A delay-based approach for TCP parallelization

TCP parallelization uses multiple virtual TCP connections to transfer data for an application process. The virtual connections can be standard or modified TCP connections. TCP parallelization has been used to effectively utilise bandwidth for data intensive applications. However, its aggressiveness to increase window and reluctance to reduce window tend to be unfair to other traffic in the network. There are solutions tackling this fairness problem. In these solutions, the fairness is achieved at the cost of losing the robustness of TCP parallelization. Other solutions use a single TCP connection to emulate the behaviour of a set of parallel TCP connections and therefore take advantage of the robustness of TCP parallelization and its fairness bias. These approaches can be used to achieve proportional fairness (e.g. MulTCP), or just to give a better throughput. However, such solutions may not be able to achieve the desired results. In this paper, we analyse these solutions and show their potential problems. A delay-based parallelization approach is then suggested to address the fairness concern by decoupling fairness and effectiveness. It shows that the desired fairness can be achieved in presence of congestion while the bandwidth can be effectively used in absence of congestion.