On the variability of internet traffic
Victor S. Frost, Georgios Y. Lazarou · 2000
As the Internet continues to grow in size, so do the number of diverse applications with different quality of service (QoS) requirements that make use of the network to transport their information. Therefore, significant effort is being directed toward extending the current Internet architecture by deploying all kind of emerging technologies so that the resultant network will be capable of providing the desired integration of the various heterogeneous traffic types and supporting diverse levels of quality of services. However, the provision of QoS guarantees over the Internet requires the understanding of traffic characteristics which are relevant to network performance. Many empirical studies on a variety of networks have shown that traffic exhibits high variability over a broad range of time scales. This new phenomenon triggered a vast amount of research in understanding the nature and the cause of this high burstiness of traffic. Several studies claim that this is because of the long-range dependence (LRD) property of traffic processes, and thus conventional traffic models are no longer valid in modeling this type of traffic. Therefore, this assertion that network traffic has LRD initiated a large effort in creating new traffic models that have LRD. Comparing the results from various studies conducted to evaluate the impact of traffic high variability on queueing performance, we believe that none of available measures of traffic variability can accurately capture the degree of traffic variability across time scales. In this dissertation we provide a new theoretical and practical framework based on the theory of point processes that can be used to accurately characterize the variability and correlation structure of traffic. We show through analysis and simulation experiments that the empirically observed high variability of network traffic over a wide range of time scales can be captured by traditional traffic models. In addition, through simulation experiments we evaluate the impact of the Transmission Control Protocol (TCP), the most widely used transport protocol in the Internet, on traffic burstiness (variability). Our results show that (a) our novel measure of variability is a better measure for capturing the burstiness of network traffic than the Hurst parameter, (b) the dynamics of TCP alone can not cause considerable variability over a substantial range of time scales, and (c) the presence of high variability over a wide range of time scales does not necessarily depend on whether a reliable and flow- and congestion-controlled protocol, such as TCP, is employed at the transport layer.