A Study onthe Cause of Long- Range Dependence Observed in Empirical TCP Traffic Traces

Several recent studies on a wide variety of networks have empirically observed that aggregate packet ows exhibit long-range dependence, i.e., the correlation between neighboring exclusive blocks of data does not asymptotically vanish when the block size is increased. Thus, actual network tra c is bursty over a broad range of time scales, in sharp contrast to conventional Markovian-type tra c models. Conventional models show tra c burstiness at only short time scales while tra c is smooth at large time scales. Additional studies have demonstrated that long-range dependence in trafc can have serious e ects on network performance, but none of the previous work gives a complete answer of what is causing long-range dependence in network tra c. Since the majority of empirical studies were performed with TCP tra c, the initial hypothesis was that the dynamics of TCP, such as ow and congestion control algorithms, are the primary factor contributing to long-range dependence (LRD) in TCP tra c. Preliminary simulation results supported our assumption. However, futher investigation and further results from a large number of simulations contradict our assumption; the dynamics of TCP are not the main cause of the LRD observed in TCP tra c. Our results show that the presence of long-range dependence in network tra c does not necessarily depend on whether or not a reliable and owand congestion-controlled protocol is employed at the transport layer. Further, the results of this study show that in a client/server network environment, if tra c exhibits long-range dependence, then the distribution of message sizes is not necessary heavy-tailed.