Poster: Time Window ECN Filtering for Multiple-bottleneck RDMA Congestion Control

In RDMA data center networks with multiple bottlenecks, a sender may overestimate level of congestion because ECNs are marked independently at multiple bottleneck switches along the path. This congestion overestimation problem is difficult to address because the sender cannot identify where congestion signals are originated. In this paper, we propose a selective ECN marking method in which each switch filters ECN marks based on local congestion severity. Simulation results show that our method improves fairness and prevents excessive rate reduction in multiple-bottleneck scenarios.