On the use of fuzzy logic controllers to comply with virtualized application demands in the cloud

Abstract

As virtualization technologies enable real-time CPU allocation, it is important to build controllers that adjust the allocation in a timely fashion avoiding resource saturation and hence, dissatisfaction of the end users of services. In this work, adaptive neuro-fuzzy inference, trained on Kalman and H∞ filters, has been used to adjust the CPU allocations based on observations of past utilization. When evaluating the performance of the proposed controller it is demonstrated that it provides even better performance than the filters it is trained on. In addition, there are no assumptions on the noise characteristics and due to the fact that the neuro-fuzzy controller can, in general, capture non-linear level processes, our controller is more robust than linear model based approaches, such as the Kalman and the H∞ filters.