On the Dynamics and Significance of Low Frequency Components of Internet Load

Dynamics of Internet load are investigated using statistics of round-trip delays, packet losses and out-of-order sequence of acknowledgments. Several segments of the Internet are studied. They include a regional network (the Jon von Neumann Center Network), a segment of the NSFNet backbone and a cross-country network consisting of regional and backbone segments. Issues addressed include: (a) dominant time scales in network workload; (b) the relationship between packet loss and different statistics of round-trip delay (average, minimum, maximum and standard-deviation); (c) the relationship between out of sequence acknowledgments and different statistics of delay; (d) the distribution of delay; (e) a comparison of results across different network segments (regional, backbone and cross-country); and (f) a comparison of results across time for a specific network segment. This study attempts to characterize the dynamics of Internet workload from an end-point perspective. A key conclusion from the data is that efficient congestion control is still a very difficult problem in large internetworks. Nevertheless, there are interesting signals of congestion that may be inferred from the data. Examples include (a) presence of slow oscillation components in smoothed network delay, (b) increase in conditional expected loss and conditional out-of-sequence acknowledgments as a function of various statistics of delay, (c) change in delay distribution parameters as a function of load, while the distribution itself remains the same, etc. The results have potential application in heuristic algorithms and analytical approximations for congestion control