A cause of self-similarity in TCP traffic

We analyse the statistical properties of aggregated traffic flows generated by TCP, in order to clarify a possible cause of self-similarity in Internet traffic. Using ns-2 simulation, we first show that aggregated traffic flows by TCP can be characterized by phase transition phenomena between non-congested and congested phases in statistical physics. Interestingly, although the traffic exhibits self-similarity with the Hurst parameter H approximate to 1.0 at the critical point between them, it is close to Poissonian away from the critical point. This result is consistent with results from real WAN traffic measurement. The main contribution of our work is to show that TCP itself contains a mechanism for generating self-similarity, without assuming self-similarity or long-range dependency (LRD) in the application layer (e.g. packet inter-arrival, connection arrival, and file size distribution). Moreover, we find that the value of the Hurst parameter at the critical point is determined by the simple feedback control called stop-and-wait flow control. Namely, it appears even without any packet retransmission events and is independent of the explicit rate increment algorithm such as slow-start and congestion avoidance. Additionally, we demonstrate that the value of the Hurst parameter at the critical point depends on the constant bit rate algorithm and topology of network. Finally, we indicate that the time series of the round trip time follow the same statistics at the critical state. Copyright (c) 2005 John Wiley & Sons, Ltd.