Self-similar traffic modeling and network performance analysis

Recent studies have shown that network traffic exhibits the property of self-similarity. The complexity of self-similarity has rendered existing traffic and performance models to be either analytically intractable in the evaluation of performance, or relatively inaccurate with respect to predicting dynamic queueing behavior.; Our main objectives in this dissertation are to develop tractable models for self-similar traffic and to present efficient approaches for network performance analysis. In order to gain a better understanding of self-similar processes, we evaluate the power spectral density function. We also present an improved self-similar traffic generation algorithm which is more efficient than existing methods for specific fast simulation of self-similar traffic. We develop a new traffic model based on a sum of Shifting Level Processes (SLP) aggregated over time. This model has a structure similar to that of a fractional ARIMA (Auto-R.egressive Integrated Moving Average) process, with a driven process of fBm (fractional Brownian motion), where the coefficients of the fBm are derived from the Pareto distribution of the active periods of the arrival process. This is a fairly general model with several special cases, which include SLP and On-Off processes. The queueing behavior of a single server with a self-similar input can be investigated analytically with the model developed in this study. Based on Norros' Lower Bound and Duffield's large deviation results, an upper bound and an approximate expression on the queue length distribution are found to predict network performances. The traffic models developed and performance analysis results presented have been verified using real traffic on an Arm Switch Network.; We also present a new aggregation and decomposition approach for the performance analysis of multi-stage switching networks with finite buffers and bursty traffic inputs, where existing methods cannot be applied. By iteratively aggregating various traffic inputs into an equivalent traffic stream, the buffer state expressions are obtained based on the states of neighboring switches. The switch at each stage is then decomposed from its neighboring buffers. It is proposed that the crossbars along the input-output route of interest in the switching network can be regrouped into an equivalent tandem line model. In terms of the conservation-of-flow, a set of nonlinear equations is recursively solved for equivalent buffer parameters. The typical performance metrics of interest are then evaluated, such as average throughput, buffer occupancy and packet delay. Compared to existing methods, the commonly used uniform traffic model and the independence assumptions are not necessary.