HGNN4Perf: Detecting Performance Optimization Opportunities via Hypergraph Neural Network
Ming Quan Fu, Minjie Wei, Minglang Qiao, Peng Ji, Zhihao Deng, Di Cui, Yutong Zhao · 2024
Performance optimization in software engineering is crucial for enhancing user satisfaction and maintaining a competitive advantage. Traditional methods for detecting performance issues – dynamic profiling and static analysis – often fall short in addressing complex dependencies within software architectures. This paper introduces the HyperGraph Neural Network (HGNN), a novel approach that leverages both static and dynamic program analysis to identify and prioritize performance bottlenecks effectively. By analyzing interconnected method call within fundamental patterns, HGNN and its enhanced version, HGNN+, utilize hypergraph neural network techniques to capture and learn dependency relationship features, significantly improving detection accuracy. Our initial testing on the three projects shows that HGNN+, especially when combined with several specific pre-trained models, presents satisfactory results compared to traditional methods.These findings underline HGNN+’s ability to manage complex dependencies and offer a scalable solution for software performance engineering. The benefits observed across multiple initial tests promise a broad applicability for the approach, setting a solid foundation for future research and expansion to more diverse datasets and neural network models, enhancing the reliability and effectiveness of performance optimization detection.