Can Adaptive Feedforward Control Improve Operation of Cloud Services?

International audienceTo process Big Data, the cloud computing architecture is used since several years. Service Level Objectives (SLO) become key indicators for the performances of the cloud computing providers. Controlling the cluster size is one of the main issues. Static decision procedures are available but they fail to provide an optimal decision in the presence of highly time varying service demands. The cluster has to be viewed as a dynamic system where the service demand is a disturbance to be compensated dynamically by a control action which is in this case the cluster size.PI feedback control has been already considered as a mean for dynamically controlling the size of the cluster. However since the disturbance is measurable an adaptive feedforward compensation can be added for improving the performance of a feedback controller. A general algorithm for adaptive feedforward control in the context of cloud computing is proposed and analyzed. Simplified versions are also presented and analyzed. Experimental results obtained on the Grid’5000 (French nation-wide cluster infrastructure) will be presented