Protocols Can Make Traffic Appear Self-Similar

Empirical studies have shown that self-similar traffic models may better describe traffic in many of today's computer networks than traditional Markovian models. The causes of this apparent self-similar behavior must be identified to determine how widely applicable these models are, and how network designers should respond. While some researchers have argued self similarity is an inherent property of traffic as generated by the typical applications, it is also possible that the network's own protocols may cause or at least contribute to this phenomenon. In this paper, it is shown that even if packets were to arrive according to the wellbehaved Poisson process, simple retransmission mechanisms can make traffic appear self similar over time scales of engineering interest. Moreover, some techniques intended to decrease the likelihood of congestion also have the effect of prolonging congestion when it does occur. This increases burstiness over large time-scales, reinforcing the appearance of self-similarity.