On 2011-04-22, Roger <[email protected]> wrote: > On Fri, 22 Apr 2011 15:19:31 GMT, unruh ><[email protected]> wrote: > >>On 2011-04-22, Roger <[email protected]> wrote: >>> On Thu, 21 Apr 2011 09:32:21 -0400, "Richard B. Gilbert" >>><[email protected]> wrote: >>> >>>>One server: if it fails you have nothing! >>>>Two servers: If the two differ, which one do you believe? >>>>Three servers: degenerates too easily to the two server case. >>>>Four servers: Allows the failure of one server. >>>>Five servers: Allows the failure of two. >>>>Seven servers: Allows the failure of three. >>> >>> I've seen these number quoted before and I don't understand >>> the last one. Why doesn't 6 allow for the failure of 3? Why >> >>Because 3-3 is a tie and the system cannot decide which is best. Ie by >>failure, read "bad timekeepers". If 3 fail-- ie stop responding to >>packets, 6 is pleanty. 4 would be enough. But if they fail by delivering >>the wrong time, and all three deliver the same wrong time (say because >>all three are in Chicago and all three used a cell phone system to set >>the time and .... ) then you have a tie. >>It starts to get a bit absurd, I know. > > Thank you, and David and Dave. > > I hadn't thought about a 3-3 tie. I hadn't even considered that > that might happen. But if that is possible then so is a 2-2 tie > with 4 servers. Ho hum, nothing is perfect in this life.
That is why 4 is considered to be able to withstand one bad clock, not 2 and 6 gives 2 bad not 3. Ie, you figure out the chances of getting x bac tickers which all give the same bad time, and take at least 2x+1 servers if that chance is too large for you. Which is why the current pool.ntp.org ntpd policy of grabbing 10 sources suggests that they consider the chance of a pool giving a bad ticker to be about 13% (to give only a 1% chance of getting more than 4 false tickers) (In fact increasing the number from 9 to 10 clocks increases the chances of getting more than 4 false tickers. The argument that you need 2x+2 is obscure to me. ) _______________________________________________ questions mailing list [email protected] http://lists.ntp.org/listinfo/questions
