Hi, On 09/08/2005 01:34 PM, Brad Knowles wrote: > At 1:28 AM -0700 2005-09-08, Ask Bjørn Hansen wrote: >> I thought 3 servers were supposed to be enough to protect against >> a false-ticker? > No. 3 is the absolute minimum you need to be able to reliably > select a syspeer, but if-and-only-if all three clocks are "sane". You > need at least four to protect against a single falseticker, five to > protect against two false tickers, etc....
If all are sane, you can do with one, can't you? As long as you select a sane one, it's OK. Or do I understand 'sane' incorrectly? > See > <http://ntp.isc.org/bin/view/Support/SelectingOffsiteNTPServers#Section_5.3.3.> > for details. Each time I read about this, I start wondering. From my knowledge of measurement in general, if you do one measurement, you get a result. If you do two or more independent measurements, and apply the correct math, you get a better result. With three and more measurements, you can detect out layers. One important piece of information for this to work with a few measurements is knowledge about the error margin for each measurement. The ntp-protocol replies contain some info that enables an estimate of the error, so I think all ingredients are there. I guess this is not really for a discussion on this list. So, I've put my way of thinking on http://www.xs4all.nl/~arnold/files/ntp.txt and I welcome any comments via the email address on that page. The main thing I wonder about is: is my way of thinking wrong, or is it the way ntpd is implemented (of which I have no detailed knowledge)? Kind regards, Arnold &:-) _______________________________________________ timekeepers mailing list [email protected] https://fortytwo.ch/mailman/cgi-bin/listinfo/timekeepers
