CONGESTION CONTROL IN AI DATACENTERS: LEVERAGING ECN AND PFC FOR IMPROVED NETWORK EFFICIENCY

The surge in AI applications has sparked intense demand for high-performance computing in data centers. As these workloads expand, network congestion has become a central bottleneck, limiting performance, efficiency, and reliability. In this paper, we’ll explore Explicit Congestion Notification (ECN) and Priority Flow Control (PFC) as prominent solutions for congestion management in AI data centers. We focus on how ECN and PFC behave in AI data center networks and how they can be configured to complement one another. Our discussion reviews existing deployments and research, points out the challenges that operators face in practice, and outlines design choices that balance latency, throughput, and fairness for GPU-driven applications. While ECN and PFC each have specific advantages, a combined approach is vital for addressing the complex requirements of AI workloads, balancing latency, throughput, and fairness.