Enhance Database Query Performance through Software Optimization and Hardware Adaptation
Author(s)
Cao, Jiashen
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
As database system queries incorporate more complex analytics like object detection powered by machine learning algorithms, leveraging the power of GPUs has become essential to ensure good query performance. However, without careful query planning and hardware-aware execution, database queries can underutilize GPU resources. To fully unlock GPU potential, my thesis focuses on improving query performance through a combination of software-level optimizations and hardware-aware adaptations. My solution integrates thoughtful query planning with optimized execution tailored to the GPUs.
My first exploration focuses on optimizing query planning, particularly for video database systems, where machine learning algorithms are often used for data analysis. To avoid overusing the computationally intensive algorithms, we introduce a fine-grained strategy that selects the just-enough accurate algorithm to improve the speed without compromising accuracy. It does so by evaluating an ensemble of algorithms on sampled data and choosing the fastest one that meets the accuracy requirement. We also extend the classic adaptive query processing techniques to dynamically reorder predicates involving machine learning algorithms to reduce the overall query processing time.
My second research endeavor centers on improving hardware utilization for database systems. Using roofline analysis, we study GPU resource usage during relational query execution and identify key performance bottlenecks. Based on the insights, we propose optimizations that reduce data movement and increase execution concurrency, outperforming state-of-the-art systems. As part of Aero, we further extend the adaptive query processing technique to improve the GPU utilization for machine learning queries—an area previously overlooked. Finally, to support the growing use of large language models in database systems, we introduce LIRS-M, a GPU-aware buffer management policy designed to better manage memory and enhance overall performance.
Sponsor
Date
2025-05-02
Extent
Resource Type
Text
Resource Subtype
Dissertation