Reproduce “Multi-Agent Reinforcement Learning for Network Load Balancing in Data Center” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhiyuan Yao, Zihan Ding, Thomas Clausen
Researchers from École Polytechnique, Cisco Systems, and Princeton University developed a Multi-Agent Reinforcement Learning (MARL) framework for data center network load balancing, demonstrating superior or competitive job completion times compared to heuristics on a real-world testbed with up to 24 servers. The framework reduced 90th-percentile Job Completion Time by 1.44x to 5.11x over strong baselines for different traffic types.
Key insights
- MARL, specifically QMIX with centralized training and decentralized execution, can effectively learn adaptive load balancing strategies from partial observations to optimize global performance in dynamic data center environments.
- Leveraging flow duration as a proxy for server load is an effective strategy for learning agents to make informed decisions without direct access to application-layer information.
- Decoupling the frequency of complex RL policy updates from the microsecond-level per-flow forwarding decisions is crucial for meeting stringent real-time latency requirements in high-throughput network systems.
- The communication overhead and synchronization challenges inherent in centralized training for MARL can impact performance and scalability, particularly in very large, distributed systems.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.