[ntp:questions] Re: Difference between PPS and PPS plus kernel discipline?

Ulrich Windl Ulrich.Windl at RZ.Uni-Regensburg.DE
Mon Mar 27 14:03:09 UTC 2006


jra at febo.com (John Ackermann N8UR) writes:

> Today I changed the configuration file of two of my NTP servers to
> enable the kernel timekeeping discipline, and have put plots of the
> effect of the change at
> http://www.febo.com/time-freq/ntp/stats/kernel_discipline/index.html.

The graphs could benefit from better labels; it's hard to guess what was
producing the output shown.

> 
> The data is derived from the offset and dispersion fields of the
> peerstats file.
> 
> When using kernel discipline, the offset and jitter shown by the ntp
> stats are essentially 0, which makes sense since both the reference and
> the unknown time come from the same source; they should track each other
> closely.
> 
> When not using kernel discipline, there is a wander of several hundred
> microseconds in one case, and several milliseconds in the other, and
> significantly more jitter in both cases.
> 
> My question is about that wander -- is it showing the (in)stability of
> the crystal on the motherboard, *or* is it simply showing that NTP by
> itself doesn't steer the system time as tightly as the kernel
> discipline?  Put another way, can I make any observations about the
> quality of the motherboard crystal from this data, or am I looking at
> NTP's PLL performance?

Good question. It could be also the bugs of the time-keeping code.
I'd suggest to start with this one: Echo back the PPS-pulses and record the
delay. This gives you interrupt delay and jitter. Then (at the same time
maybe) record the timestamps seen for the PPS pulses (using PPS API) and
compare. If there is additional jitter, the software clock adds these
(assuming the software clock is left alone). Then you know what the free
running software clock adds to the PPS pulses coming in. Then the rest is
basically NTP math.

I could be wrong however. Any experts?

> 
> Bottom line, I'm trying to explain what these plots tell me, other than
> the fact that I need an external reference to evaluate the performance
> of the NTP/PPS system.
> 
> Thanks,
> 
> John
> _______________________________________________
> questions mailing list
> questions at lists.ntp.isc.org
> https://lists.ntp.isc.org/mailman/listinfo/questions




More information about the questions mailing list