A Planet-Scale Computer – Abstract Away Regions via Global Service Placer (GSP)
summary
Meta builds a system that autonomously places services across regions using a placement planner and executor, formulating placement as a constrained bin packing problem solved by the rebalance solver and executed via ICSP and Sharp integrations. It enforces tenant latency SLOs, accounts for capacity and HW heterogeneity (H100 can be ~2× faster than A100 for some models), adapts to rapid regional GPU supply shifts, and reduces global GPU card counts for most inference models, delivering measurable cost and capacity efficiency.
AI-written summary · about
chapters (12)
- 0:00Planet-scale cloud vision
- 2:57GSP vision and solution
- 5:50GSP design principles
- 6:34GSP high-level architecture
- 7:59Placement planner and bin-packing
- 9:04Placement executor and safety
- 10:29GSP adoption across services
- 10:50AI inference case study
- 13:02AI inference placement flow
- 14:48ICSP execution and materialization
- 16:13GSP adoption results
- 16:35Key takeaways and future
related talks
20mGlobal Capacity Management
27mLive Panel: Building at the Roofline — Planning and Delivering Gigawatt-Scale AI Infrastructure
18mArchitecting Multi-tenant Data-center Networks for GPU Customers
02mStarcluster
27mFrom Supercomputing to Serving: A Case Study Delivering Cloud Native Foundation Models
49mScaling Data in a Sovereign AI Platform
21mCapacity Planning for AI Training Hardware
42mBuilding a green supercomputer in the cloud