DCTCP-FQ: Enhancing Fairness and Convergence Time in Data Center Congestion Control
Haoyu Wang, Xiaoqian Zhang, Allen Yang, Bo Sheng · 2025
Efficient congestion control is vital for the performance of modern data centers, which host diverse and latency-sensitive applications. Data Center TCP (DCTCP) is a well-established algorithm that utilizes Explicit Congestion Notification (ECN) to achieve low latency and high throughput. However, DCTCP has drawbacks regarding convergence time and fairness, especially when managing flows that start at different times when there is already congestion. To address these limitations, we introduce DCTCP-FQ, an enhanced version of DCTCP that incorporates an active queue management strategy. DCTCP-FQ employs selective packet marking at congestion points, targeting flows with transmission rates exceeding the average sustainable rate. This strategy enables faster convergence for laterstarted flow and improved fairness without compromising the throughput and latency benefits of DCTCP. Extensive simulations demonstrate that DCTCP-FQ achieves significant improvements in convergence time and flow fairness compared to traditional DCTCP in the scenario where a congestion point adds a new flow. According to our evaluations, DCTCP-FQ archives about an average of 64.82% faster in flow completion time than DCTCP for smaller size flow ranging from 200KB to 1000kb under a 1Gbps congestion point and about 43.91% increase in Jain’s fairness index in multi-hop multi-bottleneck flows topology for mixed starting flows, making it a robust solution for next-generation data center networks. This paper details the design, implementation, and evaluation of DCTCP-FQ, highlighting its potential to enhance network performance and fairness in data center environments.