On Mon, 16 Mar 2009, Brad Knowles wrote: > on 3/16/09 12:31 AM, [email protected] said: > >> actually, IMHO, you didn't explain very much at all in your article, you >> basicly stated that any competent person would know why having two NTP >> servers is worse than having one (with no information or links explaining >> your reasoning for this), > > The editors screwed up the article and didn't include my footnote. See > section 5.3.3 of the page at > <http://support.ntp.org/bin/view/Support/SelectingOffsiteNTPServers>. > >> P.S. there is a _very_ good reason for configuring systems to point at two >> NTP servers instead of one, > > Unfortunately, this is one of the more common misconceptions. It all has to > do with the way the NTP peer selection algorithm works. It never averages > numbers. It is designed to throw away the two outliers from the geometric > mean (the highest and the lowest), and then see who is left. It keeps doing > this until it gets down to one clock that is closer to the geometric mean > than any other.
I never claimed it would average them. > When you have only two upstream clocks, neither of them can be closer to the > mean than the other, and then you're off into LaLa Land where anything whacky > could potentially be done and probably is being done billions and billions of > times around the world on a daily or even hourly basis. define LaLa Land. >> it means that the clocks will remain synced >> much better if one NTP server dies rathat than having them all go off in >> their own directions. the time may not ever be as precise as the single NTP >> server case, but it's good enough for almost all real-world applications. > > Unfortunately, this claim is demonstrably wrong. And the fact that you claim > it is correct just goes to show the level of misunderstanding, > misconceptions, and misinformation that keeps getting propagated on a daily > basis. > > I'm sure you didn't intend to propagate a misconception, but that doesn't > reduce the damage you may have already done over the years by telling people > this same story, time after time. in practice this works. the clients will pick one of the two time servers and use it, but if it goes away they will use the other one. you can claim all you want that this doesn't work, but I have a datacenter worth of servers that are working with this. I define working as they all have the same time (within the bounds of my requirements), and I know that if I was not running NTP they would not for more than a few days. David Lang _______________________________________________ Tech mailing list [email protected] http://lopsa.org/cgi-bin/mailman/listinfo/tech This list provided by the League of Professional System Administrators http://lopsa.org/
