[ntp:questions] NTP recommendations for mid size network

G9 - Nuno Pereira nuno.pereira at g9telecom.pt
Fri Jan 23 13:14:06 UTC 2015


Hello,

 

Introduction:

At our company we have 50+ hosts, most of them are virtual (Xen) servers,
having a limited number of physical machines available to be an NTP server.

We're a small telecom company and we prefer to have the same time in all of
our hosts (time island), rather than an absolutely accurate time.

 

Current configuration:

Our normal config is with 2 prefered IPs (same server, not dedicated to NTP)
and 2 non-prefered external hosts (either from the pool or some selected
server). In some cases there aren't external hosts.

We just use server peers: no peer connections, no multicast and no broadcast.

This NTP server is also used by our internet clients.

 

We monitor 3 parameters of the NTP service: offset, reference peer (server)
and active peers.

 

But this configuration has some problems for us:

-          Once in a while we get warning offsets, mainly when one of our
external peers goes insane and we get 30+ ms offsets, when we normally have
less than 1 ms, having spikes that can go to 10-15 ms. With this, our goal of
having "the same time in all of our hosts" is compromised;

-          Also with little frequency one or many of our hosts loose a
reference peer, like yesterday, when 2 external peers (in the pool) got an
offset of ~1000 ms, having many of our hosts having 2 good peers (our internal
server) and 1 or 2 rejected peers (with ~1000 ms offset);

-          Obviously, if we do some maintenance in our server, of if it has a
problem, and then we get some problems in our ntp network, having clients
without reference clock and/or with bad offset when the server returns to
life;

-          Our goal of having excellent synchronization between our hosts
isn't achieved.

 

 

The question:

Given that, or ignoring what I've said, what are your recommendations?

.         How many peers to configure for each client, at least? Only one (the
same for all)? or a higher number?

.         Should we keep preferred servers, having alternatives?

.         Should we use peer connections?

.         Are broadcast or multicast better options?

.         Could we use virtual servers as internal alternatives in the
network?

.         Pool servers aren't available everywhere, but are they good options?

 

Our server is configured with 4-5 delay stratum 1 external servers with 60-70
ms of delay (preferred), and 2 or 3 external servers with stratum 2+ with <5
ms of delay (not preferred).

Is that a good option?

 

Thank you all,

Nuno Pereira

Software engineer at G9Telecom

 



More information about the questions mailing list