[ntp:questions] long linear network, time accuracy, and ntp strata

Miroslav Lichvar mlichvar at redhat.com
Tue Oct 19 14:52:07 UTC 2010

On Mon, Oct 18, 2010 at 04:49:47PM -0500, Miernik, Jerzy (Jerzy) wrote:
> I have a question about how to use ntp in a long linear network. Assume there are 50 nodes with the one most in the West having GPS, and every node running ntpd:
>   node1 --- node2 --- node3 --- ... --- node50
>   GPS
> Which approach to synchronization would be better in terms of the time accuracy in nodes:
> 1. configure 'server node1 ...' in nodes 2, 3, ..., 50; or
> 2. use manycast for automatic client / server self-organization?
> Approach 1 seems to me to lead to inferior accuracy due to increasing distance between a client in nodex and server in node1, where routing in nodes node2, node3, ..., node(x-1) would add to jitter. Yet we would have one node stratum1 and all others stratum2, it would seem...
> I would think approach 2 with manycast would end up in smaller offsets (if any) between clocks, but would lead to clocks of stratum 16, 17, ..., 50, and I am not sure how ntpd would react in node16, node17, ..., node50. 

Based on some simulations I did, I think the former approach will be
better in networks with low jitters and the latter (something like a
chain of NTP servers in groups of 4 nodes to stay below stratum 16) in
mid-high jitters. The jitter where this changes will depend on many
factors, it could be in order of hundreds of microseconds.

If you can describe the network properties, you could simulate
whole network in clknetsim (http://mlichvar.fedorapeople.org/clknetsim/).

Miroslav Lichvar

More information about the questions mailing list