Java Single vs. Platform vs. Virtual Threads Runtime Performance Assessment in the Context of Key Class Detection

Key classes are considered the most important classes of a software system. They represent the starting point for reengineering or documentation processes. The detection of key classes is considered crucial in the state of the art, many studies are focused on automatic detection based on class graph system representation. Studies show that class attributes computed with algorithms like Hyperlink-Introduced Topic Search (HITS) and PageRank (PR) give the best precision and recall performance values in detecting key classes. The runtime execution of the two algorithms is critical when they run on graphs having different class relationship weights. To ameliorate the time execution problem we experiment with the two algorithm parallel implementations based on Java: i) single thread, ii) platform or operating system threads and iii) virtual threads. The experiments are fulfilled on a set of 14 Java projects. The results show that single thread implementations for project having a relative small number of classes, namely under 1,200, perform better than platform threads implementations. Conversely, virtual threads perform better than any single thread implementation. We conclude that virtual thread model speeds up the computation of attributes with a runtime decrease of 58.41% against the single thread model.