On Thu, Oct 25, 2007 at 11:51:01PM +0200, Arnold Schekkerman wrote: > On 25-10-07 19:25, Sam Mason wrote: > > On Thu, Oct 25, 2007 at 05:40:44PM +0200, Adrian von Bidder wrote: > >> On Thursday 25 October 2007 17.24:58 Sam Mason wrote: > >>> But the case that Adrian was talking about was with an "intermittent" > >>> connection, so NTP probably wouldn't be running anyway. Maybe I'm > >>> missing something. > >> My idea is to have ntpd running in these cases. Every time connection goes > >> life, you tell it to do ntp packet exchanges. So if the network comes > >> online often enough, you get proper ntp sync where you usually would only > >> be able to run sntp clients. (This would have to be optimized by options > >> like "use burst as long as not synchronized" and so on.) > > > > Without knowing more about the algorithm that NTP uses to synchronise > > I don't know whether it would actually be able to do a better job than > > irregular sntp requests. It'd be nice to think so, but life is never > > that easy. > > I don't know exactly, but I'd be surprised if NTP does NOT do a better job.
I've never been particularly hot on control theory, I just know that things can go unstable very quickly if the design assumptions aren't maintained. You generally have to treat things as linear to do any useful analysis, which is where the assumptions come in; most systems behave in a nice linear way when you look at a small enough part of them but as soon as you move outside this region bad things tend to happen. > Even if you do a time request once a week, NTP can compensate a clock that > gains or looses 1 minute/week. That's about 100PPM isn't it? If it is, that's within the sort of error that the current algorithm can handle. I'm still not sure how it would cope with the varying and very intermittent additions of information though. > After a few weeks NTP will keep the clock > within a few seconds, while with sntp you will be off 1 minute each week > again. The only requirement (theoretical) is a free running clock with more > or less constant deviation. Which, apart from temperature changes, most clocks seem to have. Isn't this what NTP relies on to work at the moment? > If the free running clock gains 1 minute in one > week, but looses a minute in another, then you need small intervals. Yes, you've probably got other problems though. Sam _______________________________________________ timekeepers mailing list [email protected] https://fortytwo.ch/mailman/cgi-bin/listinfo/timekeepers
