Treffer: Data-Centric AI for Software Performance Engineering - Predicting Workload Dependent and Independent Performance of Software Systems Using Machine Learning Based Approaches

Context: Machine learning (ML) approaches are widely employed in various software engineering (SE) tasks. Performance, however, is one of the most critical software quality requirements. Performance prediction is estimating the execution time of a software system prior to execution. The backbone of performance estimation is prediction models, in which machine learning (ML) is a common choice. Two settings are commonly considered for ML-based performance prediction: workload-dependent and workload-independent performance, depending on whether or not the specific usage of the system is fed as input to the ML estimator.Problem: Developers usually manually understand the performance behaviour with respect to the workload. This process consumes time, effort, and computational resources since the developer repeats the running of the same tested system ( e.g. benchmark) many times, each with different workload values. In a workload-independent setting, predicting the scalar value of execution time based on the structure of the source code is challenging as it is a function of many factors, including the underlying architecture, the input parameters, and the application’s interactions with the operating system. Consequently, works that have attempted to predict absolute execution time for arbitrary applications from source code generally report poor accuracy.Goal: The thesis presents a modern machine learning-based approach for predicting the execution time from two angles: (a) workload-independent performance. (b) workload-dependent performance. Solution Approaches and Research Methodologies: To achieve the goal and tackle the problems mentioned earlier, we conducted a systematic empirical study to fill the gap of workload-dependant performance across five well-known projects in JMH benchmarking (including RxJava, Log4J2, and the Eclipse Collections framework) and 126 concrete benchmarks. We generated a dataset of approximately 1.4 million measurements. As for the poor accuracy challenges, we aim to increase the ...