Parsimonious Network Traffic Modeling By Transformed ARMA Models

Generating synthetic data traffic, which statistically resembles its recorded counterpart is one of the main goals of network traffic modeling. Equivalently, one or several random processes shall be created, exhibiting multiple prescribed statistical measures. In this paper, we present a framework enabling the joint representation of distributions, autocorrelations and cross-correlations of multiple processes. This is achieved by so called transformed Gaussian autoregressive moving-average models. They constitute an analytically tractable framework, which allows for the separation of the fitting problems into subproblems for individual measures. Accordingly, known fitting techniques and algorithms can be deployed for the respective solution. The proposed framework exhibits promising properties: 1) relevant statistical properties such as heavy tails and long-range dependences are manageable; 2) the resulting models are parsimonious; 3) the fitting procedure is fully automatic; and 4) the complexity of generating synthetic traffic is very low. We evaluate the framework with traced traffic, i.e., aggregated traffic, online gaming, and video streaming. The queueing responses of synthetic and recorded traffic exhibit identical statistics. This paper provides guidance for high quality modeling of network traffic. It proposes a unifying framework, validates several fitting algorithms, and suggests combinations of algorithms suited best for specific traffic types.