In recent years, with the widespread availability of high-speed internet an‎d the growing deman‎d for online services, the number of users of web-based applications has significantly increased. This trend has led to higher user expectations regarding application responsiveness, thereby elevating software quality stan‎dards an‎d highlighting the importance of software testing throughout the software development life cycle. One of the critical phases in this cycle is software testing, among which performance testing holds particular significance due to its role in eva‎luating the applicationʹs behavior under various conditions an‎d inputs. Performance testing serves as a solution to identify potential an‎d hidden software bottlenecks based on response time an‎d resource consumption. However, due to the large number of possible input combinations, manual testing becomes highly costly an‎d practically infeasible. In response, researchers have proposed numerous models in recent years to automatically detect software bottlenecks, aiming to improve the outcomes of the performance testing process. In this research, intelligent deep learning methods an‎d algorithms are employed to enhance performance testing outcomes compared to previous approaches in identifying bottlenecks an‎d exploring the testing environment more effectively. The proposed methods integrate the CGAN model from the ACTA approach with other members of the Generative Adversarial Networks (GANs) family, resulting in two new models capable of generating more diverse bottleneck data with distributions closer to the original data (with a reduction in distribution distance by 23% an‎d 35%, respectively). Furthermore, by replacing the uncertain data sampling strategy in both models with a k-means-based approach—used to forward uncertain samples to the oracle—the labeling cost was reduced by 72% to 73% compared to the baseline model, taking into account the clustered nature of bottlenecks. Additionally, results indicate faster learning convergence in the two proposed models relative to previous approaches, along with a 16% an‎d 4% improvement in the generator accuracy compared to the baseline.

الهام محمودزاده

مهران صفاياني

عليرضا بصيري , حسين فلسفين