Automatic and Accurate Performance Prediction in Distributed Systems

System performance is getting attention by industry as it affects user experience, and much research focused on performance evaluation approaches. Profiling is the most straightforward approach to performance evaluation of software systems, despite being limited to shallow analyses. Conversely, software performance models excel in representing complex interactions between components. Still, practitioners do not integrate performance models in the software development cycle, as the learning curve is too steep, and the approaches do not adapt well to incremental development practices. In this thesis, we propose three approaches towards automatic learning of performance models. The first approach employs a Recurrent Neural Network (RNN) to extract a full Queueing Network (QN) model of the system; the second one calibrates a Layered Queueing Network (LQN) using an RNN; the third one presents μP, a framework that allows the user to develop microservice systems and obtain the corresponding LQN model from source code analysis. We considered the microservices architecture as it is embraced by influential players (e.g., Amazon, Netflix). Those approaches have two advantages: i) minimal user intervention to flatten the learning curve; ii) continuous synchronization between software and performance model, such as each software development iteration is reflected on the model. We validated our approaches on several benchmarks taken from the literature. The models we generate can be queried to predict the system behavior under conditions significantly different from the learning setting, and the results show sensible advancements in the quality of the predictions