• big-data
• data-analysis
• performance-modeling
• scalable-computing
• system-management

The advent of the cloud computing paradigm has lead Internet service providers to set up large Internet systems distributed all over the world in order to efficiently serve needs of their customers. Large Internet systems are characterized by a huge number of hardware resources and software components having the goal to make computing services readily available to users on demand, like any other utility service available in today's society.

Efficient management of large Internet systems requires several strategies that decide on request dispatching, load balance, admission control, and request redirection without direct intervention of human administrators. At the basis of most autonomic management decisions there is the need of performance models for supporting system management by taking real-time decisions on the basis of information related to the state of internal system components and resources. Performance models supporting large infrastructures must be able to operate at different time scales and should support prompt reconfigurations motivated by continuous dynamic changes in system, client, and business policies. Performance models should be scalable for increasing numbers of hardware and software components, they should be adaptive to heterogeneous data streams characteristics, and they should guarantee reliable results over changing system conditions and requirements.

This thesis presents a set of novel performance models proposed for online management of large amounts of variable and heterogeneous data streams coming from several system monitors and networks. In particular, this thesis extends the state-of-the-art in performance modeling in manifold directions: (i) it presents a novel approach for improving scalability when managing large amounts of data; (ii) it introduces new adaptive models for the on-line detection of anomalies and relevant state changes in highly variable contexts, and for the identification of correlations and groups of related objects even when correlation is hidden by high variability; (iii) it uses predictive analytics for improving performance in the selection of cloud availability zones on the basis of user preferences.

Extensive evaluations of the proposed approaches in real systems demonstrate improvements in performance modeling with respect to the state-of-the-art solutions and the satisfaction of scalability, adaptivity, and reliability requirements that are mandatory for large systems management.

The advent of the cloud computing paradigm has lead Internet service providers to set up large Internet systems distributed all over the world in order to efficiently serve needs of their customers. Large Internet systems are characterized by a huge number of hardware resources and software components having the goal to make computing services readily available to users on demand, like any other utility service available in today's society. Efficient management of large Internet systems requires several strategies that decide on request dispatching, load balance, admission control, and request redirection without direct intervention of human administrators. At the basis of most autonomic management decisions there is the need of performance models for supporting system management by taking real-time decisions on the basis of information related to the state of internal system components and resources. Performance models supporting large infrastructures must be able to operate at different time scales and should support prompt reconfigurations motivated by continuous dynamic changes in system, client, and business policies. Performance models should be scalable for increasing numbers of hardware and software components, they should be adaptive to heterogeneous data streams characteristics, and they should guarantee reliable results over changing system conditions and requirements. This thesis presents a set of novel performance models proposed for online management of large amounts of variable and heterogeneous data streams coming from several system monitors and networks. In particular, this thesis extends the state-of-the-art in performance modeling in manifold directions: (i) it presents a novel approach for improving scalability when managing large amounts of data; (ii) it introduces new adaptive models for the on-line detection of anomalies and relevant state changes in highly variable contexts, and for the identification of correlations and groups of related objects even when correlation is hidden by high variability; (iii) it uses predictive analytics for improving performance in the selection of cloud availability zones on the basis of user preferences. Extensive evaluations of the proposed approaches in real systems demonstrate improvements in performance modeling with respect to the state-of-the-art solutions and the satisfaction of scalability, adaptivity, and reliability requirements that are mandatory for large systems management.