Using multi-objective domain optimization for routing in hierarchical networks

Network hierarchy makes network protocols more scalable and robust, but also makes the network more complex and reduces performance. With routing protocols, hierarchy reduces overhead, routing table size, and convergence time, but can also cause sub-optimality (stretch) of the routing path length compared to the flat networks. With OSPF, for example, adding routing areas, with route aggregation done in area border routers, can significantly reduce link state advertisements, link state database and convergence time, but can also increase interarea path length. Existing research has produced many excellent intra- and inter-domain routing schemes and different proposals for aggregation at border routers. As important, however, is the design of the routing hierarchy itself. This paper presents quantified comparisons of different hierarchical construction techniques for a simple hierarchical routing protocol in a 100 node network. When the hierarchy does not take into account routing path length suboptimality, we show the potential for significant stretch (e.g., on average more than doubling the shortest path for nodes under 6 hops apart). When we use multi-objective optimization, that includes the goal of minimizing stretch, the stretch is significantly reduced (e.g., reducing the stretch by approximately 50% for the above example). In large or dynamic environments, a modified simulated annealing is able to produce optimized hierarchies quickly by trading a small loss in optimality for a large reduction in optimization time. The paper analyzes the choice of multi-objective cost function in this environment and concludes that simpler functions produce the best overall results.