TCPBP: enhancing the TCP throughput in high-speed lossy networks based on a single mixed routing metric

Abstract

This work presents the Transmission Control Protocol Best Path (TCPBP), an alternative routing metric that aims to maximize the aggregated single-path TCP throughput on intradomain HighSpeed Lossy Networks. The enhanced performance is achieved by combining delay and packet loss on a single mixed metric as a function parameter which predicts the per-flow TCP throughput that should be maximized. Maximizing the throughput of each TCP flow traversing a given route is also expected to improve the aggregated throughput (i.e., the sum of all TCP flows). To demonstrate the feasibility of our approach concerning obtaining viable routes in any topology represented by a G(V,E) directed graph, we proved that the metric satisfies the mathematical properties of isotonicity and monotonicity. This fact implies that it is consistent, optimal, and loop-less, i.e., it is feasible regarding the shortest path computation, thus providing viable routes. For the shortest path computation, we devised a modified edge relaxation process suitable for non-additive metrics with positive edge costs in the Dijkstra algorithm. As proof of concept, TCP bandwidth tests were executed in an SDN High-Speed Lossy Network Linux-emulated topology with ten nodes. For the routing algorithm, we employed an intradomain link-state routing protocol (i.e., OSPF-like) capable of running both TCPBP and traditional metrics delay and hop-count. Such topology was chosen because of the potential to improve its TCP throughput with route selection, given the characteristics of the topology itself and the TCP throughput prediction model adopted. The shortest path computation running time in the control plane proved not significantly slower than the default Dijkstra algorithm traditionally used with standard additive metrics (such as hop-count). Also, the overall TCP throughput performance obtained with TCPBP was greater than with different routes founded by traditional metrics, without impact on the TCP fairness, thereby showing the potential of our proposal.