Meta AI has open-sourced Rebalancer, a C++ assignment solver designed to address resource allocation challenges in hyperscale datacenters. The tool is currently available as version 1.0.4 via PyPI for Python 3.12+ and supports Linux x86-64 and macOS 14+ ARM64 architectures. It is also distributed through .deb, .rpm, and Homebrew packages.
The software addresses scalability and usability hurdles by separating problem specification from the solving process. It allows users to model tasks, servers, and racks using specific constraints like CapacitySpec and BalanceSpec. The system compiles these into a directed acyclic expression graph to manage complex allocation logic.
Rebalancer offers two primary solving methods: a mixed integer program (MIP) for small to mid-size problems, compatible with FICO Xpress, Gurobi, or HiGHS, and a local search solver for large-scale tasks. The local search engine is parallelized, enabling the processing of approximately 40 million placement problems daily.
To assist with debugging and model visualization, Meta developed Rebalancer Explorer, a Dockerized web interface. This tool provides insights into binding constraints and relaxation effects. While other tools like OR-Tools and Timefold exist, Meta highlights Rebalancer's ability to use a single assignment specification across both MIP and local search solvers.
