Distinguished Lecture - Automatic exploration of computer system designs
Jinyang Li, NYU
Title: Automatic exploration of computer system designs
Modern machine learning technology has matched or surpassed human intelligence in many tasks, ranging from playing the game of Go to predicting protein structures. Can ML do better than humans when it comes to designing high performance computer systems? The answer seems to be a promising “yes,” based on our group’s experience in applying a learning-inspired approach to databases and ML systems. A key challenge of this approach is crafting a system design space in which automatic exploration, often guided by some performance optimization goal, is effective. Another challenge is ensuring the correctness of the design.
In this talk, I will show how to design for automatic exploration by discussing two projects. The first considers the design of concurrency control algorithms for databases by developing a “policy space” of fine-grained synchronization actions to search for the best algorithm. The second investigates new rewrite rules for SQL query optimization. In this project, we propose a verifier to automatically check the correctness of potential rules during exploration.
Speaker Bio: Jinyang Li is a professor of computer science at New York University. Her research is focused on developing better system infrastructure to accelerate machine learning and web applications. Her group has developed DGL, a widely used open-source library for programming graph neural networks. Her honors include a NSF CAREER award, a Sloan Research Fellowship, and Google, Facebook, AMD research awards. She received her B.S. from National University of Singapore and her Ph.D. from MIT, both in Computer Science.