Automated Learning of Quantitative Software Models from System Traces

Models are primary artifacts in software system development. In particular, performance models allow us to evaluate and reason about extra-functional properties, such as the aver- age response time and throughput, for which meeting ade- quate quality levels is increasingly important. Indeed, per- formance quality is considered as essential as correctness in many practical development scenarios. Markov processes are valuable models for the qualitative analysis of performance. In this thesis, we will present statistical methods that learn Markov models directly from the running software system traces. We will focus on two classes of processes: with and without memory. In the first scenario, we aim to learn funda- mental performance metrics, i.e., service demands and rout- ing probabilities, using queuing networks (QN). For processes with memory, instead, we will exploit variable length Markov chains (VLMC) to capture data dependencies throughout the traces of system executions. The conducted numerical eval- uations, the presented in-depth study of the literature, and the performed appropriate comparisons with similar tools al- low us to demonstrate how the approaches presented in this work constitute a significant step forward concerning state of the art.