Reproduce “Oort: Efficient Federated Learning via Guided Participant Selection” on OpenResearch

About this paper

Fan Lai, Xiangfeng Zhu, Harsha V. Madhyastha, Mosharaf Chowdhury

Oort, a federated learning system from the University of Michigan's SymbioticLab, introduces guided participant selection to enhance both training and testing efficiency. It achieves up to a 14.1x speedup in time-to-accuracy during training and improves final model accuracy by up to 9.8% by intelligently selecting clients based on their statistical and system utility, while also enabling precise control over data distribution for model testing.

Key insights

• Optimizing federated learning performance requires dynamically balancing a client's statistical utility (how much their data improves the model) and their system utility (how quickly they can complete tasks).
• An adaptive mechanism (pacer) is crucial to navigate the trade-off between speed and model quality during training, allowing the system to admit slower, high-utility clients as needed.
• Effective federated model testing demands explicit control over the test data distribution, which can be achieved efficiently even at scale through statistical bounds and heuristic optimization rather than computationally expensive methods like MILP.

Summary by alphaXiv.