[ntp:questions] help in evaluating NTP servers performance

Dennis Hilberg Jr dont at spam.me
Mon May 14 17:18:38 UTC 2007


themule wrote:
> Hello,
> I have 3 servers operating at stratum-2, providing time to a small
> population of clients (most routers and stratum-3 servers that
> redistribute time to their LANs). The only source of time sync are
> public NTP servers, and the configuration is based on this document:
> http://www.eecis.udel.edu/~mills/ntp/html/notes.html. Each server has
> 2 stratum-1 sources (total of 6 distinct servers) and 4 stratum-2
> peers (the other 2 internal servers plus 2 external ones). The reason
> for this is mostly the ability to simple disable one external server
> when it goes off-sync for some time and still have 3 external sources.
> 
> Now, I know that performance is quite a subjective matter. All we need
> is to keep lan servers and clients in sync, with most timestamps
> having 1 sec resolution. Currently I see offsets of about +/- 10ms on
> ours main NTP servers, with occasional peaks, so let's say performance
> is "good enough" for us.
> 
> But sometimes I wonder how good is that on an absolute scale, just out
> of curiosity. And of course if it can be somehow improved. I've
> plotted graphs from peerstats of the three main servers (ntp1, ntp2,
> ntp3), and what really surprises me is that the three servers show
> quite different patterns (hardware _is_ quite different tho).
> 
> May someone more experienced than me have a quick look at the graphs
> and provide a couple of comments on them? They are here: http://stats.esiway.net/NTP/
> 
> I'd like to know how they compare to, say, similar stratum-2 servers.
> I've found a few other graphs of running servers on the Internet, but
> I can't make real comparisons since they are either stratum-1 servers
> or stratum-2 servers right next (same LAN) to a stratum-1 server, and
> of course their accurancy is orders of magnitude better.
> 
> TIA,
> .TM.
> 

I operate two stratum 2s and graph the variable data and client counts, 
which you can view here and make your comparisons: 
http://saturn.dennishilberg.com/ntp.html.  The source for the variables is 
'ntpq -c rv'.  The source for the clients is ntp_clients_stats 
(http://www.schlitt.net/scripts/ntp/index.html).

I got the graphs finalized about a month ago, and after watching the plots I 
was slightly disappointed at their performance initially.  Based on what I 
read I assumed millisecond to sub-millisecond accuracy, consistently. 
However, both server's internet connections are asymmetric, the room's 
temperature can range ± 5°F or more throughout the day, both run a Linux OS, 
and are PC based.  From what I've read/heard that's about as good as it's 
going to get, considering the environment.

-- 
Dennis Hilberg Jr <timekeeper(at)dennishilberg(dot)com>

NTP Server Information <http://saturn.dennishilberg.com/ntp.html>




More information about the questions mailing list