Price of Path Diversity: in Practice and in Theory

Recent research [17,26] suggests that path diversity, in which end systems choose from a set of network-constructed paths, can achieve much of the benefit of source routing such as improved reliability, but avoids its scaling and policy issues. This work examines a less explored performance aspect of path diversity: end-to-end throughput. When an end system can access a diverse set of paths, it may use all or a subset of the paths to improve its throughput. Unfortunately, using both realistic simulations as well as a theoretic optimization framework, we discover that when all nodes can use multiple paths to send their traffic, the all-pair throughput of the network may decrease significantly (e.g., up to 50% in some cases). A main reason is that traffic sent along paths longer than the shortest paths consume more network resources than necessary. From this study, we suggest three areas of future work to reduce the downside of path diversity: 1) better path construction schemes, 2) more efficient multi-path resource allocation, and 3) a more practical optimization framework that can guide the design of multi-path resource allocation.