Phd Thesis Defense : Artificial Intelligence for Resource Allocation in Multi-Tenant Edge Computing

We consider in this thesis Edge Computing (EC) as a multi-tenant environment  where Net- work Operators (NOs) own edge resources deployed in base stations, central oces and/or smart boxes, virtualize them and let third party Service Providers (SPs) – or tenants – dis- tribute part of their applications in the edge in order to serve the requests sent by the users. SPs with heterogeneous requirements coexist in the edge, ranging from Ultra-Reliable Low Latency Communications (URLLC) for controlling cars or robots, to massive Machine Type Communication (mMTC) for Internet of Things (IoT) requiring a massive number of connected devices, to media services, such as video streaming and Augmented/Virtual Reality (AR/VR), whose quality of experience is strongly dependant on the available re- sources. SPs independently orchestrate their set of microservices, running on containers, which can be easily replicated, migrated or stopped. Each SP can adapt to the resources allocated by the NO, deciding whether to run microservices in the devices, in the edge nodes or in the cloud. We aim in this thesis to advance the emergence of real deployments of the “true” EC in real networks, by showing the utility that NOs can collect thanks to EC. We believe that this can contribute to encourage concrete engagement and investments engagement of NOs in EC. For this, we point to design novel data-driven strategies that efficiently allocate resources between heterogeneous SPs, at the edge owned by the NO, in order to optimize its relevant objectives, e.g., cost reduction, revenue maximization and better Quality of Service (QoS) perceived by end users, in terms of latency, reliability and throughput, while satisfying the SPs requirements.

This thesis presents a perspective on how NOs, the sole owners of resources at the far edge (e.g., at base stations), can extract value through the implementation of EC within a multi-tenant environment. By promoting this vision of EC and by supporting it via quantitative results and analysis, this thesis provides, mainly to NOs, findings that can influence decision strategies about the future deployment of EC. This might foster the emergence of novel low-latency and data-intensive applications, such as high resolution augmented reality, which are not feasible in the current Cloud Computing (CC) setting.

Another contribution of the thesis it that it provides solutions based on novel methods that harness the power of data-driven optimization. We indeed adapt cutting-edge tech- niques from Reinforcement Learning (RL) and sequential decision making to the practical problem of resource allocation in EC. In doing so, we succeed in reducing the learning time of the adopted strategies up to scales that are compatible with the EC dynamics, via careful design of estimation models embedded in the learning process. Our strategies are conceived in order not to violate the confidentiality guarantees that are essential for SPs to accept running their computation at the EC, thanks to the multi-tenant setting.