Learning-Based Multi-controller Coordination for Self-Optimization

A complex software system may have hundreds of tuning parameters with both itself and its runtime environments. To perform optimally at runtime, a self-adaptive system often needs to employ a series of controllers to tune different parameters. In most cases, these parameters are not independent but interact with each other, thus demanding effective coordination among different controllers. Due to system complexity and the uncertain and changing environments, there are usually no explicit relationships among relevant controlled parameters. In this paper, we propose a learning-based approach to achieve effective coordination among multiple controllers for self-optimization. Based on specific controllers for different controlled parameters, our approach uses an additional coordinator to adaptively switch among different controllers. A learning-based algorithm is adopted to continually evaluate the effectiveness of each controller, providing decision basis for the coordination. The results of our experimental study with a Web-based system show that our approach can significantly improve the effectiveness of self-optimization with coordination among multiple controllers.