Improving Programming Support for Hardware Accelerators Through Automata Processing Abstractions

Abstract

The adoption of hardware accelerators, such as Field-Programmable Gate Arrays, into general-purpose computation pipelines continues to rise, driven by recent trends in data collection and analysis as well as pressure from challenging physical design constraints in hardware. The architectural designs of many of these accelerators stand in stark contrast to the traditional von Neumann model of CPUs. Consequently, existing programming languages, maintenance tools, and techniques are not directly applicable to these devices, meaning that additional architectural knowledge is required for effective programming and configuration.

Current programming models and techniques are akin to assembly-level programming on a CPU, thus placing significant burden on developers tasked with using these architectures. Because programming is currently performed at such low levels of abstraction, the software development process is tedious and challenging and hinders the adoption of hardware accelerators.

This dissertation explores the thesis that theoretical finite automata provide a suitable abstraction for bridging the gap between high-level programming models and maintenance tools familiar to developers and the low-level hardware representations that enable high-performance execution on hardware accelerators. We adopt a principled hardware/software co-design methodology to develop a programming model providing the key properties that we observe are necessary for success, namely performance and scalability, ease of use, expressive power, and legacy support.

First, we develop a framework that allows developers to port existing, legacy code to run on hardware accelerators by leveraging automata learning algorithms in a novel composition with software verification, string solvers, and high-performance automata architectures. Next, we design a domain-specific programming language to aid programmers writing pattern-searching algorithms and develop compilation algorithms to produce finite automata, which supports efficient execution on a wide variety of processing architectures. Then, we develop an interactive debugger for our new language, which allows developers to accurately identify the locations of bugs in software while maintaining support for high-throughput data processing. Finally, we develop two new automata-derived accelerator architectures to support additional applications, including the detection of security attacks and the parsing of recursive and tree-structured data. Using empirical studies, logical reasoning, and statistical analyses, we demonstrate that our prototype artifacts scale to real-world applications, maintain manageable overheads, and support developers' use of hardware accelerators. Collectively, the research efforts detailed in this dissertation help ease the adoption and use of hardware accelerators for data analysis applications, while supporting high-performance computation.