Machine learning powers an emerging family of intelligent applications in many domains, from retail and transportation, to finance and healthcare. To support computation-intensive machine learning tasks, enterprises usually build large-scale GPU clusters, which presents challenges on the performance and efficiency of the machine learning systems. On the other hand, machine learning itself is also used to improve the system performance. In this talk, I will share two of my projects. The first one is PipeSwitch, which is published in OSDI’20. It allows multiple deep learning applications to time-share the same GPU with millisecond-scale switching overhead. With PipeSwitch, GPU utilization can be significantly improved without sacrificing SLOs.  The second project is RegexNet, which is published in S&P’21. It adopts a deep learning model, which is updated constantly in a feedback loop during runtime, to detect ReDoS attacks, which is a Denial-of-Service attack exploiting vulnerable regular expressions.

Zhihao Bai is a 5^th year Ph.D. student at Johns Hopkins University, advised by Professor Xin Jin. He works on computer systems, with a focus on machine learning systems. Before entering Johns Hopkins University, he received B.S. from Shanghai Jiao Tong University in 2017, and he was a member of ACM Honored Class, Zhiyuan College. He published many papers on top conferences, including USENIX OSDI (2020) and IEEE Security and Privacy (2021). He also received USENIX FAST Best Paper Award in 2019.