Selective randomized load balancing and mesh networks with changing demands

We consider the problem of building cost-effective networks which are robust to dynamic changes in demand patterns. We compare several architectures based on using demandoblivious routing strategies. Traditional approaches include single-hop architectures based on a (static or dynamic) circuitswitched core infrastructure, and multi-hop (packet-switched) architectures based on point-to-point circuits in the core. To address demand uncertainty, we seek minimum cost networks that can carry the class of hose demand matrices. Apart from shortest-path routing, Valiant’s randomized load balancing (RLB), and VPN tree routing, we propose a third, highly attractive approach: selective randomized load balancing (SRLB). This is a blend of dual-hop hub routing and randomized load balancing which combines the advantages of both architectures in terms of network cost, delay, and delay jitter. In particular, we give empirical analyses for the cost (in terms of transport and switching equipment) for the discussed architectures, based on three representative carrier networks. On these three networks, SRLB maintains the resilience properties of RLB while achieving significant cost reduction over all other architectures, including RLB and multi-hop IP/MPLS networks using VPN tree routing.