Parameterisation of Markovian queueing models for IT systems

Modern IT systems are continuously growing both in size and complexity, thus making performance analysis and modelling an increasingly difficult task, if not intractable. As a result, in some cases the task has to be significantly simplified by focusing on selected system components or by reducing it to a smaller scale. One of the biggest challenges in system performance modelling is the accurate determination of service requirements. This thesis presents an investigation into novel statistical inference methods to effectively parameterise queueing models from system traces which exhibit diverse characteristics. I present two case studies, namely one of an Enterprise Resource Planning (ERP) application and one of a web server. Firstly, I propose a modelling methodology to address the limitations of service demand estimation based on CPU utilisation measurements by directly calibrating queueing models based on response time measurements. Secondly, with the aim of capturing more representative web workloads, I propose a methodology to parameterise queueing models for web server performance analysis that extracts hidden Markov models from real HTTP traffic via fitting. Experimental results indicate that all the methods developed as part of this thesis are highly competitive against state-of-the-art techniques.