Impact of flow control windows in TCP on fractal scaling of traffic exiting from a server pool
Heung-No Lee · 2002
We provide an analytical and network-systematic framework to characterize the self-similar, fractal scaling phenomenon which is believed to be ubiquitously present in modern high speed data network traffic. We show that the self-similar network traffic is due mainly to the use of closed-loop flow control at the transport layers, such as the use of the classes of protocols from the TCP family. For in-depth investigation on the subject, we inject synthetically generated application-level traffic which is completely short-range traffic, into a very simple network simulated with NS2.0, and examine the influence of different parameters of a TCP algorithm on a variety of different fractal scaling behaviors observed at the packet-level traffic traversing a link in the simulated network. We provide a very simple - but intuitive - mathematical explanation of the observed phenomenon using the shot-noise processes. Specifically, different kernel filters of the shot-noise are constructed to model the behavior of the window process cwnd in different stages of the congestion avoidance algorithms employed in a TCP algorithm. With the use of exponential-law shot-noise process, for example, we indicate that the cwnd process in the slow start phase results in a unique scaling behavior from RTT to a finer time scale, having a scaling slope of /spl gamma/ = 2. From RTT to a coarse time-scale, the more conventional fractal scaling behavior with the Hurst parameter less than 1.0 is observed; we compare this with the rectangular-Pareto shot noise process.