Design of Congestion Control Algorithm based on Switch Queue Length Change Detection

Remote Direct Memory Access (RDMA) has emerged as the predominant networking technology and plays an important role in the Artificial Intelligence Generated Content (AIGC) industry, which requires effective congestion control algorithms to unleash its performance. Since the Data Center Quantized Congestion Notification (DCQCN) protocol was introduced in 2015, it has quickly become the standard congestion control solution across RDMA networks and an essential component of high-speed clusters and data center networks. However, as a congestion control (CC) algorithm that primarily relies on explicit congestion notification (ECN), DCQCN is challenged by delayed bandwidth convergence and sluggish congestion detection. To address these challenges, this paper proposes an optimized congestion control algorithm solution based on DCQCN. Specifically, we propose more accurate switch queue length monitoring, which is divided into two main parts: a congestion detection phase and a congestion control phase. Our algorithm utilizes the change rate of the switch queue length to refine and identify the degree of congestion and adjust the data sending rate after congestion occurs based on this change rate. Experimental results demonstrate that the proposed algorithm achieves better average flow completion time (FCT) in different scenarios including incast and actual network load, compared with DCQCN and other CC algorithms.