Optimizing Congestion Avoidance and Congestion Control in Wired and Wireless Networks

Abstract

Internet over the past few years has transformed from an experimental system into a gigantic and decentralized source of information. The success of the Internet can be partly attributed to the congestion control mechanisms implemented in Transmission Control Protocol (TCP). TCP has been the de-facto transport protocol for Internet since its inception. Although TCP constantly evolved over a period of two decades, the diversity in the characteristics of present and next generation networks and a variety of application requirements have posed several challenges to TCP congestion control mechanisms. As a result, the shortcomings in the fundamental design of TCP have become increasingly apparent. In this dissertation, we propose solutions to overcome these shortcomings and increase the robustness of TCP by carefully optimizing the fundamental TCP congestion control algorithms. We focus on the middle ground between end-to-end transport protocols and network based transport protocols. The major goal is to optimize the performance of TCP while ensuring minimum deployment complexity. The motivation stems from the fact that the need for deployment of Active Queue Management (AQM) and Explicit Congestion Notification (ECN) has become apparent, owing to the drastic impact of “persistently full buffers” on the performance of Internet. We aim to leverage the benefits of AQM/ECN mechanisms and provide a richer explicit feedback to the end-hosts to aid them in making efficient congestion control decisions. Although Random Early Detection (RED) has proved to be an effective AQM mechanism, its performance is highly sensitive to the appropriate settings of the parameters. Rather than tuning the parameters of original RED, this dissertation instead, aims to improve the performance of Adaptive RED (ARED) and Refined Adaptive RED (Re-ARED) by proposing two new AQM mechanisms. First, we idemonstrate that neither ARED’s conservative approach alone nor Re-ARED’s aggressive approach alone suffices to improve the throughput and reduce the packet drop rate. Hence, we have designed and implemented two new AQM mechanisms, namely “Fast Adapting RED (FARED)” and “Cautious Adaptive RED (CARED)” to combine the benefits of ARED and Re-ARED. Extensive simulation results show that while FARED fails to achieve the desired goal, CARED offers robust performance in a wide variety of scenarios and outperforms ARED and Re-ARED. Unlike other RED variants, CARED requires only algorithmic modifications and is easy to deploy. Second, a new congestion signaling mechanism called “eXtended ECN (XECN)” is designed and implemented to provide richer feedback to the end-hosts. XECN does not require additional bits in TCP or IP header since it re-uses the bits already allocated for ECN efficiently and unambiguously. Moreover, it requires modifications only at the sender and the receiver. It does not require any modification in the working of the router. Third, we develop a new variant of TCP called “TCP Surathkal” which leverages the benefits of AQM and XECN. TCP Surathkal takes congestion control decisions based on the severity of congestion in the network. We present a modified design of a fluid model which is based on Poisson Counter Driven Stochastic Differential Equations to validate the working of TCP Surathkal. Results obtained by conducting extensive simulations and mathematical modeling show that TCP Surathkal achieves high utilization, reduces the packet drop rate and incurs less oscillations in the router queues. The advantages of TCP Surathkal make it suitable for deployment in a wide variety of networks.