Abstract
TCP is essential for reliable Internet data transfers. However, its slow-start mechanism, designed to estimate the optimal transmission rate, can lead to bandwidth under-utilization and severe congestion in modern networks with large Bandwidth-Delay Products (BDP). This issue is particularly problematic for applications relying on small data transfers, such as web browsing and social networking.
This thesis presents a new theoretical framework that reveals the intricate dynamics of TCP slow-start, explores the interactions among TCP’s operational parameters and provides insights into their collective impact on performance. It highlights the formation of queues along the connection path, offering a deeper understanding of a congestion event and the extent of its dispersion along the path. Empirical experiments validate the framework’s practical relevance and accuracy.
Building on the insights obtained from the theoretical framework, SUSS, a sender-side lightweight mechanism, is designed to enhance the performance of small-size data transfers. SUSS mitigates bandwidth under-utilization during the slow-start phase by accelerating the increase in the transmission rate based on whether continual exponential growth is predicted for sub- sequent round-trip times (RTTs). Implemented in the Linux kernel, our real-world experiments across various device types and Internet locations demonstrate that SUSS consistently outperforms the traditional slow-start, achieving over a 20% improvement in flow completion time in all experiments with flow sizes less than 5 MB in large-BDP networks.
This thesis also presents a novel mechanism called StopEG to accurately detect the optimal time for stopping the exponential growth of the data transmission rate in the TCP slow-start phase. We show that theoretically the number of in-flight packets in the forward path is no more than 56.8% of all the in-flight packets when the bottleneck link is unsaturated, and use this value as the threshold to stop the exponential growth. StopEG is integrated into Google’s BBR algorithm and evaluated in the ns-3 simulator. Simulation results show that StopEG significantly reduces the bottleneck queue length by approximately 68% during the initiation of new connections, thereby mitigating queuing delays and the risk of buffer overflow, and improving the overall efficiency of the network.