Date of Award

2026

Document Type

Open Access Dissertation

Degree Name

Doctor of Philosophy in Computer Science (PhD)

Administrative Home Department

Department of Computer Science

Advisor 1

Zhenlin Wang

Advisor 2

Junqiao Qiu

Committee Member 1

Soner Onder

Committee Member 2

Benjamin Ong

Abstract

Pattern matching, driven by Finite Automata (FA) and Finite State Machines (FSM), is a fundamental computational kernel in modern unstructured data processing, spanning data analytics, bioinformatics, and network security. While Graphics Processing Units (GPUs) offer massive hardware parallelism, their potential is often hindered by the inherent sequential dependencies of state transitions and irregular memory access patterns. This dissertation addresses these bottlenecks by proposing a suite of speculation-centric parallelization frameworks, systematically transforming traditionally memory-bound automata processing into efficient GPU operations.

The research is presented through three progressive frameworks, which systematically advance from standard deterministic models, to alternative bitwise execution paradigms, and finally to complex non-deterministic processing:

First, to optimize foundational Deterministic Finite Automata (DFA) computations, this dissertation introduces GSpecPal. This framework addresses the high costs of mis-speculation and inefficient sequential recovery. By utilizing speculative data forwarding and heuristic thread scheduling, GSpecPal increases parallelism during verification and recovery and improves thread utilization, achieving an average speedup of 7.2× over the evaluated state-of-the-art implementation.

Second, to further mitigate memory bottlenecks by shifting to a compute-oriented execution model, we present SpecBit. This framework reformulates DFA transitions into compact bitwise computations via logic synthesis and transition reduction. To break sequential dependencies in bitstream evaluations, SpecBit employs a dual-level speculative parallelization scheme, improving throughput by up to 495.4× and tail latency by up to 16× over the evaluated GPU- and CPU-based regex engines.

Third, to tackle the increased computational complexity of Non-deterministic Finite Automata (NFA), we introduce ANG. Advancing beyond deterministic models, ANG addresses the performance degradation caused by insufficient state-level parallelism. By employing a tailored speculative parallelization scheme, ANG exploits fine-grained, input-chunk-level parallelism to significantly reduce processing latency for real-time analytics.

Extensive evaluations using diverse, real-world datasets demonstrate the impact of these frameworks. Experimental results show substantial throughput improvements and latency reductions over the evaluated state-of-the-art GPU-based automata engines across a broad range of workloads, although the magnitude of these gains varies by workload and baseline. Overall, the results demonstrate that speculation-centric parallelization is a scalable and effective approach to improving GPU utilization in complex automata processing workloads.

Download

Available for download on Tuesday, August 03, 2027

Share

COinS