Global Capacity Management
summary
Meta explains a system that turns global datacenters into a single schedulable computer by offering Global reservations that use an assignment solver to allocate machines (example: a 300‑server request becomes 150+150+150 across three regions = 450 total for one‑region fault tolerance or 30 in each of 11 regions = 330 with more regions), and a Regional fluidity pipeline that attributes demand to front‑end drivers, plans multi‑region shifts, and an orchestrator that sequences safe moves; next work is a capacity orchestrator for stateful and sharded systems.
AI-written summary · about
chapters (13)
- 0:00Global Capacity Management intro
- 1:29From regional to global thinking
- 2:54Meta cluster management
- 3:55Latency sensitive vs tolerant
- 4:37Global reservations for tolerant
- 7:28How global reservations work
- 9:34Global capacity for sensitive
- 10:37Demand attribution & fluidity
- 13:03Solver-driven regional placement
- 15:08Beyond two service types
- 16:12Vision for automated global
- 17:39Future orchestration plans
- 19:01Open challenges & next steps
related talks
16mGlobal Capacity Management With Flux
20mGlobal Capacity Management with Flux
30mCapacity Management for Fun & Profit
18mA Planet-Scale Computer – Abstract Away Regions via Global Service Placer (GSP)
21mCapacity Planning for AI Training Hardware
18mAvoiding and Breaking Out of Capacity Prison
29mCapacity Constraints Unveiled: Navigating Cloud Scaling Realities
20mManaging Demand and Efficiency for Meta Multi-tenant Services