On Friday, June 8, 2012 12:43:12 PM UTC+10, David L. Mills wrote: > Julian, > > Thanks for the paper reference. Your ideas on feed-forward are similar > to the ideas in Greg Troxel's MIT dissertation. These ideas were > partially implemented in NTPv3 a very long time ago. > > There are some minor misinterptretations in the paper. The NTP > discipline loop is not critically damped; it is purposely underdamped > with a minor overshoot of a few percent in order to improve the > transient response. The impulse response was slavishly copied in the > microkernel and nanokernel code that left here over a decade ago. The > microkernel survives in Solaris and the nanokernel in most BSD systems > with varying degrees of fidelity; however, many systems have elected to > modify the original BSD tickadj semantics, which result in an extra > pole. The result is a moderate instability at the longer poll intervals, > especially if the step threshold is increased or eliminated. In any > case, the response has no serious misbehavior as the paper described. > Note that in no case are the daemon and kernel algorithms cascaded as > the paper implies. Either one or the other is used, but not both. > > The system behavior with multiple servers is indeed as the paper > suggests, but there is considerable algorithm horsepower to diminish the > effects, including the cluster and combine algorithms, plus the > anti-clockhop and prefer peer mechanisms. These provisions were first > implementd in the Internet of twenty years ago when the congestion on > overseas links frequently reached over one second. Perhaps today these > algorithms could be more carefully tuned for LANs and even wifi netorks. > > As the paper describes, NTP algorithms are designed for traditional > mathematical analysis, but with both linear and nonlinear components. > However, the FLL algorithm is based on a model described by Levine as > predictive. The model in the documentation describes both the PLL and > FLL in predictive terms, but that doesn't change the conclusions in the > paper. > > The paper suggests possible improvements in data filtering and analysis. > The clock filter and popcorn spike suppressor algorithms in NTP > represent one approach. A persistent observation is that NTP does not > effectively use offset/delay samples other than at the apex of the > scattergram. While it does indeed do that for the huff-n'-puff fiilter, > the possible improvement in other cases is problematic. The paper does > not mention the implications of roundtrip delay in the maximum error > statistic, such as in Cristian's model, as used by NTP. It is a natural > error bound for asymmetric paths such as mentioned in the paper. > > In summary, the NTP algorithms have evolved over thiry years in response > to major changes in Internet service models and surely could use some > further evolution. I am glad there is continuing interest in improvements. > > Dave
Dear David, Thanks a lot for mentioning Greg Troxel's dissertation, I was not aware of it and will read it avidly. I really appreciate your comments and details you provided on ntpd design. I believe this is valuable information for readers interested in our paper, and I am willing to add your comment to our webpage where the paper is referenced. Would you agree to this? Julien _______________________________________________ questions mailing list [email protected] http://lists.ntp.org/listinfo/questions
