Adaptive Learning For Model Driven Control Of Cloud Applications

System performance tracking, planning and on-demand provisioning are common techniques employed today for management of cloud applications. This work explores the use of machine learning algorithms to tune Kalman filters, which track system performance metrics to make predictions about the future state of running applications. The benefit of accurate predictions is to enable planning algorithms, for example model-based schemes like layered queueing networks, to determine potential system choke points with high granularity and specificity. The focus of this work is to evaluate the impact of long short term neural networks on improving the accuracy of Kalman filter predictions via tuning, to use these predictions to drive layered queueing network model driven decision control actions, and to evaluate the performance of this controller in different virtual machine types in the cloud. In addition, a new scheme for layered queueing network model parameter tracking using principal component analysis is proposed and evaluated.