Design Challenges on Machine-Learning Enabled Resource Optimization

Nowadays, service providers' (SPs) need for efficient resource utilization solutions is more demanding than ever. The optimal use of the physical and virtual infrastructures guarantees that the waste of resources due to overdesign is minimized while the provided services enjoy the required quality of service levels. However, the prediction of the exact amount of the required resources per service at any time of its lifecycle is not an easy process. For this purpose, we propose a solution that handles the infrastructure in a holistic manner introducing a novel architecture that exploits the monitoring data from three layers (hardware, virtualization, and application) and uses them to train machine learning models, which can accurately predict the exact amount of the required resources per service. Its implementation using open-source tools and its performance are also presented.