[SunHELP] NTP design

[Jim Pennino]

On Thu, Jul 18, 2002 at 06:27:43PM -0700, Tim Gallagher wrote:
> We are redesigning our NTP system and I was hoping to get some expert input
> from the list.
> 
> There are two geographically remote data centers, latency between the two
> locations is 50-60 ms (using ping -s) typically. The plan is to have a GPS
> netclock in each location, directly connected to two machines each - for a
> total of 4 stratum 1 NTP servers (running Solaris 8 with the bundled xntpd)
> operating as peers.
> 
> First thought is to have slightly different client files on the data center
> boxes at each location in order to designate one of the local servers as the
> preferred NTP server. This is a large environment.
> 
> The questions;
> 1- Does this seem like a good plan for use with two different locations?
> 2- Any problem to be expected with the servers syncronizing across the WAN?
> 3- Is it workable to have a different server preferred in each location?
> 
> Thanks,
> -- Tim
> _______________________________________________
> SunHELP maillist  -  SunHELP at sunhelp.org
> http://www.sunhelp.org/mailman/listinfo/sunhelp

A technical nit; if the GPS netclocks are network ntp devices, they will be
stratum 1 and the 4 solaris machines will be stratum 2, not that it will
make any difference in the long run.

Some general thoughts based on the results of doing something very similar,
but with 5 locations:

The only practical reason I've found for setting a preferred server is to
manage network traffic given the quality of ntp implementations and GPS
clocks these days. The ntp client should be able to sort out the
best time source by itself.

Setting all 4 Solaris boxes to point to both GPS boxes will keep everything
synced if 1 of the GPS boxes fails. Under normal circumstances, the Solaris
boxes will automatically pick the local GPS box as its preferred server.

Pointing local clients to servers across the WAN will only increase network
traffic, though only a little unless you have a HUGE number of machines,
but serve no purpose. The local clients need only point to the local Solaris
boxes because the clients can't get any better time then the local Solaris
can get anyway.

People get too fixated on stratum levels. For a given stratum 1 server,
a client at stratum 4 will have better time with a consistant, stable
chain back to stratum 1 than a stratum 2 client with a flaky link to the
stratum 1 server.

Consistancy of network latency to a server is in general more important
than the absolute value.

-- 
Jim Pennino