Congestion avoidance in overlay networks through multipath routing

Abstract

Overlay networks relying on traditional multicast routing approaches use only a single path between a sender and a receiver. This path is selected based on latency, with the goal of achieving fast delivery. Content is routed through links with low latency, ignoring slower links of the network which remain unused. With the increasing size of content on the Internet, this leads to congestion, messages are dropped and have to be retransmitted. A multicast multipath congestion-avoidance routing scheme which uses multiple bottleneck-disjoint paths between senders and receivers was developed, as was a linear programming model of the network to distribute messages intelligently across these paths according to two goals: minimum network usage and load-balancing. The former aims to use as few links as possible to perform routing, while the latter spreads messages across as many links as possible, evenly distributing the traffic. Another technique, called message splitting, was also used. This allows nodes to send a single copy of a message with multiple receivers, which will then be duplicated by a node closer to the receivers and sent along separate paths only when required. The model considers all of the messages in the network and is a global optimisation. Nevertheless, it can be solved quickly for large networks and workloads, with the cost of routing remaining almost entirely the cost of finding multiple paths between senders and receivers. The Gurobi linear programming solver was used to find solutions to the model. This routing approach was implemented in the NS-3 network simulator. The work is presented as a messaging middleware scheme, which can be applied to any overlay messaging network.