Network traffic characterization of TCP
Wu-chun Feng · 2002
Networks from wireless to high-speed integrated services require support for the fluctuating and heterogeneous demands of end users. The ability to characterize the behavior of the resulting aggregate network traffic can provide insight into how traffic should be scheduled to make efficient use of the network, and yet still deliver expected quality-of-service to end users. These issues are of fundamental importance to the design of the next-generation Internet-from wireless communication to high-performance distributed computational grids such as NASA's Information Power Grid. Many research efforts in network traffic characterization conclude that network traffic is self-similar (i.e., fractal or bursty), and thus not amenable to the statistical-multiplexing techniques currently found in the Internet. In particular they claim that the heavy-tailed distributions of file size packet interarrival, and transfer duration solely contribute to the self-similarity of aggregate network traffic. In contrast we demonstrate that it is the TCP stack itself that induces much of the self-similar behavior even when aggregated application traffic should smooth out as more applications' traffic are multiplexed. Furthermore, if random early detection (RED) gateways/routers are used, we show that network performance degrades even further due to the extra burstiness induced by the gateway itself.