[c-nsp] NTP network design considerations

[Saku Ytti]

On Fri, 14 Oct 2022 at 23:32, Gert Doering via cisco-nsp
<cisco-nsp at puck.nether.net> wrote:

> For a true time geek, the time the rPIs provide is just not good
> enough (fluctuates +/- 20 usec if the rPI has work to do and gets
> warm) :-)

Meinberg does not do HW timestamping, nor does NTP. Almost every NIC
out there actually does support HW timestamping, but you'd need chrony
to actually enable the support.

When using Meinberg and clocking your host, ~all of the inaccuracy is
due to SW timestamping, oscillator doesn't matter. Basically with NTP
you are measuring the host context switch times in your jitter. This
is because networks have become organically extremely low jitter
(because storing packets is expensive).
We see across our network single digit microsecond jitters (in my M1
computer I see loopback pings having >50us stddev in latency), and
because the timing we use is SW timestamp, our one-way delay
measurement precision is measuring the NTP host kernel/software
context switch delays.
The most expensive oscillator would do nothing for us, but HW
timestamping and cheap 2eur OXCO would radically improve the quality
of our one-way delay measurements.


-- 
  ++ytti