facebook / rebalancer

项目摘要

Rebalancer is a C++-based assignment solver library and domain-specific language (DSL) developed by Meta for solving large-scale resource allocation problems, such as placing ML training/inference workloads, balancing server loads, and routing traffic. It supports both local search heuristics for scalability and mixed-integer programming for optimal solutions.

为什么值得关注

Backed by Meta and used for critical large-scale infrastructure placement and load balancing, it offers a generic DSL for assignment problems with multi-threaded C++ performance and Python bindings, capable of handling problems with up to a million objects.

优势

• Scalable C++ core handling ~1M objects
• Flexible DSL for defining complex assignment rules and objectives
• Multiple solving algorithms including local search and MIP (Gurobi, Xpress, HiGHS)
• Python bindings available for broader accessibility

局限性

• Lacks automated tests in the repository
• No examples or Docker support provided
• Optimal solving relies on commercial solvers (Gurobi, Xpress)

使用场景

• ML training and inference workload placement
• Server and hardware resource allocation
• Traffic routing and load-balancing migrations