5781 Evolving FBOSS for the Next Gen AI Fabric

This podcast explores how Meta has evolved its FBOSS network management system to tackle the demands of next-generation AI workloads. The innovative solution, known as Disaggregated Scheduled Fabric (DSF), employs a credit-based scheduling approach and cell spraying across fabric links to achieve near-optimal load balancing and enhanced bandwidth. Significant upgrades were made to both the DSF SDK and FBOSS, including a new control plane for neighbor resolution and improved failure management in the distributed system. Performance tests demonstrated impressive improvements in bandwidth-heavy scenarios and a smooth response to failures, setting the stage for larger and more scalable data center architectures.