Final Defense

GPU-Accelerated Processing of Graph Queries and Multi-Vector Retrieval

The Hong Kong University of Science and Technology (Guangzhou)

Data Science and Analytics Thrust

PhD Thesis Examination

By Mr. Guanghua LI

ABSTRACT

Graphics processing units (GPUs) have become an important platform for accelerating computation-intensive and data-intensive workloads. Their massive parallelism and high memory bandwidth benefit many data processing tasks, including graph query processing and multi-vector retrieval. Graph queries involve neighbor traversal and subgraph matching, which can benefit from parallel processing of graph data. Multi-vector retrieval represents each query and document with multiple vectors and aggregates similarities between query and document vectors to estimate relevance. It involves many vector similarity calculations that can be performed in parallel. However, efficient GPU utilization remains challenging due to their different computational characteristics, memory access patterns, and resource requirements.

This dissertation studies GPU-accelerated processing of graph queries and multi-vector retrieval. We first develop TenGraph, a tensor-based graph query engine that compactly represents graph topology and intermediate results and expresses subgraph matching and relational-style operations through optimized tensor operators. To address TenGraph’s memory traffic and scalability limitations, we develop VORA. Its VECTORFLUX library supports programmer-controlled GPU vector operator fusion. VORA also adopts workload decomposition for out-of-core subgraph query processing. We further develop VECTORHIT for multi-vector retrieval. It retrieves similar document vectors for each query vector and filters candidate documents based on their support, namely, the number of distinct query vectors that retrieve at least one vector from a given document. It further reduces document scoring cost by using a subset of query vectors for initial scoring of candidate documents. In summary, this dissertation develops query processing algorithms and execution strategies that account for both workload characteristics and GPU hardware capabilities. By exploiting parallelism while reducing redundant computation, data movement, and memory consumption, our techniques improve query processing efficiency and scalability.

TEC

Chairperson: Prof Jun YANG
Prime Supervisor: Prof Qiong LUO
Co-Supervisor: Prof Wei ZHANG
Examiners:
Prof Jeffrey Xu YU
Prof Jing TANG
Prof Hao LIU
Prof Hong CHENG

Note

Audience may attend only on-site.

Date

26 November 2026

Time

14:00:00 - 16:00:00

Location

E3-105, HKUST(GZ)

Event Organizer

Data Science and Analytics Thrust

Email

dsarpg@hkust-gz.edu.cn