Routing optimization and capacity assignment in multi-service IP networks

This dissertation investigates routing optimization and capacity assignment in IP networks with multiple services. Both are key processes of traffic engineering and network planning, needed for cost-efficient, quality of service (QoS) aware operation of IP networks. Routing optimization provides a means to balance the traffic load in the network with the goal to improve quality of service. The main objective of our routing optimization procedures is the minimization of the maximum link utilization in the network. Two fundamentally different strategies of routing optimization are investigated. In the first case, optimization is based on conventional routing protocols, which implement shortest-path routing based on static link metrics. By adjusting the values of these metrics (i.e., the link weights), the route selection process can be influenced and, thus, the path pattern optimized. A significant novelty of our approach is the consideration of additive as well as concave link metrics. This approach allows a more flexible route selection and, therefore, provides a greater optimization potential. The link weight setting problem is formulated as a mixed-integer linear program, which for small networks can be solved with standard tools. For larger networks, a hybrid genetic algorithm is presented. In the second case, source routing is utilized as a complementary technology to improve an existing routing configuration. In order to avoid having to establish dedicated paths between all nodes of a network, a hybrid approach is pursued. The majority of traffic is forwarded along shortest paths in native IP fashion and only a small number of flows is selected and subject to source routing. For this type of routing optimization two mixed-integer linear programs are presented. Starting with a given configuration of shortest-path routing, an optimal set of flows is identified for rerouting and the corresponding paths are computed. For the capacity assignment process we distinguish between two basic types of traffic, elastic and stream traffic, which are associated with data applications and real-time services, respectively. Since their characteristics and QoS requirements are different, two distinct approaches are presented. The fundamental property of elastic traffic is its rate adaptability, which is caused by the feedback mechanism of TCP. The transmission rate of a sender is adjusted to the available bandwidth along the route towards the receiver. As a consequence, the overall throughput, which is the relevant QoS measure for elastic traffic, depends not only on the link capacities, but also on the number of concurrent flows. In this dissertation, the theory of processor sharing is applied to the dimensioning of networks for elastic traffic. In our context, processor sharing is interpreted as ideal bandwidth sharing among flows, whose arrival instances form a Poisson process and whose service times are generally distributed. The basic model is investigated and deficiencies are identified. In networks with longer round trip times the