Multipath Routing From a Traffic Engineering Perspective: How Beneficial is It?

Multipath routing gives traffic demands an opportunity to use multiple paths through a network. In a single-demand situation, its benefits are easy to see. In a multi-commodity case, when potentially all node-pairs (demands) generate traffic, they compete for the same network resources. In this work, we consider multipath routing in communication networks in a multi-commodity setting from a traffic engineering perspective. Based on a result from linear programming, we show that at an optimal solution, the number of demands that can have multiple paths with nonzero flows is of the order of the number of network links for three commonly used traffic engineering objectives. We introduce a multipath measure (MPM) and show that under certain traffic conditions and topological structures, the MPM is zero or close to zero; i.e., multipath routing provides little or limited gain compared to single-path routing. For the all-pair traffic case, multipath routing is observed to be advantageous for small networks. When the number of nodes is about 25 or higher and all node pairs have traffic, this advantage drops as the number of nodes in a network increases. For the fat-tree data center topology, the benefit of multipath routing also drops as the number of pods increases. Our findings are somewhat against a common belief (expressed by the term "load sharing") that multipath routing is significantly better in effective distribution of traffic over the network resources.