The author proposes to characterize the burstiness of packet-arrival processes with indexes of dispersion for intervals and counts. These indexes, which are functions of the variance of intervals and counts, are relatively straightforward and can estimate and convey much more information than simpler indexes (such as the coefficient of variation) that are often used to describe burstiness quantitatively. The author defines and evaluates the indexes of dispersion for some of the simple analytical models that are frequently used to represent highly variable processes. The indexes for a number of measured point processes that were generated by workstations communicating to file servers over a local area network are also estimated. It is shown that nonstationary components in the measured packet-arrival data distort the shape of the indexes, and ways to handle nonstationary data are proposed. To show how to incorporate measures of variability into analytical models and to offer an example of how to model the measured packet-arrival processes, the author describes a fitting procedure based on the index of dispersion for counts for the Markov-modulated Poisson process.>

The traffic on a 10-Mb/s Ethernet local-area network that connects diskless workstations to file servers in a university environment is analyzed. The traffic is substantially heavier than has been recorded in previous studies; measured over 1-s intervals, it frequently exceeds 30% of the network bandwidth. The distribution of packet lengths and the patterns of packet interarrival times are interpreted for the three protocols that carry significant traffic: the transmission control protocol (character traffic), the network disk protocol (paging traffic), and the network file system protocol (remote file access traffic). The two latter protocols account for 68% of the packets and 94% of the data bytes on the network. File access to a remote file server generates bursts of traffic that can last several seconds and that demand bandwidths on the order of 120 kB/s, or about 10% of the Ethernet bandwidth.>

The authors discuss the upper and lower bounds on the accuracy of the time synchronization achieved by the algorithm implemented in TEMPO, the distributed service that synchronizes the clocks of the University of California, Berkeley, UNIX 4.3BSD systems. The accuracy is shown to be a function of the network transmission latency; it depends linearly upon the drift rate of the clocks and the interval between synchronizations. TEMPO keeps the clocks of the VAX computers in a local area network synchronized with an accuracy comparable to the resolution of single-machine clocks. Comparison with other clock synchronization algorithms shows that TEMPO, in an environment with no Byzantine faults, can achieve better synchronization at a lower cost.>