System identification and control system design for relative performance management and resource provisioning of virtualized software system

The increasing range of applications and services maintained by software systems has motivated the growing popularity of virtualization technology as a framework for the intensification of computing performance. Virtualization enables multiple independent systems to use a shared infrastructure at the same time. It is very challenging for multiple virtual machines (VMs) to run applications with different performance objectives and under unpredictable workload changes. Many concerns have been raised, especially regarding the increasing resource utilization and the sensitivity of performance properties. Consequently, it is essential to automate the management tasks such as managing performance properties dynamically at runtime while sharing a limited amount of resources. Some of the main challenges include the nonlinear characteristics of the system, limited resources, differentiated performance objectives, and workloads uncertainty. This thesis aimed to address these issues by implementing system identification and control engineering techniques for relative performance management and dynamic resource provisioning using the principles of optimization and feedback control. An experimental testbed of virtualized software system is established to generate real observational data and to confirm performance of the proposed approaches.<br><br>In this thesis, the dynamic of a virtualized software system is characterized in linear and nonlinear functions, through block-oriented system identification. The nonlinear functions from input and output elements are utilized as nonlinear compensator functions in the structure of a feedback control loop. The novelty of the proposed system identification is the model estimation of an output nonlinear model in a reduced parameter model of B-Spline function based on $k-means$ clustering approach. This approach reduces the impact of nonlinearities on the output response stability of feedback control system.<br><br>In addition, three control methods have been designed and implemented for performance management; PI-based feedback control, Data-driven control and Finite Control Set - Model Predictive Control. The control performances are evaluated in the testbed with different scenarios of workload and performance objective references. The experimental results have shown that control systems with pre-input and post-output nonlinear compensation provide robust performance and significant improvement in the stability of relative performance management in virtualized software system.