The scores page (e.g. <http://www.pool.ntp.org/scores/103.4.109.106>) states 
that:

"The monitoring system [...] is more susceptible by random network latencies 
between the server and the monitoring system than a regular ntpd client would 
be. In other words: Don't be alarmed by the occasional large offset and please 
don't use the offset as an absolute performance metric."


When I send an ntp-request to some server and get an answer, I can compute the 
offset between my computer and the real time, which is an interval with the 
length of the round trip time (minus time spend inside the server).

When I do this for several ntp-servers (A) I get something like:

(this should be viewed in a mono-spaced font - the "===" are meant to be 
aligned)
      ----=======-----
              =======-                   
                 -======= 
                 -=======---
   -------=======------         
                 
This means that the offset between my computer and the real time is somewhere 
in the [=======] interval, which is the intersection of all [--...--] intervals.


Now when I query another ntp-server (B) and get:
         =======        (intersection of all A-servers)
                            -------------       (server B)      
Then I know that:
(1) either all (A) servers are bad.

This might happen, if all depend on the same bad stratum 1 server.
So assume this is not the case: they come from different countries, which all 
have their own stratum 1 server.
Let's also assume that the likelihood of several stratum 1 servers being all 
bad (and all by about the same amount) is very close to zero.

In this case we necessarily have:
(2) server B is almost certainly bad and should removed from the pool at once.


This test does not use offset values (which are indeed susceptible to "random 
network latencies") but offset intervals which are not.
There is no real need for a "Score" value - either the offset-interval of a 
server intersects with the intersection of the offset-intervals of all other 
servers (ok), or it does not (bad).


Kind regards,

Gerriet.


P.S. I don't know if the mailing list supports attachments; but here is a 
picture which illustrates my case with real data:

The intersection between all good (blue) servers is between the green vertical 
lines (ca. 35 msec wide).
The bad (brown) server 103.4.109.106 / ntp0.1asiacom.net does NOT intersect the 
green interval.

Each server is queried 4 times.
_______________________________________________
pool mailing list
[email protected]
http://lists.ntp.org/listinfo/pool

Reply via email to