Decouple and Decompose - Scaling Resource Allocation with DeDe

Learn about DeDe, a scalable optimization framework for large-scale resource allocation in cloud systems, presented as a conference talk at OSDI '25. Discover how researchers from Harvard University and University of Illinois Urbana-Champaign address the challenge of growing optimization problems that have outpaced commercial solvers in production environments. Explore the key insight that most real-world resource allocation problems are inherently separable, optimizing aggregate utility of individual resource and demand allocations under separate constraints. Understand the decouple-and-decompose approach that forms DeDe's core, which decouples entangled resource and demand constraints and decomposes overall optimization into alternating per-resource and per-demand subproblems that can be solved efficiently in parallel. Examine the implementation details of DeDe as a Python package with a familiar modeling interface and review experimental results across three representative resource allocation tasks: cluster scheduling, traffic engineering, and load balancing, demonstrating significant speedups while generating higher-quality allocations compared to existing approaches.