CUDA Accelerated Graph Algorithms for Key Class Detection

Key classes are deemed as pivotal elements within a software system, serving as focal points for reengineering or documentation endeavors. The identification of these key classes holds significant importance in contemporary practices, with numerous studies dedicated to automating their detection based on representations within class graphs. Research indicates that employing algorithms such as Hyperlink-Introduced Topic Search (HITS) and PageRank (PR) yields optimal precision and recall performance in identifying key classes. However, the runtime execution of these algorithms becomes critical, particularly when operating on graphs with varied weights attributed to class relationships. To address the challenge of runtime execution, we explore parallel implementations of these algorithms utilizing CUDA, invoked from a Java application through JCuda. Specifically, we investigate two approaches: i) employing Java virtual threads and ii) utilizing CUDA threads within the context of the JCuda library. CUDA has fundamentally transformed how we harness GPU acceleration across diverse computational tasks, spanning parallel processing, deep learning, and high-performance computing. Our experiments are conducted on a data set comprising 14 Java projects. Our findings reveal that the hardware parallel CUDA threading model significantly accelerates attribute computation, achieving a runtime reduction of 95% to 97% compared to the virtual threading model.