[ntp:questions] NTP servers redundancy

unruh unruh at wormhole.physics.ubc.ca
Tue Jan 19 20:24:31 UTC 2010


On 2010-01-19, pc <Paul.Croome at softwareag.com> wrote:
> Jean,
>
> It's already been said (and others may say it again) that an NTP
> configuration with
> two servers is generally considered bad. That's true in the sense that
> a
> two-server setup won't give you better time than a single-server
> setup,
> and may indeed give you slightly worse time; you would have to do some
> testing to see if that is an issue for you. In many cases, it's
> sufficient
> if the time is accurate to plus/minus a second or two, and your setup
> should certainly achieve that. Using two (or more) servers is a good
> idea for achieving redundancy in case a server fails.

It would still give slightly better time than one. The statistical
uncertainties will cancel. The problem with 2 is that if one of the
servers goes nuts and its time deviates from true time, there is no way
the client can decide which of the two servers is the better claock.
With three, it can use majority voting to decide which is bad. But if
one goes off the air, then it is clear that it is bad. So 2 is bad
against a rogue clock, but OK for redundancy against total failure of
one. 
With one or two the time at the servers should be good to a few usec. I
used to say 1usec from my experience, but on Christmas eve my one
suddenly began to experience +- 10usec periodic fluctuations. Since the
temp of the motherboard also seems to have experienced similar
fluctuations with similar time period, it may be that that is the cause,
but it is stange that this only began Christmas eve. So lets say +- 10
usec. the clients will then be able to do about +- 20-30 usec over the
network ( assuming no weird assymmetric network behaviour)

So, you need to decide which is the more likely failure mode. One server
going off air ( crash, etc) or one server going rogue ( suddenly
delivering the wrong time, but with great authority). 

>
> Paul




More information about the questions mailing list