An efficient framework of congestion control for next-generation networks

The success of the Internet can partly be attributed to the congestion control algorithm in the Transmission Control Protocol (TCP). However, the tremendous growth in the range of bandwidth-delay products, Bit-Error Rates, and the increased diversity in applications has stressed TCP and the need for new transport protocol designs has become increasingly important. Prior research has focused on the design of either end-to-end protocols (e.g., CUBIC) that rely on implicit congestion signals such as loss and/or delay or network-based protocols (e.g., XCP) that use precise per-flow feedback from the network. While the former category of schemes have performance limitations, the latter are hard to deploy, can introduce high per-packet overhead, and open up new security challenges. This dissertation explores the middle ground between these designs and makes four contributions. First, we study the interplay between performance and feedback in congestion control protocols. We argue that congestion feedback in the form of aggregate load can provide the richness needed to meet the challenges of next-generation networks and applications. Second, we present the design, analysis, and evaluation of an efficient framework for congestion control called Binary Marking Congestion Control (BMCC). BMCC uses aggregate load feedback to achieve efficient and fair bandwidth allocations on high bandwidth-delay networks while minimizing packet loss rates and average queue length. BMCC reduces flow completion times by up to 4x over TCP and uses only the existing Explicit Congestion Notification bits. Next, we consider the incremental deployment of BMCC. We study the bandwidth sharing properties of BMCC and TCP over different partial deployment scenarios. We then present algorithms for ensuring safe co-existence of BMCC and TCP on the Internet. Finally, we consider the performance of BMCC over Wireless LANs. We show that the time-varying nature of the capacity of a WLAN can lead to significant performance issues for protocols that require capacity estimates for feedback computation. Using a simple model, we characterize the capacity of a WLAN and propose the usage of the average service rate experienced by network layer packets as an estimate for capacity. Through extensive evaluation, we show that the resulting estimates provide good performance.