] Re: NTP server source needed
Though I don't have a Time Machines NTP server I've monitored them for several years and have seen a number of "off by N-seconds" errors. The manufacturer reproduced the problem and released new firmware for the TM2000B. I don't know the status of other models. I would be hesitant to trust the server when the GPS signal is marginal. I have a local LeoNTP server. Though it uses a different chipset than the Time Machines TM, it also exhibits off by N-seconds. I've seen similar errors for several other public LeoNTP servers, including some of the ntpX.leontp.com servers. When I degrade the GPS signal the N-second errors become more frequent. A typical NTP server will indicate potential loss of accuracy by setting the alarm bit or increasing root dispersion. The LeoNTP does neither, instead it decreases the NTP "precision" field. [The value itself is bogus.] The machine's physical design is nice, the performance is excellent. I wish the timestamps were more trustworthy. Steve Sommars On Thu, Apr 21, 2022 at 8:20 PM Paul Watts wrote: > Thanks for the reply, Forrest. > > I do have a TM1000(model number?) but I thought they were out of stock. > Their website currently claims otherwise and I'll probably get another. > > I do have systems at each site running NTP against my local LeoNTP devices > and the US NTP pool. > > I just like to have a local source as well as remote. > > Recently, I learned about the Centerlclick NTP250. Anyone have any > experience? > > On Thu, Apr 21, 2022 at 12:09 Forrest Christian (List Account) < > li...@packetflux.com> wrote: > > > Time machines makes a similar box. > > > > However, one would expect that with that many devices at a site one could > > find a couple of devices which could run a ntp server along with their > > normal workload. > > > > On Thu, Apr 21, 2022, 2:23 AM Paul Watts wrote: > > > > > Long time lurker, first time poster. > > > > > > In the past, I've purchased the LeoNTP server to act as a local time > > server > > > on my work network. Unfortunately, those NTP servers aren't available > to > > > order. > > > > > > I'm in need of two - three more NTP servers for various locations. > > > > > > Do any of you have an alternate NTP server recommendation? > > > > > > 1. I'd like to keep the price less than $1,000. The Leo's were a great, > > > self contained choice at less than $400. > > > 2. Support for IPv4 is required and IPv6 is desired. The main > > shortcoming I > > > have with the Leo's is no support for IPv6. > > > 3. I'll serve about 150 clients per site from a given NTP server. My > NTP > > > clients are a combination of Windows and Linux computers along with > > various > > > types of network gear (Cisco, Palo Alto, Aruba, Juniper). The Leo's > don't > > > even breathe hard at that level. > > > 4. No PTP support is required. My timing needs are IT related - log > > > consistency, security protocol requirements, etc. > > > 4. I'm not interested in building and using something like a Raspberry > > Pi. > > > I'll be sending these to remote sites and don't want to worry about > > issues > > > like memory card failures and the hobbyist feel of many Pi cases and > > > cooling options. That's fine for my home but not for work. > > > 5. Availability is important. > > > > > > Any suggestions are welcome. > > > > > > Thanks, > > > > > > Paul Watts. > > > ___ > > > time-nuts mailing list -- time-nuts@lists.febo.com -- To unsubscribe > > send > > > an email to time-nuts-le...@lists.febo.com > > > To unsubscribe, go to and follow the instructions there. > > > > > ___ > > time-nuts mailing list -- time-nuts@lists.febo.com -- To unsubscribe > send > > an email to time-nuts-le...@lists.febo.com > > To unsubscribe, go to and follow the instructions there. > > > -- > Paul Watts. > rpwa...@gmail.com > ___ > time-nuts mailing list -- time-nuts@lists.febo.com -- To unsubscribe send > an email to time-nuts-le...@lists.febo.com > To unsubscribe, go to and follow the instructions there. > ___ time-nuts mailing list -- time-nuts@lists.febo.com -- To unsubscribe send an email to time-nuts-le...@lists.febo.com To unsubscribe, go to and follow the instructions there.
[time-nuts] Re: GPSDO Control loop autotuning
On Dienstag, 12. April 2022 00:52:42 CEST Magnus Danielson via time-nuts wrote: > A trivial state-space Kalman would have phase and frequency. Assuming > you can estimate the phase and frequency noise of both the incoming > signal and the steered oscillator, it's a trivial exercise. It's > recommended to befriend oneself with Kalman filters as a student > exercise. It can do fairly well in various applications. I'm wondering, can the frequency really be part of the state? It is not a property we can measure independently, we can only derive it from the TIC measurements. There is the EFC value of course, but that doesn't directly correspond with the output frequency... ___ time-nuts mailing list -- time-nuts@lists.febo.com -- To unsubscribe send an email to time-nuts-le...@lists.febo.com To unsubscribe, go to and follow the instructions there.
[time-nuts] Re: Question about GPS 1PPS ADEV
On Donnerstag, 21. April 2022 13:48:42 CEST Markus Kleinhenz via time-nuts wrote: > Hello Matthias, > > Could you explain further how enabled GLONASS messed things up for you? Noticeably more transients in the phase error. Phase jumps of 15ns or more. Since then I only enable GPS and Galileo, to mitigate my incomplete sky view a bit. I only have view west to east over south and I filter all SVs below 20° elevation. > > Regards, > > Markus > ___ > time-nuts mailing list -- time-nuts@lists.febo.com -- To unsubscribe send an > email to time-nuts-le...@lists.febo.com To unsubscribe, go to and follow > the instructions there. ___ time-nuts mailing list -- time-nuts@lists.febo.com -- To unsubscribe send an email to time-nuts-le...@lists.febo.com To unsubscribe, go to and follow the instructions there.
[time-nuts] Re: Question about GPS 1PPS ADEV
Private email from the folks at NRCan indicated that having both constellations in the data would potentially result in (more) clock offset jumps in the PPP results. Remember that GLONASS and GPS are synced to different master clocks which are some small (and potentially variable) number of nanoseconds apart. But part of it may also be that their modeling of GLONASS clocks and orbitals isn't as robust as for GPS due to having less experience with it. If you're purely interested in location, multiple constellations can be helpful, but the precision timing application is a bit different. Also, note that this is probably a non-issue for simple non-post-processed GPS PPS, as the other error sources would swamp any timescale differences. John On Apr 21, 2022, 12:09 PM, at 12:09 PM, Markus Kleinhenz via time-nuts wrote: >Hello Matthias, > >> Of course curiosity got the better of me. I switched the GPSDO to >hold-over >> and recorded about 4000 seconds of data. Result is attached. BTW I >terminated >> the experiment after the temperature compensation kicked the DAC one >LSB down. >> I am quite pleased with the performance of the OCXO, though. I was >not >> expecting that, it is really nothing special. Not a surplus Chinesium >OCXO, >> but something you can buy for quite reasonable money from Digikey. >> >> So, the bulge is getting a bit more prominent, but it is in no way as > >> prominent as the figure 26 in John Ackermann's paper. >> >> John, what constellations did you have enabled during the test? Just >GPS, or >> also others? I usually run with GPS and Galileo enabled (and I avoid >GLONASS, >> it messes everything up). Can this make a difference? >> >> >Could you explain further how enabled GLONASS messed things up for you? > >Regards, > >Markus >___ >time-nuts mailing list -- time-nuts@lists.febo.com -- To unsubscribe >send an email to time-nuts-le...@lists.febo.com >To unsubscribe, go to and follow the instructions there. ___ time-nuts mailing list -- time-nuts@lists.febo.com -- To unsubscribe send an email to time-nuts-le...@lists.febo.com To unsubscribe, go to and follow the instructions there.
[time-nuts] Re: Question about GPS 1PPS ADEV
I have yet to find a reliable source for this, but something I have heard in the past is that since the different constellations' satellites may be controlled and disciplined differently (and to different standards), it is not wise to use more than one constellation on receiver that is used to discipline a time server. Like I said though, I have not found a reliable source for this, and I haven't even been able to confirm it in my own testing. I would love to know more on the subject. Regards, John Miller > On Apr 21, 2022, at 7:48 AM, Markus Kleinhenz via time-nuts > wrote: > > Hello Matthias, > >> Of course curiosity got the better of me. I switched the GPSDO to hold-over >> and recorded about 4000 seconds of data. Result is attached. BTW I >> terminated >> the experiment after the temperature compensation kicked the DAC one LSB >> down. >> I am quite pleased with the performance of the OCXO, though. I was not >> expecting that, it is really nothing special. Not a surplus Chinesium OCXO, >> but something you can buy for quite reasonable money from Digikey. >> >> So, the bulge is getting a bit more prominent, but it is in no way as >> prominent as the figure 26 in John Ackermann's paper. >> >> John, what constellations did you have enabled during the test? Just GPS, or >> also others? I usually run with GPS and Galileo enabled (and I avoid >> GLONASS, >> it messes everything up). Can this make a difference? >> >> > Could you explain further how enabled GLONASS messed things up for you? > > Regards, > > Markus > ___ > time-nuts mailing list -- time-nuts@lists.febo.com -- To unsubscribe send an > email to time-nuts-le...@lists.febo.com > To unsubscribe, go to and follow the instructions there. ___ time-nuts mailing list -- time-nuts@lists.febo.com -- To unsubscribe send an email to time-nuts-le...@lists.febo.com To unsubscribe, go to and follow the instructions there.
[time-nuts] Re: NTP server source needed
Thanks for the reply, Forrest. I do have a TM1000(model number?) but I thought they were out of stock. Their website currently claims otherwise and I'll probably get another. I do have systems at each site running NTP against my local LeoNTP devices and the US NTP pool. I just like to have a local source as well as remote. Recently, I learned about the Centerlclick NTP250. Anyone have any experience? On Thu, Apr 21, 2022 at 12:09 Forrest Christian (List Account) < li...@packetflux.com> wrote: > Time machines makes a similar box. > > However, one would expect that with that many devices at a site one could > find a couple of devices which could run a ntp server along with their > normal workload. > > On Thu, Apr 21, 2022, 2:23 AM Paul Watts wrote: > > > Long time lurker, first time poster. > > > > In the past, I've purchased the LeoNTP server to act as a local time > server > > on my work network. Unfortunately, those NTP servers aren't available to > > order. > > > > I'm in need of two - three more NTP servers for various locations. > > > > Do any of you have an alternate NTP server recommendation? > > > > 1. I'd like to keep the price less than $1,000. The Leo's were a great, > > self contained choice at less than $400. > > 2. Support for IPv4 is required and IPv6 is desired. The main > shortcoming I > > have with the Leo's is no support for IPv6. > > 3. I'll serve about 150 clients per site from a given NTP server. My NTP > > clients are a combination of Windows and Linux computers along with > various > > types of network gear (Cisco, Palo Alto, Aruba, Juniper). The Leo's don't > > even breathe hard at that level. > > 4. No PTP support is required. My timing needs are IT related - log > > consistency, security protocol requirements, etc. > > 4. I'm not interested in building and using something like a Raspberry > Pi. > > I'll be sending these to remote sites and don't want to worry about > issues > > like memory card failures and the hobbyist feel of many Pi cases and > > cooling options. That's fine for my home but not for work. > > 5. Availability is important. > > > > Any suggestions are welcome. > > > > Thanks, > > > > Paul Watts. > > ___ > > time-nuts mailing list -- time-nuts@lists.febo.com -- To unsubscribe > send > > an email to time-nuts-le...@lists.febo.com > > To unsubscribe, go to and follow the instructions there. > > > ___ > time-nuts mailing list -- time-nuts@lists.febo.com -- To unsubscribe send > an email to time-nuts-le...@lists.febo.com > To unsubscribe, go to and follow the instructions there. > -- Paul Watts. rpwa...@gmail.com ___ time-nuts mailing list -- time-nuts@lists.febo.com -- To unsubscribe send an email to time-nuts-le...@lists.febo.com To unsubscribe, go to and follow the instructions there.
[time-nuts] GPS Control Loop
25 years ago Brooks Shera developed a GPSDO that had a Frequency accuracy of 1 E-11. A year later in 1998 he published his work in July 11998 in QST. The total package included circuits, PC board and in depth analysis of every part of his design including his filter.. I was one of the first buying the board, GPS receiver and working with an other Time Nut building a couple of units. Using my Tracor 527E,, HP5345A and aHP5061A Cs in parallel to an Austron Loran C system using multiple HP 10811's in all cases observed better than 1 E-11 frequency. All 10811's had previously been tested to have an AV below E-12. Subsequent tests with Austron 1150 and Datum 1000 showed even better results. Over time Brooks added features but the design remained the same. His last effort was to make other DAC's available, but at that time his Cancer had spread to his brain. Lets remember this was the time of S/A and it is safe to say GPS has improved by a factor of 100. DAC's compared to the Audio DAC are also improved by a factor 50. Using later circuits we still use his design and Richard McCorkle did a PIC based on his Filter for us. Very little else has changed. Reviewing commercial units like End Run short term is a function of the OCXO or Rb, long term GPS. There was a time when Time Nuts had access to quality surplus OCXO's and Rb's at a 'reasonable" price, that tine is gone. Before the time I did get an OSA 8607 I did look at a Morion OCXO, their best with an AV of 2 E-13 at 1 second. The would not respond to my RFQ but thanks to a Time Nut that buys from them regular the answer was 6000 Euro and 1 year delivery after order! To us Shera is the Gold Standard. worth a review for any one doing their own GPSDO, we are presently focusing on the best GPS receiver, my problem is Ublox 9T is not an option because at 80 I am not able to get safely on the roof. Hire some one is an option but liability may be a problem Bert Kehren ___ time-nuts mailing list -- time-nuts@lists.febo.com -- To unsubscribe send an email to time-nuts-le...@lists.febo.com To unsubscribe, go to and follow the instructions there.
[time-nuts] Re: GPSDO Control loop autotuning
Hi Erik,
when I familiarized myself with the topic back in 2018, I found this paper:
https://ieeexplore.ieee.org/document/7324685 (IEEE paywall)
https://www.researchgate.net/publication/308730123_Synchronization_robustness_using_Kalman-based_clock_servos
They found that Kalman filters and PI controllers perform equally when
used digital clock servos if tuned appropriately. While the paper is
about network clock synchronization, the problem is very similar to OCXO
disciplining. Instead of GNSS PPS jitter you have packet delay
variation, which is more complex than simple jitter. I haven't checked
any scientific literature after 2018.
Best regards,
Carsten
On 21.04.22 11:00, Erik Kaashoek wrote:
Hi Tobias,
Using the excellent material found here:
https://nbviewer.ipython.org/github/rlabbe/Kalman-and-Bayesian-Filters-in-Python/blob/master/table_of_contents.ipynb
I went ahead and implemented the first order filter as described in
chapter 8.
The simulation was changed to use a real recorded PPS as measurement
noise and a real recorded TIC of a TCXO as the to be estimated data
and a Vtune steering an added frequency
The STDEV of the PPS was 5.6e-9
The kalman filter predicts the phase and frequency fairly well
STDERR of phase is 4e-9
STDERR of freq is 1.75e-9
The loop was then closed with a PI controller tasked to minimize the
phase error.
Tuning was a bit of a problem but the current performance is
STDEV phase error 1.25e-8
STDEV frequency error 1.2e-9
Unfortunately the resulting ADEV is still worse compared to a well
tuned PI controller acting directly on the TIC of the TCXO.
More work to do
Erik.
On 21-4-2022 9:44, Pluess, Tobias via time-nuts wrote:
Hi Erik,
yes I also found this webpage, and the examples are good.
I had the exact same idea as you, except I thought that the model for
the
VCO output is either phase only or phase and frequency. Yours is even
a bit
more sophisticated, thats quite cool!
And yes, of course it will be possible to have a one dimensional filter,
even if your state space equations have more than one state. However, in
this case, A, B and C are a matrix and two vectors (sometimes also
called
F, G and H). But the output that is measured ("observed") is just
scalar,
i.e. 1D.
I am not sure whether your state extrapolation is correct, I don't
see it
at the moment unfortunately. But I will think about it a bit more.
To your question, about how to close the loop.
Well. What do you do when you don't have the alpha-beta-gamma filter?
you take your phase error readings and feed them into some sort of
controller, for example a PID controller, do you.
Now think of the alpha-beta-gamma filter (or also the Kalman filter) as
some sort of "prefilter". (do you remember when you asked me why I
prefilter my 1PPS measurements? :-D at that time, I had a 2pole IIR
filter
with adjustable time constant as prefilter). So you feed your actual
1PPS
measurements into the alpha-beta-gamma filter, and the filter outputs
the
1PPS measurements, but with noise mostly removed (hopefully!). Then, use
this as the error signal for your controller.
As how to choose the alpha, beta and gamma coefficients, you need to
know
the variance of your input signal. i.e. determine the standard
deviation of
the 1PPS readings and square that. I am not sure at the moment how to
find
the coefficients from the variance, but I believe somewhere in the
kalmanfilter.net tutorial, there is an equation for it if I remember
correctly.
Also worth mentioning is the fact that some GPS modules also output an
estimate of the timing accuracy. This can also be used as an additional
input to calculate the coefficients and adapt them, as receiving
conditions
change.
does that help a bit?
I am unfortunately not yet so far as you, I have made some calculations
using matlab, but I am not sure whether they are correct. I do see some
filtering effect of the Kalman filter, though. (I used the Kalman
filter,
with the error covariance matrices, not the alpha-beta-gamma filter,
like
you did, but maybe this is a better idea!). Are you interested in
sharing
your excel sheet?
thanks,
best
Tobias
HB9FSX
On Wed, Apr 20, 2022 at 1:18 AM Erik Kaashoek wrote:
Tobias,
Your Kalman post triggered me to study this "dummy" level introduction
to Kalman filters: https://www.kalmanfilter.net/
and now I'm trying to write the Kalman filter.
An example of a Kalman filter described in the link above uses
distance,
speed and acceleration as predictors for a one dimensional Kalman
filter
So I wondered if it would be possible to do the same for the phase, eg
use phase, frequency and drift in the prediction model and still have a
one dimensional model?
The Kalman filter thus uses the normalized phase as input (the measured
time difference between the 1 PPS and the 1 second pulse derived from
the VCO output) to predict the next phase, frequency and drift
p = phase
f = frequency (d Phase/d T)
d = drift (d Frequency / d T)
m =
[time-nuts] Re: GPSDO Control loop autotuning
Kalman filters are optimal only if they work with statistical processes
that can be described with derivatives or integrals, i.e. white phase noise
(f^0), white FM (f^-2), random walk FM (f^-4) and so on.
Most of VCOs are heavily flicker FM dominated
Il giorno gio 21 apr 2022 alle ore 19:09 Erik Kaashoek
ha scritto:
> Hi Tobias,
> Using the excellent material found here:
>
> https://nbviewer.ipython.org/github/rlabbe/Kalman-and-Bayesian-Filters-in-Python/blob/master/table_of_contents.ipynb
> I went ahead and implemented the first order filter as described in
> chapter 8.
> The simulation was changed to use a real recorded PPS as measurement
> noise and a real recorded TIC of a TCXO as the to be estimated data and
> a Vtune steering an added frequency
> The STDEV of the PPS was 5.6e-9
> The kalman filter predicts the phase and frequency fairly well
> STDERR of phase is 4e-9
> STDERR of freq is 1.75e-9
>
> The loop was then closed with a PI controller tasked to minimize the
> phase error.
> Tuning was a bit of a problem but the current performance is
> STDEV phase error 1.25e-8
> STDEV frequency error 1.2e-9
>
> Unfortunately the resulting ADEV is still worse compared to a well tuned
> PI controller acting directly on the TIC of the TCXO.
> More work to do
> Erik.
>
> On 21-4-2022 9:44, Pluess, Tobias via time-nuts wrote:
> > Hi Erik,
> >
> > yes I also found this webpage, and the examples are good.
> > I had the exact same idea as you, except I thought that the model for the
> > VCO output is either phase only or phase and frequency. Yours is even a
> bit
> > more sophisticated, thats quite cool!
> > And yes, of course it will be possible to have a one dimensional filter,
> > even if your state space equations have more than one state. However, in
> > this case, A, B and C are a matrix and two vectors (sometimes also called
> > F, G and H). But the output that is measured ("observed") is just scalar,
> > i.e. 1D.
> > I am not sure whether your state extrapolation is correct, I don't see it
> > at the moment unfortunately. But I will think about it a bit more.
> >
> > To your question, about how to close the loop.
> > Well. What do you do when you don't have the alpha-beta-gamma filter?
> > you take your phase error readings and feed them into some sort of
> > controller, for example a PID controller, do you.
> > Now think of the alpha-beta-gamma filter (or also the Kalman filter) as
> > some sort of "prefilter". (do you remember when you asked me why I
> > prefilter my 1PPS measurements? :-D at that time, I had a 2pole IIR
> filter
> > with adjustable time constant as prefilter). So you feed your actual 1PPS
> > measurements into the alpha-beta-gamma filter, and the filter outputs the
> > 1PPS measurements, but with noise mostly removed (hopefully!). Then, use
> > this as the error signal for your controller.
> >
> > As how to choose the alpha, beta and gamma coefficients, you need to know
> > the variance of your input signal. i.e. determine the standard deviation
> of
> > the 1PPS readings and square that. I am not sure at the moment how to
> find
> > the coefficients from the variance, but I believe somewhere in the
> > kalmanfilter.net tutorial, there is an equation for it if I remember
> > correctly.
> > Also worth mentioning is the fact that some GPS modules also output an
> > estimate of the timing accuracy. This can also be used as an additional
> > input to calculate the coefficients and adapt them, as receiving
> conditions
> > change.
> >
> > does that help a bit?
> > I am unfortunately not yet so far as you, I have made some calculations
> > using matlab, but I am not sure whether they are correct. I do see some
> > filtering effect of the Kalman filter, though. (I used the Kalman filter,
> > with the error covariance matrices, not the alpha-beta-gamma filter, like
> > you did, but maybe this is a better idea!). Are you interested in sharing
> > your excel sheet?
> >
> > thanks,
> > best
> > Tobias
> > HB9FSX
> >
> >
> >
> >
> > On Wed, Apr 20, 2022 at 1:18 AM Erik Kaashoek wrote:
> >
> >> Tobias,
> >> Your Kalman post triggered me to study this "dummy" level introduction
> >> to Kalman filters: https://www.kalmanfilter.net/
> >> and now I'm trying to write the Kalman filter.
> >> An example of a Kalman filter described in the link above uses distance,
> >> speed and acceleration as predictors for a one dimensional Kalman filter
> >> So I wondered if it would be possible to do the same for the phase, eg
> >> use phase, frequency and drift in the prediction model and still have a
> >> one dimensional model?
> >> The Kalman filter thus uses the normalized phase as input (the measured
> >> time difference between the 1 PPS and the 1 second pulse derived from
> >> the VCO output) to predict the next phase, frequency and drift
> >> p = phase
> >> f = frequency (d Phase/d T)
> >> d = drift (d Frequency / d T)
> >> m = measured phase
> >> with a 1 second interval between observations the state
[time-nuts] Re: EFTF/IFCS 2022 - time-nuts meetup?
Hi Attila, I will attend the EFTF/IFCS in Paris together with the AXTAL Chief Engineer Henry. We will arrive Monday afternoon and will leave on Thursday at noon. Will be glad to meet you both! Best regards Bernd -Ursprüngliche Nachricht- Von: Attila Kinali [mailto:att...@kinali.ch] Gesendet: Mittwoch, 20. April 2022 20:47 An: Discussion of precise time and frequency measurement Betreff: [time-nuts] EFTF/IFCS 2022 - time-nuts meetup? Hey there! As some of you might be aware, next week is EFTF/IFCS in Paris. Magnus and I will be there and we were wondering if any other time-nuts would be present. Attila Kinali -- The driving force behind research is the question: "Why?" There are things we don't understand and things we always wonder about. And that's why we do research. -- Kobayashi Makoto ___ time-nuts mailing list -- time-nuts@lists.febo.com -- To unsubscribe send an email to time-nuts-le...@lists.febo.com To unsubscribe, go to and follow the instructions there. ___ time-nuts mailing list -- time-nuts@lists.febo.com -- To unsubscribe send an email to time-nuts-le...@lists.febo.com To unsubscribe, go to and follow the instructions there.
[time-nuts] Re: Question about GPS 1PPS ADEV
Hello Matthias, > Of course curiosity got the better of me. I switched the GPSDO to hold-over > and recorded about 4000 seconds of data. Result is attached. BTW I terminated > the experiment after the temperature compensation kicked the DAC one LSB > down. > I am quite pleased with the performance of the OCXO, though. I was not > expecting that, it is really nothing special. Not a surplus Chinesium OCXO, > but something you can buy for quite reasonable money from Digikey. > > So, the bulge is getting a bit more prominent, but it is in no way as > prominent as the figure 26 in John Ackermann's paper. > > John, what constellations did you have enabled during the test? Just GPS, or > also others? I usually run with GPS and Galileo enabled (and I avoid GLONASS, > it messes everything up). Can this make a difference? > > Could you explain further how enabled GLONASS messed things up for you? Regards, Markus ___ time-nuts mailing list -- time-nuts@lists.febo.com -- To unsubscribe send an email to time-nuts-le...@lists.febo.com To unsubscribe, go to and follow the instructions there.
[time-nuts] Re: NTP server source needed
Time machines makes a similar box. However, one would expect that with that many devices at a site one could find a couple of devices which could run a ntp server along with their normal workload. On Thu, Apr 21, 2022, 2:23 AM Paul Watts wrote: > Long time lurker, first time poster. > > In the past, I've purchased the LeoNTP server to act as a local time server > on my work network. Unfortunately, those NTP servers aren't available to > order. > > I'm in need of two - three more NTP servers for various locations. > > Do any of you have an alternate NTP server recommendation? > > 1. I'd like to keep the price less than $1,000. The Leo's were a great, > self contained choice at less than $400. > 2. Support for IPv4 is required and IPv6 is desired. The main shortcoming I > have with the Leo's is no support for IPv6. > 3. I'll serve about 150 clients per site from a given NTP server. My NTP > clients are a combination of Windows and Linux computers along with various > types of network gear (Cisco, Palo Alto, Aruba, Juniper). The Leo's don't > even breathe hard at that level. > 4. No PTP support is required. My timing needs are IT related - log > consistency, security protocol requirements, etc. > 4. I'm not interested in building and using something like a Raspberry Pi. > I'll be sending these to remote sites and don't want to worry about issues > like memory card failures and the hobbyist feel of many Pi cases and > cooling options. That's fine for my home but not for work. > 5. Availability is important. > > Any suggestions are welcome. > > Thanks, > > Paul Watts. > ___ > time-nuts mailing list -- time-nuts@lists.febo.com -- To unsubscribe send > an email to time-nuts-le...@lists.febo.com > To unsubscribe, go to and follow the instructions there. > ___ time-nuts mailing list -- time-nuts@lists.febo.com -- To unsubscribe send an email to time-nuts-le...@lists.febo.com To unsubscribe, go to and follow the instructions there.
[time-nuts] Re: GPSDO Control loop autotuning
Hi Tobias,
Using the excellent material found here:
https://nbviewer.ipython.org/github/rlabbe/Kalman-and-Bayesian-Filters-in-Python/blob/master/table_of_contents.ipynb
I went ahead and implemented the first order filter as described in
chapter 8.
The simulation was changed to use a real recorded PPS as measurement
noise and a real recorded TIC of a TCXO as the to be estimated data and
a Vtune steering an added frequency
The STDEV of the PPS was 5.6e-9
The kalman filter predicts the phase and frequency fairly well
STDERR of phase is 4e-9
STDERR of freq is 1.75e-9
The loop was then closed with a PI controller tasked to minimize the
phase error.
Tuning was a bit of a problem but the current performance is
STDEV phase error 1.25e-8
STDEV frequency error 1.2e-9
Unfortunately the resulting ADEV is still worse compared to a well tuned
PI controller acting directly on the TIC of the TCXO.
More work to do
Erik.
On 21-4-2022 9:44, Pluess, Tobias via time-nuts wrote:
Hi Erik,
yes I also found this webpage, and the examples are good.
I had the exact same idea as you, except I thought that the model for the
VCO output is either phase only or phase and frequency. Yours is even a bit
more sophisticated, thats quite cool!
And yes, of course it will be possible to have a one dimensional filter,
even if your state space equations have more than one state. However, in
this case, A, B and C are a matrix and two vectors (sometimes also called
F, G and H). But the output that is measured ("observed") is just scalar,
i.e. 1D.
I am not sure whether your state extrapolation is correct, I don't see it
at the moment unfortunately. But I will think about it a bit more.
To your question, about how to close the loop.
Well. What do you do when you don't have the alpha-beta-gamma filter?
you take your phase error readings and feed them into some sort of
controller, for example a PID controller, do you.
Now think of the alpha-beta-gamma filter (or also the Kalman filter) as
some sort of "prefilter". (do you remember when you asked me why I
prefilter my 1PPS measurements? :-D at that time, I had a 2pole IIR filter
with adjustable time constant as prefilter). So you feed your actual 1PPS
measurements into the alpha-beta-gamma filter, and the filter outputs the
1PPS measurements, but with noise mostly removed (hopefully!). Then, use
this as the error signal for your controller.
As how to choose the alpha, beta and gamma coefficients, you need to know
the variance of your input signal. i.e. determine the standard deviation of
the 1PPS readings and square that. I am not sure at the moment how to find
the coefficients from the variance, but I believe somewhere in the
kalmanfilter.net tutorial, there is an equation for it if I remember
correctly.
Also worth mentioning is the fact that some GPS modules also output an
estimate of the timing accuracy. This can also be used as an additional
input to calculate the coefficients and adapt them, as receiving conditions
change.
does that help a bit?
I am unfortunately not yet so far as you, I have made some calculations
using matlab, but I am not sure whether they are correct. I do see some
filtering effect of the Kalman filter, though. (I used the Kalman filter,
with the error covariance matrices, not the alpha-beta-gamma filter, like
you did, but maybe this is a better idea!). Are you interested in sharing
your excel sheet?
thanks,
best
Tobias
HB9FSX
On Wed, Apr 20, 2022 at 1:18 AM Erik Kaashoek wrote:
Tobias,
Your Kalman post triggered me to study this "dummy" level introduction
to Kalman filters: https://www.kalmanfilter.net/
and now I'm trying to write the Kalman filter.
An example of a Kalman filter described in the link above uses distance,
speed and acceleration as predictors for a one dimensional Kalman filter
So I wondered if it would be possible to do the same for the phase, eg
use phase, frequency and drift in the prediction model and still have a
one dimensional model?
The Kalman filter thus uses the normalized phase as input (the measured
time difference between the 1 PPS and the 1 second pulse derived from
the VCO output) to predict the next phase, frequency and drift
p = phase
f = frequency (d Phase/d T)
d = drift (d Frequency / d T)
m = measured phase
with a 1 second interval between observations the state extrapolation
(or prediction) is
p[k+1] = p[k] + f[k] + 0.5*d[k]
f[k+1] = f[k] + d[k]
d[k+1] = d[k]
and the state update is
p[k] = p[k-1] + alpha*( m - p[k-1])
f[k] = f[k-1] + beta*(m-p[k-1])
d[k] = d[k-1] + gamma*(m-p[k-1])/0.5)
In a small simulation (in excel) I used a VCO that initially is on the
correct frequency, after some seconds Vtune has a jump and after some
more seconds the Vtune starts to drift.
The Vtune of the VCO can have noise(system error) and the m can have
noise (measurement error).
The p,f and m nicely track the input values even when measurement noise
is added but after this I'm stuck as I do not yet understand how to
calculate the
[time-nuts] Re: GPSDO Control loop autotuning
Hi Erik,
yes I also found this webpage, and the examples are good.
I had the exact same idea as you, except I thought that the model for the
VCO output is either phase only or phase and frequency. Yours is even a bit
more sophisticated, thats quite cool!
And yes, of course it will be possible to have a one dimensional filter,
even if your state space equations have more than one state. However, in
this case, A, B and C are a matrix and two vectors (sometimes also called
F, G and H). But the output that is measured ("observed") is just scalar,
i.e. 1D.
I am not sure whether your state extrapolation is correct, I don't see it
at the moment unfortunately. But I will think about it a bit more.
To your question, about how to close the loop.
Well. What do you do when you don't have the alpha-beta-gamma filter?
you take your phase error readings and feed them into some sort of
controller, for example a PID controller, do you.
Now think of the alpha-beta-gamma filter (or also the Kalman filter) as
some sort of "prefilter". (do you remember when you asked me why I
prefilter my 1PPS measurements? :-D at that time, I had a 2pole IIR filter
with adjustable time constant as prefilter). So you feed your actual 1PPS
measurements into the alpha-beta-gamma filter, and the filter outputs the
1PPS measurements, but with noise mostly removed (hopefully!). Then, use
this as the error signal for your controller.
As how to choose the alpha, beta and gamma coefficients, you need to know
the variance of your input signal. i.e. determine the standard deviation of
the 1PPS readings and square that. I am not sure at the moment how to find
the coefficients from the variance, but I believe somewhere in the
kalmanfilter.net tutorial, there is an equation for it if I remember
correctly.
Also worth mentioning is the fact that some GPS modules also output an
estimate of the timing accuracy. This can also be used as an additional
input to calculate the coefficients and adapt them, as receiving conditions
change.
does that help a bit?
I am unfortunately not yet so far as you, I have made some calculations
using matlab, but I am not sure whether they are correct. I do see some
filtering effect of the Kalman filter, though. (I used the Kalman filter,
with the error covariance matrices, not the alpha-beta-gamma filter, like
you did, but maybe this is a better idea!). Are you interested in sharing
your excel sheet?
thanks,
best
Tobias
HB9FSX
On Wed, Apr 20, 2022 at 1:18 AM Erik Kaashoek wrote:
> Tobias,
> Your Kalman post triggered me to study this "dummy" level introduction
> to Kalman filters: https://www.kalmanfilter.net/
> and now I'm trying to write the Kalman filter.
> An example of a Kalman filter described in the link above uses distance,
> speed and acceleration as predictors for a one dimensional Kalman filter
> So I wondered if it would be possible to do the same for the phase, eg
> use phase, frequency and drift in the prediction model and still have a
> one dimensional model?
> The Kalman filter thus uses the normalized phase as input (the measured
> time difference between the 1 PPS and the 1 second pulse derived from
> the VCO output) to predict the next phase, frequency and drift
> p = phase
> f = frequency (d Phase/d T)
> d = drift (d Frequency / d T)
> m = measured phase
> with a 1 second interval between observations the state extrapolation
> (or prediction) is
> p[k+1] = p[k] + f[k] + 0.5*d[k]
> f[k+1] = f[k] + d[k]
> d[k+1] = d[k]
> and the state update is
> p[k] = p[k-1] + alpha*( m - p[k-1])
> f[k] = f[k-1] + beta*(m-p[k-1])
> d[k] = d[k-1] + gamma*(m-p[k-1])/0.5)
>
> In a small simulation (in excel) I used a VCO that initially is on the
> correct frequency, after some seconds Vtune has a jump and after some
> more seconds the Vtune starts to drift.
> The Vtune of the VCO can have noise(system error) and the m can have
> noise (measurement error).
> The p,f and m nicely track the input values even when measurement noise
> is added but after this I'm stuck as I do not yet understand how to
> calculate the alpha, beta and gamma values from the observed measurement
> error as the above website does not yet explain how to do this.
> And I do not understand how to close the loop to either keep the phase
> error or frequency error as low as possible.
> So much to learn.
> Erik.
>
> On 16-4-2022 22:44, Pluess, Tobias via time-nuts wrote:
> > Hallo all,
> >
> > In the meantime I had to refresh my knowledge about state-space
> > representation and Kalman filters, since it was quite a while ago since I
> > had this topic.
> >
> > So I looked at the equations of the Kalman filter. To my understanding,
> we
> > can use it like an observer, and instead of using the phase error and
> > feeding it to the PI controller, we use the output of the Kalman filter
> as
> > input to the PI controller. And the Kalman filter gets its input from the
> > phase error, but we also tell it how much variance this phase error has.
> >
[time-nuts] NTP server source needed
Long time lurker, first time poster. In the past, I've purchased the LeoNTP server to act as a local time server on my work network. Unfortunately, those NTP servers aren't available to order. I'm in need of two - three more NTP servers for various locations. Do any of you have an alternate NTP server recommendation? 1. I'd like to keep the price less than $1,000. The Leo's were a great, self contained choice at less than $400. 2. Support for IPv4 is required and IPv6 is desired. The main shortcoming I have with the Leo's is no support for IPv6. 3. I'll serve about 150 clients per site from a given NTP server. My NTP clients are a combination of Windows and Linux computers along with various types of network gear (Cisco, Palo Alto, Aruba, Juniper). The Leo's don't even breathe hard at that level. 4. No PTP support is required. My timing needs are IT related - log consistency, security protocol requirements, etc. 4. I'm not interested in building and using something like a Raspberry Pi. I'll be sending these to remote sites and don't want to worry about issues like memory card failures and the hobbyist feel of many Pi cases and cooling options. That's fine for my home but not for work. 5. Availability is important. Any suggestions are welcome. Thanks, Paul Watts. ___ time-nuts mailing list -- time-nuts@lists.febo.com -- To unsubscribe send an email to time-nuts-le...@lists.febo.com To unsubscribe, go to and follow the instructions there.
[time-nuts] Re: EFTF/IFCS 2022 - time-nuts meetup?
Hi Attila and Markus, I wobei at the Symposium in Paris together with our Chief Engineer Henry. We will arrive on Monday afternoon and will leave Thursday at noon time. Will be glad to meet you both - and more folks. Best regards Bernd sent by Mobile phone Am 20.04.2022 20:47 schrieb Attila Kinali : Hey there! As some of you might be aware, next week is EFTF/IFCS in Paris. Magnus and I will be there and we were wondering if any other time-nuts would be present. Attila Kinali -- The driving force behind research is the question: "Why?" There are things we don't understand and things we always wonder about. And that's why we do research. -- Kobayashi Makoto ___ time-nuts mailing list -- time-nuts@lists.febo.com -- To unsubscribe send an email to time-nuts-le...@lists.febo.com To unsubscribe, go to and follow the instructions there. ___ time-nuts mailing list -- time-nuts@lists.febo.com -- To unsubscribe send an email to time-nuts-le...@lists.febo.com To unsubscribe, go to and follow the instructions there.
