On 2013-04-01, David Woolley <[email protected]> wrote: > unruh wrote: > >> >> Well, modulo Mills' prejudices as to how ntpd should work. In short, if >> the user has different requirements from those that Mills had, there may >> be a way of using the data differently to get better results. But you >> really really have to know what you are doing. >> > As the OP mentioned PIDs (proportional, integral, differential > controllers), I think he is assuming the Millsian process control > algorithm type of structure, rather than the linear regression structure > used by chrony. He's just assuming a much smaller loop time constant > than is actually used.
I think he really does not know what he wants or what ntp, or chrony, does. It may be that for his purposes a different control structure would work better. Eg different weights in the PID or different time constants. Or perhaps a linear regression structure would work better or even some non-linear structure. But he needs to know what he wants, and needs to know how to get it from the data, keeping in mind all of the things that can go wrong. Eg, how will his control structure handle an hour of high one way delays in the network? Or sporadic delays. Of the remote server going down and the clock having the freewheel. Or one of the many other things that can go wrong. Thus a system which 90 % of the time works better than ntpd, but 10 % goes into totally unstable oscillation would probably not satisfy him. (or even 1%). And any looking at the system over some fairly short period of time ( eg the 15 min he used as his base) might show that his system works much better. _______________________________________________ questions mailing list [email protected] http://lists.ntp.org/listinfo/questions
