Disaggregated Scheduled Fabric: Scaling Meta’s AI Journey

This post introduces Disaggregated Scheduled Fabric (DSF), Meta's next-generation network fabric technology designed for AI training networks. It addresses challenges in traditional Clos-based networks by disaggregating line cards and fabric cards into distinct, interconnected hardware devices, creating a scalable, distributed system. DSF is a VOQ-based system powered by OCP-SAI and FBOSS, featuring packet spraying and a credit-based congestion control algorithm for near-optimal load balancing and lossless delivery. The post details the architecture for building massive clusters, including L1 zones (scaling units), L2 zones (dual-stage fabric), and L3 regions (super-spine interconnectivity), enabling clusters of up to 18K x 800G GPUs. It also introduces 'Input Balanced Mode' to manage traffic and prevent congestion during link failures by ensuring devices have equal or less input bandwidth compared to output bandwidth.