Credit-R: Enhancing Credit-Based Congestion Control in Cross-Data Center Networks
Lunsheng Li, Yuang Chen, Hancheng Lu, Li He, Lei Gao, Ningcheng Wang · 2024
Hop-by-hop credit-based congestion control offers advantages of rapid convergence and strong congestion avoidance. Its low queue occupancy is particularly significant for data centers interconnected with shallow-buffered switches, ensuring zero packet loss and supporting long-distance link extension of any length. However, credit-based protocols suffer from excessive wastage of credits, leading to low credit utilization for small flows, which in turn results in low bandwidth utilization for long-distance links, making it difficult to apply to cross-data center networks. In this paper, we propose an innovative scheme named Credit-R, which aims at improving credit utilization and bandwidth utilization in cross-data center networks. In Credit-R, we maintain flow state information on external switches to identify and reuse wasted credits based on the sender phase. By leveraging the flow state information recorded in two kinds of flow tables, we reallocate these credits to the senders of active flows that can trigger data transmission, greatly enhancing the effectiveness of credit-based congestion control in long-distance cross-data center networks. Extensive packet-level simulations demonstrate that compared to state-of-the-art congestion control protocols, our scheme effectively reduces average flow completion time (FCT) by 25% to 79% under different realistic loads, and improves 99th tail performance by 27% to 81 % respectively. Moreover, Credit-R is able to safely speed up transmission during the pre-credit phase.