CCCR: Combining CNP and RTT for congestion control in datacenter networks

Abstract

With the rapid development of cloud computing, AI, and big data, data center networks face challenges in achieving ultra-low latency, high bandwidth, and stability. Many data centers still rely on traditional switches, which lack programmable features for advanced congestion control algorithms. In this environment, existing algorithms like DCQCN and TIMELY face two major challenges: (1) a single congestion signal (such as ECN or RTT) struggles to accurately reflect network conditions, leading to delayed congestion detection; (2) heuristic rate control strategies are prone to causing network fluctuations and slow convergence, making it difficult to meet the demands of high-bandwidth links. To address these issues, we propose CCCR, a congestion control algorithm that combines ECN (via CNP) and RTT signals. CCCR enables rapid, accurate rate reduction using receiver-side feedback and employs a adaptive rate increase based on minimum, average, and target RTT. It also adjusts in-flight data using per-flow BDP estimation. Simulations show that compared to DCQCN, TIMELY, and Swift, CCCR reduces the average flow completion time by 11%, 20%, and 12% respectively in incast scenarios, with better fairness than HPCC, and achieves up to 82% reduction in tail flow completion time for medium flows and up to 74% for long flows. In large-scale simulations, CCCR achieves comparable performance to programmable switch-based HPCC algorithms.