DSCC: Dynamic synergistic congestion control for lossless RDMA datacenter networks

Published in Computer Networks, Volume 280, 112151, 2026

Large-scale incast traffic generated by AI training tasks presents significant challenges to RDMA networks. We propose DSCC, a dynamic and synergistic congestion control algorithm that combines PFC and ECN. Building upon the AI ECN framework, DSCC adjusts the PFC threshold according to the degree of incast while adhering to the PFC and ECN threshold constraints. It dynamically tunes both the growth rate and triggering threshold of PFC in response to the current congestion level at the switch. Simulation results in ns-3 demonstrate that under incast traffic scenarios, DSCC achieves up to 19.2% reduction in flow completion time (FCT) compared to AI ECN.

Recommended citation: Jianxing Zhuge, Zeming Gao, Ye Tian, Jun Wang, Shaoxuan Yun, and Xiangyang Gong. (2026). "DSCC: Dynamic synergistic congestion control for lossless RDMA datacenter networks." Computer Networks, 280, 112151.
Download Paper