Content area
As data in a database management system (DBMS) becomes increasingly complex, there is a growing need for sophisticated analytical techniques—such as deep neural networks—capable of extracting meaningful insights. However, these methods are often compute-intensive, necessitating the use of specialized accelerators such as the graphic processing unit (GPU) to achieve practical performance.
Given this trend, modern DBMSs frequently integrate GPU-based query execution. For a good performance, we argue that query planning must be approached judiciously, ensuring that advanced algorithms are employed cautiously so as not to abuse the underlying hardware. Once an optimal plan is devised, the next critical step is to refine its implementation to leverage the GPU’s computational resources fully. By combining judicious query planning with optimized execution, we can unlock the full potential of modern hardware accelerators in database systems.
My first research endeavor focuses on improving the query plan to avoid overusing compute intensive algorithms for machine learning-based queries, particularly queries in video DBMSs. FiGO proposes a fine-grained approach where just-enough accurate deep learning algorithms are used to analyze the data without compromising accuracy. To pick the right algorithm, FiGO uses an ensemble of algorithms, evaluates all algorithms on sampled data, and applies the fastest but yet still accurate algorithm to analyze. Aero proposes to apply the classic adaptive query processing (AQP) framework on those emergent queries. It utilizes runtime statistics to dynamically reorder predicates that involve deep learning to minimize the query processing time.
My second research endeavor focuses on improving hardware utilization, particularly GPUs in DBMSs. The GPU database systems characterization and optimization paper analyzes the resource utilization of GPUs during query execution and their performance bottlenecks. Based on the roofline analysis, we propose optimizations to reduce the total data access and increase the query execution concurrency to improve the performance, delivering better performance than the state-of-the-art systems. As part of Aero, we also extend the AQP framework to also promote the GPU utilization during machine learning (ML) query executions, which previously was not done through the AQP framework. Lastly, to facilitate the recent demand of using the large language model (LLM) in DBMSs, we propose tailored buffer management policy called LIRS-M to better utilize the GPU memory resource for a good performance.