ML-ECN: Multi-Level ECN Marking for Fair Datacenter Traffic Forwarding

Abstract

Recently, Explicit Congestion Notification (ECN) has been leveraged by most Datacenter Network (DCN) protocols for congestion control to achieve high throughput and low latency. However, the majority of these approaches assume that each switch port has one queue while current industry trends towards having multiple queues per each switch port. To this end, we propose ML-ECN, a Multi-Level probabilistic ECN marking scheme for DCNs enabled with multiple-service, multiple-queue switch ports. The main design of ML-ECN relies on the separation between small, medium, and large flows by dedicating multiple queues for each flow class to ensure fairness enqueueing. ML-ECN employs a single threshold for each queue in small service-queue class and multiple thresholds with a probabilistic marking for each queue in medium and large service-queue classes to achieve low latency for mice (small) and high throughput for elephant (large) flows. In addition, ML-ECN performs fairness-aware ECN marking that ensures small flows never get marked at early queue build up. Large-scale ns-2 simulations show that ML-ECN outperforms existing approaches for different performance metrics.