[time-nuts] NTP jitter with Linux

Steve . iteration69 at gmail.com
Thu Apr 5 06:11:57 UTC 2012


> I stand firm that the only proper way to do this is with a 100%
> deterministic architecture.


> ..only proper way to do what?  The goal is to discipline the internal
> software clock to GPS.  A typical application is a database server
> that is running a web e-comerse site so that transactions get time
> tagged.  So you would run a web, file or database server on a
> deterministic, no cache micro controller?
>

In the original message, Mike is trying to get the jitter better than 20e-6
on a pc. I don't believe this falls in the typical use of NTP category.
What i assume here is that the jitter is local to the machine sampling the
1pps signal. Otherwise this is moot.

 Numerically... 1/(20e-6) = 50e3,  likewise with nyquist variance (50e3 x2)
= 10e3, or time domain that is 100khz wide. This tells us that the input of
the pc must be rate monotonic to 100khz. The only thing that i can think of
that comes close to this monotic bw is a 192khz sound card, i have to
assume these cards are rated in raw sampling rate and not nyquist, so
correcting we see (192khz /2) = 96khz bw. Pretty close to 100khz. Maybe
close enough when considering both sides of figure contain the correction
for twice the highest bw.  Assuming i did not fat finger a calculation, a
192khz sound card is another possible solution to the problem.

The sound card approach would permit conversion to the frequency domain and
i think it would be fun to apply fourier analysis to something like a 1pps
signal. Does anyone have any thoughts as to how stable an approach like
this would be?

Steve


More information about the time-nuts mailing list