An Investigation on Data Center Congestion Control Algorithms

Abstract

With the rapid growth of data-intensive applications, congestion control algorithms for datacenter networks under RDMA over Converged Ethernet protocol have become vital in managing various traffic patterns that demand ultra-low latency and high end-to-end throughput. Although many rule-based and learning-based algorithms have been proposed to enhance throughput and reduce latency, challenges still remain in ensuring fair and efficient control under dynamic, bursty, and variable-sized flows which are prevalent in datacenter networks. In this thesis, we analyze the traffic control mechanism existing in modern data center networks and propose the Fair Datacenter Congestion Control (FDCC) algorithm based on deep reinforcement learning that leverages Long Short-Term Memory (LSTM) and historical memory augmentation to enable predictive control. FDCC operates from a window-based perspective, and introduces a staged reward design to effectively learn from network states based on in-band network telemetry acquired from various sources along the flow path to achieve fair congestion control under dynamic traffic. We perform extensive experiments under a variety of datacenter traffic patterns, including incast flows, long-short flows, and real-world traffic flows. The results suggest that our approach significantly enhances fairness for dynamic flow patterns over a range of state-of-the-art rule-based and learning-based congestion control algorithms, while maintaining comparable flow completion time and goodput.