[ntp:questions] what happens when sys.peer turns stratum 16?
brian.utterback at gmail.com
Mon May 31 19:49:37 UTC 2010
When we talk about "insane" or "broken" clocks, the context is
important. What we are really talking about is what we call a
"falseticker", that is, a clock that does not have the actual time in
the range of its offset plus or minus its dispersion. So, with that
definition the amount by which it is "broken" could be very small, in
the range of milliseconds. Obviously, any clock that is further off
than the tolerance for error that your particular application has is
going to be a problem. If your tolerance is tighter than the range of
dispersions of your "truechimer" servers, then you will not be able to
reach your goal. If the tolerance is in the range, then any server
that moves your idea of the time outside the range is a problem and
"broken" as far as you are concerned.
As we saw before, the key is "independent" servers. We saw before that
4 servers is not enough to counter the failure of two servers if the
two servers are either not independent or happen to fail in the same
manner. In the same token, 3 servers are not enough to counter even a
single failure if the failure is in the right range. That is the real
reason for having four servers, not so that you will have three left
after the failure. Three is just not enough to outvote certain types
of single server failures.
More information about the questions