Flow-based Networking and Quality of Service

Abstract

During the past two decades the Internet has been widely deployed and integrated into the society, radically altering the way people communicate and exchange information. Although the Internet was intended as a research network between few a institutions in the United States, it has grown to take a central role in day-to-day communications and is by now considered indispensable. The last decade we have seen a migration of the classic telecommunication services, telephony, radio and television, to the Internet and the emergence of new services, such as online gaming and e-government. Hence, the Internet is replacing the existing, dedicated telecommunication networks and embodies a single medium for all (wired) communications. The low latencies and losses for which the dedicated networks were optimized, cannot be guaranteed by the Internet, which philosophy relies on best-effort routing. Packets in the Internet can literally travel across the entire network, making it very hard to predict where the packets may travel, yielding tremendous uncertainties regarding packet-delays and loss. In flow-based communications, the packets from a source to a destination that belong the same stream are forced along the same path, offering a better control over the packets and an improved quality control. In this thesis we study the performance of networks using flow-based communication and we show by means of measurements that the current Internet can yield highly unpredictable behavior in packet delivery. Based on traceroute measurements we show that the best-effort environment of the Internet can lead to highly unpredictable behavior of packets. The measurements indicate that the routes that the packets in the Internet follow are very changeable due to dynamics in the routing plane. This unpredictable behavior may hamper the performance of real-time applications and serves as a motivation for the flow-based networking used in this thesis. We introduce a new model that describes the network performance by using an analogy with queueing theory. The analysis of network performance is complicated by the many dependencies between the properties of a network. The objective is to express the network performance in terms of these properties and reveal the root of the problem. The model facilitates to express the emergent performance characteristics, such as the blocking rate of traffic flows and the maximum throughput of the network, in terms of network parameters that are given by design, such as the number of nodes. The model considers the network as a black box and minimizes the degrees of freedom in order to reduce the dependencies and improve the comprehensibility of the model. Due to the small number of parameters, we are able to discern the influence of these parameters on the performance. In addition, we address the difficulties of studying network dynamics, which can be traced back to the dependencies between the links and the importance of traffic fluctuations in the network on the overall performance.