Energy Efficient Vehicular Fog Computing

More than 2 billion smart vehicles are predicted to be on the roads by 2048. These vehicles, equipped with computation and communication units, can be integrated with cloud and fog computing to provide services with reduced power consumption and latency at the network edge, referred to as vehicular fog computing (VFC). Vehicular fogs (VFs) can serve various services including artificial intelligence (AI)-based services in a virtualised environment.
This thesis investigates the placement of multiple services over an architecture with VFs formed by clustering stationary vehicles in addition to cloud and fog resources. We investigate the energy efficiency of software matching, where user tasks require a specific software package installed in processing nodes in the proposed architecture. A mixed integer linear programming (MILP) model is developed to optimise the service placement considering software matching in the architecture so the power consumption is minimised. We evaluate the impact of software package availability, processing demand, vehicle density of VFs, and task splitting on the power consumption. The results show power savings of up to 81.8% when enabling task splitting and deploying software packages following a Zipf distribution in the high-density VF, compared to processing all requests at the cloud.
We also investigate the energy efficiency of virtual networking embedding (VNE), considering mapping the processing and networking requirements of virtual network requests (VNRs) into the proposed architecture. Furthermore, we investigate the embedding of neural network (NN)-based services abstracted as virtual NN-based service requests (NSRs) over the proposed architecture. We use a multi-objective MILP model to optimise the resource allocation so the trade-off between power minimisation and latency minimisation is investigated. The results show that the joint minimisation of power consumption and latency achieves a power saving of 67% compared to latency minimisation and a latency reduction of 38% compared to power minimisation.