An experimental analysis of TCP congestion control algorithms within virtualized environments

Abstract

In the growing landscape of virtualized networks, the performance of TCP congestion control algorithms remains a critical factor in ensuring efficient data transmission. This dissertation presents a comparative and experimental analysis of four prominent TCP congestion control algorithms - Vegas, CUBIC, BBRv2, and DCTCP – in virtualized environments. Motivated by the need to understand how these algorithms work in vir- tualized environments, this study investigates their behavior in various scenarios with varying network conditions, including baseline performance, under basic network fail- ures, and in competitive scenarios. This study differs from others found in the literature by evaluating virtualization scenarios and using a physical testbed environment instead of simulations to evaluate the performance of TCP congestion control algorithms. The testbed consists of dedicated physical servers and network devices configured to emu- late various network conditions. This configuration enables precise control and repro- ducibility of experiments, providing accurate measurements of key evaluation metrics: sending rate, throughput, throughput fairness, round trip time (RTT), and retransmis- sion rates. The findings indicate that, in virtualized environments, algorithms such as Vegas, CUBIC, DCTCP, and BBRv2 exhibit unique performance characteristics that af- fect network efficiency and reliability. Factors such as resource sharing and overhead between virtual machines impact the algorithm's performance. Delay-based algorithms such as Vegas are more affected by virtualization-induced latency. At the same time, CUBIC's window growth strategy can lead to suboptimal performance due to increased queuing delays in virtual switches. BBRv2's balanced approach is well suited to the dy- namic conditions imposed by virtualization but can be affected by additional processing overhead and variable latency. The study concludes that no algorithm universally out- performs the others in all scenarios. Instead, the choice of congestion control algorithm should depend on the context, considering specific network conditions and performance requirements. This dissertation contributes to understanding the dynamics of TCP con- gestion control in virtualized environments, offering insights that can guide the selection and optimization of these algorithms to improve network performance. Based on the findings, network administrators managing virtualized environments should select TCP congestion control algorithms according to specific operational needs. Vegas is ideal for minimizing latency, CUBIC and DCTCP for maximizing throughput, and BBRv2 for maintaining fairness and adaptability in dynamic network conditions. Furthermore, the study reveals that the virtualization context introduces an additional layer of complexity when deploying these algorithms in cloud-based scenarios. This critical distinction highlights the need to account for the unique challenges posed by virtualization when evaluating and optimizing TCP performance in modern data center environments.