Re-architecting Congestion Management in Lossless Ethernet
Abstract
The lossless Ethernet is attractive for data centers and cluster systems, but various performance issues, such as unfairness, head-of-line blocking and congestion spreading, etc., impede its large-scale deployment in production systems. Through fine-grained experimental observations, we inspect the interactions between flow control and congestion control, and are aware that the radical cause of performance problems is the ineffective elements in the congestion management architecture for lossless Ethernet, including the improper congestion detection mechanism and inadequate rate adjustment law.Inspired by these insights and findings obtained in experiment investigations, we revise the congestion management architecture, and propose the Photonic Congestion Notification (PCN) scheme, which consists of two basic components: (i) a novel congestion detection and identification mechanism to recognize which flows are really responsible for congestion; (ii) a receiver-driven rate adjustment method to alleviate congestion in as short as 1 RTT. We implement PCN using DPDK NICs and conduct evaluations using testbed experiments and simulations. The results show that PCN greatly improves performance under concurrent burst workload, and significantly mitigates PFC PAUSE messages and reduces the flow completion time under realistic workload.