[ntp:hackers] A stop-gap authenticated time service

Miroslav Lichvar mlichvar at redhat.com
Wed Nov 11 14:24:10 UTC 2015


On Wed, Nov 11, 2015 at 08:32:05AM +0000, Poul-Henning Kamp wrote:
> >What has changed significantly since the old days is the network
> >jitter. It went down by orders of magnitude, so the Allan intercept is
> >much shorter than before and ntpd is too slow now.
> 
> Do you have measurements to back this up, and if so, are they
> representative of the median network paths rather than one particular
> installation ?

I don't have any data from the time when NTP was developed. My
impression from what I have read about NTP is that jitter in hundreds
of milliseconds was considered good back then.

>From user's reports and my own observations I'd say these days a
typical jitter between an NTP client and its servers on the internet
is in single digit milliseconds.

As a data point, I can show you a distribution of jitter as measured
in the last few weeks on a server I use to monitor the pool.ntp.org
servers. This includes about two thousand servers spread across the
globe. The median jitter is about 5 milliseconds.

https://mlichvar.fedorapeople.org/tmp/ntp_pool_jitter.png

The pool DNS uses geolocation, so a typical client should be using
servers that are much closer than an average pool server, have a
smaller delay and also jitter. When I limit the selection to the few
tens of servers in the Czech zone, which is a fairly small area, the
median jitter goes down to 1.8 milliseconds. When I include also
Germany, Austria, Slovakia and Poland to cover a larger area, it's
about 3.2 milliseconds.

Do you see something significantly different in your data?

-- 
Miroslav Lichvar


More information about the hackers mailing list