Analysis and stochastic characterization of TCP flows

Since the most Internet services use TCP as a transport protocol there is a growing interest in the characterization of TCP flows. However, the flow characteristics depend on a large number of factors, due to the complexity of the TCP. As a result, the TCP characteristics are normally studied by means of simulations or controlled network setups. In this paper we propose a TCP characterization based on a generic model based of stochastic flow with burstiness and throughput ((σ,ρ)-constrains), which is useful in order to characterize flows in atm and other flow-switched networks. The model is obtained through extensive analysis of a real traffic trace, comprising an approximate number of 1,500 hosts and 1,700,000 TCP connections. The results suggest that TCP connections in the wide area Internet have low throughput while the packet bursts do not suffer an exponential increase, as indicated by the slow-start behavior. On the other hand, the impact of the connection establishment phase is striking. We note that the throughput of the TCP flow is approximately half the throughput which is obtained in the data transfer phase, namely after the connection has been established.