[time-nuts] Re: Evaluating counter performance, seeking feedback.

[Erik Kaashoek]

Tom, others.
An important reason for me to write these posts is to try to try to 
understand what I am doing. writing helps me to discover flaws in my 
thinking. If its too much feel free to block me.
To test systematic errors when measuring the ratio between two 
frequencies a dual channel signal generator was used. One channel set at 
static 10MHz, the other channel set to sweep +/- 500 Hz around 10MHz.
Both channels where send to the two inputs of the U6200A and the 
tinyCount (own build counter). Both counters measured the ratio between 
the two frequencies with a 1 second gate time. Both counters used their 
internal reference for measuring the ratio.
Timelab was used to record the frequency difference [1][2] and after 
measuring one full sweep and removing the global linear trend a very 
weird pattern appeared [3]. Both the U6200A and the tinyCount had a 
cyclic deviation from the global linear trend depending on the 
difference in frequency. For a span of 1kHz the U6200A showed 1.5 cycle 
up to 1.3e-9 error and the tinyCount  25 cycles of up to 2.5e-9 error.
When measuring two static 10 MHz signals with only a small frequency 
difference the U6200A and the tinyCount agreed within their measurement 
noise.
Further testing showed the amount of cycles in the full sweep was 
independent of the sweep span. Reducing the span with a factor 10 also 
reduced the maximum deviation with a factor of 10. Reducing the sweep 
time did change the amount of cycles in but not with a linear relation 
to the change in sweep time.
As the tinyCounter uses regression the amount of points used for the 
regression was changed from 20000 to 50000 but this did not create any 
visible difference in the frequency deviation cyclic pattern.
Some further work is required to understand the cause of these cyclic 
deviations.
Erik.

[1] http://athome.kaashoek.com/time-nuts/1kHz%20sweep%20U6200A.tim
[2] http://athome.kaashoek.com/time-nuts/1kHz%20sweep%20tinyGTC.tim
[3] http://athome.kaashoek.com/time-nuts/1kHz%20sweep.png

On 8-4-2022 21:11, Tom Van Baak wrote:
> Thanks for posting both TIM files along with the plots. Here are 
> replies to 3 of your questions:
>
> > Is this setup meaningful in assessing performance differences?
> > If not, how to improve?
>
> It's probably better to compare two different counters using the same 
> setup. So give them both the same Rb as ext REF and give them both the 
> same DUT and then collect data simultaneously (apples and apples). But 
> it sounds like you can't use or don't have an ext REF input for your 
> DIY counter? In that case, right, you have to resort to the unusual 
> arrangement that you're using (apples and oranges). This is one reason 
> why almost every frequency counter has an external REF input.
>
> > How can one compress or expand a TIM file to correct for the 
> difference in gate time?
> > A better approach would be to ensure gate times where identical.
>
> Right, I noticed the two TIM files don't line up. That's a problem. 
> They are off by several seconds at the beginning of the run and spot 
> on at the end of the run. I suspect some manual editing? Note this 
> doesn't affect the ADEV plots, but it messes up the phase and 
> frequency plots. It's easy to fix.
>
> It looks like you didn't input the correct sample rate when you loaded 
> the data into TimeLab. There's a box in the acquisition menu to set 
> the sample rate. Normally close enough is good enough, but when you're 
> working with simultaneous data you need to be much more precise. This 
> isn't a problem with timestamp data because the actual sample rate is 
> implicit in the timestamp. And it's not a problem for zero-dead-time 
> frequency measurements either because one of the clocks does the 
> pacing. But for traditional gated measurements, yes, sample rate may 
> be inexact, and may also vary depending on the measurement data or 
> auto trigger settings.
>
> A while ago I collected simultaneous data on a several oscillators for 
> many days. When my PC reads serial data from a counter I always prefix 
> lines with a MJD timestamp [1]. Days or years later it tells me when I 
> did the experiment. It can also be used to detect gaps in the data. It 
> also makes it easy to make x-y scatter plots using MJD as an axis. It 
> allows multiple runs to be correlated (e.g., environmental data on one 
> PC, counter data on another PC, GPS data from a third PC, etc.). But 
> most importantly it allows me to compute the actual sample rate, the 
> tau, for any data that I ever collect. For example 5 "identical 
> counters" set for 10 s gate time had actual sample rates of:
>
> 10.3475 s
> 10.3496 s
> 10.3496 s
> 10.3494 s
> 10.3475 s
>
> This doesn't matter for short runs, doesn't matter to ADEV, doesn't 
> matter for phase or frequency plots of one counter, but matters a lot 
> when you have multiple counters over an extended period of time.
>
> > How can one use the two TIM files to calculate the RMS of the 
> differences in frequency?
> > My hope is to use this RMS calculation as a single number quality 
> indicator.
>
> Perhaps explain more what you're trying to do. Remember that frequency 
> depends very much on the averaging time so you can't just use a single 
> number. ADEV works because it's always ADEV(tau). There is a special 
> case when ADEV measurements are strictly linear, usually with a slope 
> of -1 or -1/2. Then it is customary to use a single number. For 
> example 1e-9/tau for 1 ns of WPM, or 1e-6/√tau for 1 ppm of WFM.
>
> /tvb
>
> [1] see comcat1 and comcat2 in my www.leapsecond.com/tools/ directory.
>
>
> On 4/7/2022 12:44 AM, Erik Kaashoek wrote:
>> To better understand the performance of a home build counter a 
>> comparison was done with a Picotest U6200A
>> The two channel home build counter was setup to measure the frequency 
>> of the 10MHz output from a Rb on one channel and the 10MHz output 
>> from a not so good OCXO on the other channel.
>> The ratio between the two frequencies was measured with a 1 second 
>> gate time, multiplied by 1e+7 and send to Timelab.
>> The U6200A had the Rb output as 10MHz reference and the 10MHz from 
>> the OCXO into channel 1. Gate time was also set to 1 second.
>> In Timelab the data from the Counter under test and the U6200A where 
>> recorded simultaneously over a 1000 second period
>> The recorded data was saved and adjusted for the difference in start 
>> time of the measurements and loaded back into Timelab.
>> U6200A TIM file http://athome.kaashoek.com/time-nuts/U6200A.tim
>> Own counter TIM file http://athome.kaashoek.com/time-nuts/tinyGTC_2.tim
>> A first performance check was done by plotting the unwrapped linear 
>> residue of the phase of both measurements. (see: 
>> http://athome.kaashoek.com/time-nuts/tinyGTCvsU6200A_Phase.png ) . 
>> The measurements did show some differences in instantaneous phase but 
>> the differences where small and even at 980 seconds the two 
>> measurements agree rather well.
>> A second performance check was done using the frequency plot. (see 
>> http://athome.kaashoek.com/time-nuts/tinyGTCvsU6200A_freq.png ). 
>> Overall the two measured frequencies agreed with sometimes up to 
>> 1e-10 difference. The difference in gate time of the two counters was 
>> very visible as a gradual shift. As the measurements where aligned in 
>> time at the end of the measurement the time difference at the start 
>> was about 3 seconds.
>> A detailed plot of the measured frequencies over the last 100 seconds 
>> (see 
>> http://athome.kaashoek.com/time-nuts/tinyGTCvsU6200A_freq_detail.png 
>> ) showed an occasional difference between the frequency measurements 
>> of the two counters up to 2e-10
>> Questions:
>> 1: Is this setup meaningful in assessing performance differences? If 
>> not, how to improve?
>> 1: How can one compress or expand a TIM file to correct for the 
>> difference in gate time? A better approach would be to ensure gate 
>> times where identical.
>> 2: How can one use the two TIM files to calculate the RMS of the 
>> differences in frequency?  My hope is to use this RMS calculation as 
>> a single number quality indicator.
>> Erik.
>>
> _______________________________________________
> time-nuts mailing list -- time-nuts at lists.febo.com -- To unsubscribe 
> send an email to time-nuts-leave at lists.febo.com
> To unsubscribe, go to and follow the instructions there.