Software-Hardware Co-Optimization for High-Performance, Resource-Efficient, and Secured GPU Cloud Platforms

PhD Qualifying Examination


Title: "Software-Hardware Co-Optimization for High-Performance, 
Resource-Efficient, and Secured GPU Cloud Platforms"

by

Mr. Yongkang ZHANG


Abstract:

To maximize resource utilization in data centers, cloud service providers 
often colocate high-priority, latency-sensitive (LS) GPU tasks with 
low-priority, best-effort (BE) GPU tasks—referred to as tenants—on 
the same GPU. While recent research has explored improving the quality of 
service (QoS), resource efficiency, and security in multi-tenant GPU cloud 
platforms through software and hardware innovations, there remains a lack of 
a systematic review from a software-hardware co-optimization perspective. 
This gap limits researchers' ability to holistically optimize multi-tenant 
GPU cloud platforms.

This survey addresses this need by first analyzing commodity GPU 
architectures and identifying key bottlenecks in guaranteeing QoS and 
security of GPU sharing. It then reviews existing optimization approaches for 
multi-tenant GPU cloud platforms—including software-based, OS-level, 
and hardware-level approaches—that aim to enhance performance, reduce 
resource costs, and bolster the security of multi-tenant GPU platforms. 
Finally, this work highlights promising research opportunities for 
co-optimizing software and hardware stacks, providing a comprehensive 
framework to bridge research gaps and inspire advancements in multi-tenant 
GPU sharing.


Date:                   Wednesday, 12 March 2025

Time:                   4:00pm - 6:00pm

Venue:                  Room 3598
                        Lifts 27/28

Committee Members:      Dr. Shuai Wang (Supervisor)
                        Prof. Xiaowen Chu (Co-supervisor, HKUST-GZ)
                        Dr. Binhang Yuan (Chairperson)
                        Dr. Xiaomin Ouyang