[time] outlier

Tim Shoppa shoppa
Thu Nov 2 12:24:05 UTC 2006

Kiss Gabor <kissg at niif.hu> wrote:

> >> Could you explain me what the word 'outlyer' means in all NTP related
> >> documents? Should I worry about remote servers marked with space in
> >> output of "ntpq -p" or not?
> > It's where one system is grossly different to all the rest.  Suppose that you 
> > had 10 server.  Your offsets with 9 servers are in the range -0.1s to +0.1s, 
> > but one server shows an offset of 0.9s.  It would be the "outlier" -
> Look this:
>       remote           refid      st t when poll reach   delay   offset  jitter
> ==============================================================================
> *host1           .GPS.            1 u  202  256  377    0.553   -1.249   0.343
> +host2           .GPS.            1 u  182  256  377    3.904   -0.879   0.073
> -host3           xxx.xxx.xxx.xxx  2 u    9  256  377   33.926   -0.450   0.152
> +host4           xxx.xxx.xxx.xxx  2 u    8  256  377   39.030   -1.050   0.252
> -host5           xxx.xxx.xxx.xxx  2 u  202  256  377    0.564   -0.947   0.152
> Host3 and host5 are marked as outliers.
> Why?
> Host4 has worse paremeters.

Ignore what others tell you about the classic textbook definition of
outlier. NTP's server selection algorithm (see it in the sourcecode,
ntp_loopfilter.c and ntp_proto.c) will by design often throw out a host
or two as "outlyers" because they just aren't as good as the rest.
They don't have to be grossly out of range, they just have to be
not quite as good (unless you follow the algorithm in very close
detail you may think that they are actually as good or better than,
but the algorithm thinks differently!) as the other ones.

Then it'll poll them again and change its mind again, picking
a whole different set of (apparently good) servers to throw out as

In this case, NTP can afford to be picky.

We had extended discussions on this a few years back on this very
same mailing list, when it came to server selection algorithms.


More information about the pool mailing list