Lightweight Traffic Engineering ​for the AI/ML Fabrics

Kevin Wang from Juniper Networks presents lightweight traffic engineering solutions for AI/ML fabrics, focusing on Global Load Balancing (GLB) and Deterministic Path Forwarding (DPF). GLB optimizes load balancing decisions using link node feedback, while DPF proactively plans traffic to avoid congestion by separating a physical IP fabric into multiple logical fabrics using BGP session coloring. The presentation covers the technical details of both solutions, including topology discovery, link quality exchange, and implementation in three-stage and five-stage clonal networks. Following the presentation, an audience Q&A addresses simulation validation, handling message failures in link quality exchange, performance metrics for different AI/ML workloads, standardization efforts for the router information protocol, and the node-level focus of traffic engineering.