Recommended Number of NTP Servers

Achim Gratz Stromeko at
Mon Nov 4 18:35:17 UTC 2019

Richard Laager via devel writes:
> The default for maxclock is 10. This includes the "pool" entries, of
> which the stock Debian configuration has four:
> pool iburst
> pool iburst
> pool iburst
> pool iburst
> As a result, the default is to use 6 pool servers.

That is a cargo-culted configuration from an earlier (buggy) pool
implementation in NTP classic that didn't mobilize enough pool servers
when only a single pool line was present.  With recent NTPd you really
only need one pool ine and (especially if you have multiple server lines
as well) a modified tos setup so that the selection algorithm does not
kill all of them.  Here's what I currently use:

tos minsane 7 minclock 15 maxclock 35 ceiling 4

I have around ten fixed servers (most of them internal), so that setup
leaves me with 5 internal and 10 external (mostly from the pool)
candiodate servers when locked to the GPS refclock.  The pool statement
initially provides about twice as many servers as the surviving set
finally contains and it does continue to weed out bad clocks and
mobilize new ones if necessary (doesn't happen often, but it's nice to
see that it works).

> If that's true, then it seems like odd numbers of servers are better,
> all things being equal. From there, it seems like I should set maxclock
> to either 11 (so 11-4 = 7 servers) or 9 (so 9-4 = 5 servers). Any flaws
> in this logic?

No, at least if all your servers come from the pool.  If not, the
selection algorithm will prefer closer clocks with lower stratum (those
on the LAN with their own refclock) anyway.  Failover from a single
local server going away is much smoother if you have at least five of
them, I'd think that this should hold for pool-only setups.

> Any thoughts on whether 7 servers or 5 servers is better? The same page
> says, "Conventional wisdom is that using at least five upstream time
> servers would probably be a good idea, and you may want more." The
> conservative choice is 7, but is that likely to make a meaningful
> difference compared to 5?

Depends on how uniform the routing is across your pool.  But I'd
personally go for more than five, that makes it much more unlikely to
get dealt an initial set that is off in the same way and thus looks good
to NTP while it actually isn't.

+<[Q+ Matrix-12 WAVE#46+305 Neuron microQkb Andromeda XTk Blofeld]>+

SD adaptation for Waldorf rackAttack V1.04R1:

More information about the devel mailing list