On 2011-12-25, David J Taylor <[email protected]> wrote: > "unruh" <[email protected]> wrote in message > news:[email protected]... > [] >>> Question: would you expect the reported jitter to increase over the >>> first 30 minutes or so? >> >> Could be somone switched on a vacuum cleaner for example. > > No. I've seen something like this behaviour before, with the initial few > tens of minutes producing a more stable results than a full run. > >> That is now ntp works. All it knows is the current offset, and tries to >> get rid of it by changing the frequency. >> It does not know that there is a sudden step. I does not remember the >> old offset values. > > This behaviour seems wrong to me. Unless it's known that the frequency > can suddenly change, NTP should not be altering it in crash-bang steps, or > it should take a more long-term view before doing so.
ntpd does NOT take a long term veiw. The decision made in designing it wasw that it would be a short term Markovian feedback loop. It does not have any memory If the offset is positive speed up the clock, if negative, slow it down. If you want a program with a different design philosophy, get chrony. It remembers up to 64 of the last measurements and uses them to determine what the best estmate is of the actual offset and rate error in the local clock is. (It uses linear regression, corrects past measurements for current offset and rate changes, and tests to make sure that a linear fit is a good estimate, decreasing the number of remembered items if it is not.) > >> You might look at the peerstats file and also look at the "roundtrip" >> time. I might be that occasionally one of the paths from wireless to >> computer gets shorter ( clearer signal?) and ntpd will then take that as >> a good value, and an earlier value, and try to correct for that offset-- >> which it does by stepping the frequency. > > I can imagine an occasional longer delay, but not a shorter one. I > haven't been collecting peerstats data. Signal strength is high and > unlikely to drop, although I accept that's not impossible. Sure it can. HIgh noise levels with occasional bursts of low noise. The ntpd algorithm saves the last 8 offsets and roundtrip times, and uses only the one with the smallest roundtrip, on the theory that is the best with smallest assymetry in out and in trips. This is not a good estimate if there is generally noise but occasionally a clear spell which affects one of the paths but not the other. > > Cheers, > David > _______________________________________________ questions mailing list [email protected] http://lists.ntp.org/listinfo/questions
