Events

A Holistic View on Machine Learning for Systems

Improving computer system performance and resource efficiency are long-standing goals. Recent approaches that use machine learning methods to achieve these goals rely on a predictor that predicts the latency, throughput, or energy consumption of a sub-computation to, for example, aid hardware resource management or scheduling.

In this talk, I will present a holistic view on machine learning for systems. I will demonstrate that the optimization goals between machine learning methods and systems problems do not always align, and this misalignment means that optimizing machine learning prediction accuracy does not optimize system behavior. Instead, my research vision focuses on a holistic view of machine learning for systems pipeline. The key insight in achieving this vision is making proper tradeoffs between different stages within the pipeline. Based on this vision, I will introduce a couple of machine learning for systems solutions to meet different systems’ goals including energy, performance, and interpretability. I will conclude the talk with my future directions.