MultiPaths Revisited - A novel approach using OpenFlow-enabled devices

This thesis presents novel approaches enhancing the performance of computer networks using multipaths. Our enhancements take the form of congestion-aware routing protocols. We present three protocols called MultiRoute, Step-Route, and finally PathRoute. Each of these protocols leverage both local and remote congestion statistics and build different representations (or views) of the network congestion by using an innovative representation of congestion for router-router links. These congestion statistics are then distributed via an aggregation protocol to other routers in the network. For many years, multipath routing protocols have only been used in simple situations, such as Link Aggregation and/or networks where paths of equal cost (and therefore equal delay) exist. But, paths of unequal costs are often discarded to the benefit of shortest path only routing because it is known that paths of unequal length present different delays and therefore cause out of order packets which cause catastrophic network performances. Further, multipaths become highly beneficial when alternative paths are selected based on the network congestion. But, no realistic solution has been proposed for congestion-aware multipath networks. We present in this thesis a method which selects alternative paths based on network congestion and completely avoids the issue of out of order packets by grouping packets into flows and binding them to a single path for a limited duration. The implementation of these protocols relies heavily on OpenFlow and NOX. OpenFlow enables network researchers to control the behavior of their network equipment by specifying rules in the routers flow table. NOX provides a simple Application Programming Interface (API) to program a routers flow table. Therefore by using OpenFlow and NOX, we are able to define new routing protocols like the ones which we will present in this thesis. We show in this thesis that grouping packets together, while not optimal, still provides a significant increase in network performance. More precisely we show that our protocols can, in some cases, achieve up to N times the throughput of Shortest Path (SP), where N is the number of distinct paths of identical throughput from source to destination. We also show that our protocols provide more predictable throughput than simple hash-based routing algorithms. %extend this with steproute and pathroute results. Todays networks provide more and more connections between any source-destination pair. Most of these connections remain idle until some failure occurs. Using the protocols proposed in this thesis, networks could leverage the added bandwidth provided by these currently idle connections. Therefore, we could increase the overall performance of current networks without replacing the existing hardware.