Selective randomized load balancing outperforms VPN-tree routing for disparate demand and link granularities

Abstract

We find that selective randomized load balancing (SRLB) and Valiant's fully-randomized load balancing (RLB) can be lower-cost routing templates than VPN-tree routing for hose-constrained random network traffic, provided that link bandwidths have a significantly coarser granularity than individual demands, as is typically the case in today's optical networks. Our results build on the previous VPN conjecture and subsequent proof that VPN-tree routing provides the optimal capacity solution for the VPN design problem under the assumptions that the bandwidth demands and link capacities are both fractional or both integral with the same granularity. We consider ring and mesh network topologies and use both an analytic approach as well as a numerical exhaustive search to find the lowest-cost routing templates. In the analytic analyses, we find that using RLB on both full-mesh and ring networks can achieve lower cost than VPN-tree routed networks for fully hose-model traffic. In the exhaustive-search analyses, we calculate the cost of a general mesh network and study the effect of adding varying amounts of deterministic traffic to the hose traffic. We find that a `background' of deterministic traffic tends to shift the optimal routing template away from VPN tree and towards SRLB and RLB solutions.