Time: 3pm Friday 4/25
Location: Klaus 2100
Online: Microsoft Teams
teams.microsoft.com
Jiashen Cao
Ph.D. Candidate
School of Computer Science
College of Computing
Georgia Institute of Technology
Committee
Dr. Joy Arulraj (co-advisor), School of Computer Science, Georgia Institute of Technology
Dr. Hyesoon Kim (co-advisor), School of Computer Science, Georgia Institute of Technology
Dr. Divya Mahajan, School of Electrical and Computer Engineering and Computer Science, Georgia Institute of Technology
Dr. Kexin Rong, School of Computer Science, Georgia Institute of Technology
Dr. Rathijit Sen, Microsoft Gray System Lab
Dr. Tarique Siddiqui, Microsoft Research
Abstract
Modern database systems incorporate machine learning (ML) algorithms within queries to deliver richer insights and support complex business logic. However, the inclusion of these compute-intensive algorithms introduces new performance bottlenecks. My thesis focuses on addressing these challenges through a combination of software-level optimizations and hardware-aware adaptations. In my talk, I will highlight two projects.
The first project, Aero, introduces an enhanced adaptive query execution framework to produce better query plans and more efficient GPU resource management. Traditional databases rely on static, ahead-of-time query optimization, which struggles to optimize ML queries due to the difficulty of collecting accurate statistics beforehand. Moreover, optimal hardware utilization often requires dynamic, runtime decisions, which is impossible in the static optimization setting. Aero addresses both of these challenges in a unified framework and achieves significant performance improvements over prior approaches.
The second project shifts focus toward supporting large language model (LLM) analytics within database systems. Specifically, we target scenarios where LLMs are too large to fit entirely in GPU memory. In such cases, prior offloading-based approaches suffer from CPU-GPU bandwidth limitations, which become the new bottleneck. To overcome this, we propose a GPU-centric buffering strategy paired with an LLM-aware buffer replacement policy significantly, improving both performance and buffer hit rates.
]]>