AN ENHANCED ECN-BASED CONGESTION MITIGATION WITH TRAFFIC REROUTING FOR CLOS DATA CENTER NETWORKS

Abstract

Explicit Congestion Notification (ECN) is widely used in modern data center networks to signal congestion without incurring packet loss. ECN is effective in reducing queue buildup and preserving high throughput. However, congestion signals are generated only after queue buildup has begun, which may delay mitigation in large-scale AI data centers where synchronized, high-throughput traffic flows are common. This paper examines the limitations of conventional ECN behavior in spine-leaf data center fabrics and proposes an enhanced ECN-based mechanism that utilizes in-network feedback and dynamically adapts Equal-Cost Multi-Path (ECMP) paths. By proactively rerouting traffic away from congested paths before widespread congestion propagation, the proposed approach aims to improve job completion time, reduce unnecessary rate throttling, and better utilize available network resources. Additionally, incorporating real-time buffer telemetry into this path-selection process provides a powerful extension that directly elevates fabric stability and load distribution. This work shows the potential of a closer integration of congestion notification and path selection mechanisms may help shorten congestion resolution and improve the efficiency of AI training and inference workloads in modern data center environments.