high performance computing on graphics processing units: hgpu.org

This thesis introduces the development of a new GPU-based database to accelerate queries of Digital Humanities data to extract document texts that are then data-mined to produce visualizations of aspects of the humanities data. The goal is to advance the state-of-the-art in massively parallel database work by investigating methods for utilizing graphical processing units in database systems. This thesis advances prior work done in the field of high-performance, massively-parallel databases. Some prior work focused on fixed length data types such as integers and doubles, often coupled with straight-forward single table queries. Other work focused on using primitives that are not a component of standard SQL databases. This thesis introduces an efficient virtual database engine that executes the majority of database operations directly on the GPU. The GPU database executes a subset of SQLite’s SELECT queries, which are typically the most computationally expensive operations in a transactional database. This database engine extends existing research by exploring methods of string operations, multiple table joins, indexing, and varying length data types and sets. This thesis discusses the design of a new GPU virtual database engine, which is developed using the CUDA extensions to C/C++, from loading data from the file on disk to processing the program on the GPU. This thesis focuses on the development of new improvements to deal with caching data for the GPU, processing and coalescing varying length data, and performing joins between multiple tables in the GPU database. The GPU database is demonstrated in a real world application. The application wraps the database in a graphical user interface which facilitates data selection. The application performs data mining of Humanities data for common text mining algorithms. The mined data is processed into visualizations to illustrate the resulting data digests.