|Thesis abstract: |
This thesis proposes a methodology to assess and improve energy efficiency and quality of service in a data center. After an analysis and classification of the main approaches related to green information systems, we focus on the application level and propose an approach based on four steps: monitoring, evaluation, adaptation and updating. Our context consists in a business process in which each activity is placed on a dedicated virtual machine. Each virtual machine is deployed on a server of a monitored data center. The first step consists in assessing the state of the system, both in terms of energy efficiency and quality of service. This has been done by defining and selecting a set of metrics able to describe these aspects. Thresholds are defined over the values of these metrics: if violated, the system is in a sub-optimal situation and a repair strategy needs to be enacted. We explored relations between metrics in order to be able to act directly and indirectly over the violated variables. This has been done using a Bayesian Network describing relations among the states of the variables in the system. These relations are automatically learned from data. A learning strategy is also proposed to select the best adaptation action to react to a non optimal situation. This selection depends on the context in which the action is taken. The effect of the action is observed and used to update information about its impact. Since learning on a real and running system can be dangerous for its performance, a simulation environment has been developed. In the simulation, the system and the interactions between its components have been modeled. Its parameters can be learned collecting data from a real system. Once validated, the simulation environment can be used to analyze the relations between indicators and the impact of the actions in an off-line mode. Only when reliable, actions will be applied to solve inefficiencies in the real system.