Correlations Induced in a Packet Stream by Background Traffic in a Multiplexed Environment S. Srivastava , K. Mitchell , and A. van de Liefvoort Division of Computer Science & Electrical Engineering, School of Interdisciplinary Computing and Engineering, University of Missouri–Kansas City, Kansas City, MO 64110 USA. The authors present an analytic model that shows that the service requirements at the server for background classes of traffic can have a significant effect on a tagged class, impacting not only the variance, but more importantly, introducing slowly decaying correlations in the tagged stream. The model is general in that it incorporates first- and second-order statistics for all processes involved and focuses on the correlations introduced by multiplexing. The results show that correlations present in background classes, combined with differing service requirements for the different classes, have enormous impact on the departure stream of all classes involved. 1. Introduction Over the past 20 years, IP networks have been extremely successful in carrying delay insen- sitive data. These networks have been relatively cheap to build, simple to manage, and scale very well. Recently, there has been much debate on the ability of IP networks to provide quality of service (QoS) in the Internet. In order to provide service comparable to flow based networks such as ATM, two issues must be addressed; efficient use of bandwidth, and ensuring QoS for a service class. Efficient use of bandwidth can be ensured by incorporating traffic engineering. Fortz and Thorup [1] show that traffic engineering with open shortest path first (OSPF) performs within a few percent of what flow based networks (optimal general routing) can achieve. Pioro et al. [2] provide mechanisms to find weight systems that ensure effective traffic engineering in an OSPF network. Such approaches make efficient use of bandwidth within reach in an IP network. Ensuring QoS for a service class in IP networks poses the same issues and considerations as in flow based networks, except that the traffic characteristics can vary more widely. In flow based networks, traffic streams share bandwidth with their own classes only, and hence it is reasonable to assume a certain homogeneity in terms of the arriving streams and their service requirements. In IP networks, bandwidth is usually shared between all arriving streams of all classes, and assumptions of homogeneity may not be realistic. In fact, we will show that common assumptions regarding outflows may not be valid. This work has been supported in part by US NSF under grant No. ANI 0106640.