Thesis title: Proposal and Investigation of Artificial Intelligence Algorithms for Resource Allocation in Network Function Virtualization Architectures
The objective of the research work is the study and evaluation of allocation and reconfiguration systems of virtual resources in networks based on the Network Function Virtualization paradigm, for the creation and management of network services. In a first phase, traces of known traffic are taken into consideration and the allocation and reconfiguration of resources in a multi-provider cloud environment are studied. Orchestration systems are evaluated by creating optimization algorithms for resource allocation, together with a scalable architecture for their addressing, and the issue of reconfiguration of resources is studied, analyzing methods, costs and advantages, and proposing algorithms that take into account both the cost of resources and the effect that infrastructure management has on the Quality of Service.
The high computational complexity of the optimization algorithms leads to the proposal and evaluation of heuristics whose performances are compared in terms of execution time as the number
of service requests varies. The high time needed to reconfigure cloud resources in Network Function Virtualization network environments has led to the proposal of solutions in which a prediction-based resource allocation is performed. The further contribution introduced by the research work consists in proposing and evaluating an ETSI-compliant architecture for the resource allocation and reconfiguration, which integrates the prediction and allocation procedures by applying a prediction procedure with a custom loss function that instead of minimizing the prediction error, optimizes an asymmetric cost function modeled on the cost terms of the particular use case that the Service Provider has to face.
The proposed methodology can be applied for any prediction technique, both traditional and those based on more modern AI-integrated system. In this work the proposed methodology for the following two prediction techniques is analyzed: the first one based on Seasonal Autoregressive Integrated Moving Average (SARIMA) traditional models, the second one based on the application
of Long Short Term Memory (LSTM) neural networks. Performances of the proposed solutions will be analyzed in backbone and metropolitan network and traffic scenarios, in particular the proposed solution is tested with real mobile traffic of an Italian Operator.