RAS: A Resource Allowance System for Perpetual Region-wide Resource Allocation
summary
Facebook builds a region-wide allocation system called Resource Allowance System (RAS) that replaces many server pools with a two-level allocator and a move service that safely drains and reassigns servers. RAS uses mixed-integer programming with symmetry grouping to reduce the model to one to ten million variables and meets a one-hour SLO for solving. It enforces capacity, per-failure-domain buffers and rack-spread goals, tolerates ~10k correlated failures, and is already deployed on 80% of servers, achieving about 4.2% overhead vs a 2.8% ideal.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:12Traditional allocation problems
- 3:20Goals of the Resource Allowance System
- 5:06Regional capacity challenges
- 7:54RAS capacity management
- 10:45Algorithm design choice
- 13:10RAS placement policies
- 14:58Failure domain policies
- 17:05Failure domain example
- 19:16Rack spread results
- 20:22RAS status and future
related talks
17mApplication of Online Resource Allocation Based on Ray
06mRISC: Resource-Constrained Urban Sensing Task Scheduling Based on Commercial Fleets
21mR2P2: Making RPCs first-class datacenter citizens
17mTowards High-Performance Transactional Stateful Serverless Workflows with...
21mSvalinn: Overload Control in Large-Scale Servers with Multiple Resource Bottlenecks
20mGlobal Capacity Management
16mDistRS: Disaggregated Reward Service for RLVR with Batch-Level Constraint
18mA Planet-Scale Computer – Abstract Away Regions via Global Service Placer (GSP)