Using nonlinear control of resources to achieve differential performance objectives in software systems

In many competitive business domains, software systems have become vital to achieve the business objectives efficiently. In such software systems, maintaining performance properties such as response time and throughput at runtime is important to avoid customer dissatisfaction and violation of service level agreements. This is a challenging task as service providers typically need to share computing resources between service consumers in order to deliver those services efficiently under dynamic and unpredictable environmental conditions. Managing such systems using human-in-the-loop decision making methods at runtime is neither efficient nor cost-effective. As a result, runtime performance management tasks need to be automated. Closed-loop approaches based on control engineering methodologies have been widely investigated, as a way to achieve relative and absolute performance management objectives at runtime, while sharing a limited amount of resources. These approaches are based on linear modelling and control methods. However, linear approaches neglect the prominent nonlinear dynamics of the relative and absolute performance management systems and provide effective control only in a limited operating range. In this thesis, we classify the nonlinearities that exist in the relative and absolute performance management schemes. We then introduce two novel nonlinear feedback control methods to reduce the runtime impact of nonlinearities on the control system. In the first approach, compensators are integrated into the control system to reduce the impact of nonlinearities. In particular, a Hammerstein-Wiener block-oriented model is used for relative performance management while a MIMO Wiener model is used for absolute performance management. In the second approach, we represent the dynamics of the nonlinear system with multiple linear models. Multiple models and multiple linear controllers are implemented together with a switching scheme, to select the most suitable controller to provide control under the current operating conditions. In addition, we present a class library of control components, to facilitate the implementation of complex control systems for software systems. The evaluations conducted using simulation studies and experimental real-world case studies indicate that the proposed nonlinear approaches can significantly improve the performance and resource management capabilities compared to other state-of-the-art approaches. We further demonstrate that the class library significantly improves the efficiency of the control system engineering process.