>... >From: "List Mail User" <[EMAIL PROTECTED]> > >> > ... The person with two clocks is never really sure of >> > the current time. >> >> OT, but... above - *not* a good quote, but it sounds nice) >> To be `sure' of the time, you need at least three clocks (look at the >> documentation for ntp/ntpd). >
jdow wrote (quite correctly) at about 13:15:23 on Sun, 13 Mar 2005 >And even that is a gross oversimplification. (And I COULD setup my >system at one time, at least, to be approximately 1 second off by >picking the wrong ntp servers. Seems GTEI.NET's time server used for >their DNS machines and first hop routers was off, considerably.) > >Reading the ntp/ntpd documentation is desirable in any case if one is >interested in precision time keeping. > >{^_-} > Of course, you are correct; I was just "making fun" of the original quote; I actually use 3 GPS receivers on three different machines and between the servers, in total use a overlapping set of 12 public peers/servers - I still have to "fudge" (for the non-ntpd literate RTFM) due to bad skew in my cheap receivers and actually find the best time source for many years has been clepsydra.dec.com (long ago and far away, it was timekeeper,dec,com). I also set my own GPSs to stratum 1, because the skew is so bad. Unfortunately, it has been nearly ten years since all the U. Del. servers have been badly overloaded and True-Time no longer allows public free access. So, yes, I do take "time" seriously; But you are more than just simply correct when you point out that my misstatement was almost as incorrect as the quote I was poking fun at. (I've seen the lousy GTE servers; until recently SBC ran some really good, but unpublished servers - but they seem to have been taken down, or at least don't answer public queries as of last month - they were the "old" pacbell servers - and the SBC published servers "false tick" about once a hour). Basically, I'm happy with keeping my net (under a hundred machines), synced within 5-8 msec. and a estimated error on all servers (not just time servers) under 2 msecs. Better than that would mean buying "real" equipment. At this moment, on my primary time server (i.e. "prefer"'d by most of my machines), I'm within 12 msec for all the public peers/servers I use except for otc2.psu.edu (who I should drop), louie.udel.edu, and (one of the rotating x.pool.ntp.org servers) splinter.bowdoin.edu all of whom have "false tick"'d in the past 25 minutes - for most of the rest I'm within 1.5 msec and my own three time servers are all within 600 usec of each other (two on the same subnet, one two subnets away). Yes, what ntp does used to seem like magic, now I think the best improvement will come with a better filtering function for "smoothing" (e.g. a Widrow style adaptive filter instead of the simple FIR used). Several people have shown that just adding taps doesn't provide any improvement (I seem to remember an old paper showing the six or seven was as good as or better than the eight tap filter actually used - but I haven't done any `real' DSP work in a nearly decade). Also IIRs look better at times, but the `bad' cases entirely rule them out (now if you used a Widrow filter to generate the IIR coefficients and did a fall back to FIR whenever the output approached instability,... I think I'll leave these things to people like Eric Fair, etc. who have more time to spend thinking about it than I do). Anyway, I'd love to discuss this off list, but this is way OT for the SA group. Bye, Paul Shupak [EMAIL PROTECTED]