Thesis title: Dynamic Edge Computing and Learning
We live at the edge of a new revolution of wireless networks, that are going to transform from pure communication systems, to service enablers that build on the tight integration of communication, computation, caching and control. The aim of this work is to investigate possible synergies between communication and computing. In particular, we propose novel dynamic resource allocation strategies for computation offloading in the context of edge computing-aided wireless networks, to jointly orchestrate radio and computation resources. In the first part, we focus on the offloading of general applications, devising minimum energy strategies that meet average and probabilistic end-to-end delay constraints, by exploring a typical concept of wireless communication networks: the energy-delay trade-off. Differently from traditional communication problems, in edge computing scenarios, the delay comprises transmission and computation delays, so that it is convenient to consider, in this trade-off, both sources, possibly jointly. Then, in this work, constraints will be imposed on the overall delay experienced by the offloaded data, from its generation at the end devices (sensors, mobile users, car, etc.) until its computation at nearby edge servers. In dynamic scenarios, this translates into constraints on the sum of transmission queues at the end devices, containing data to be transmitted, and computation queues at the edge server.
In the first part, we will focus on user-centric computation offloading, in which our objective only takes into account end devices' energy consumption. Starting from long-term optimization problems (presented in separate sections and addressing different challenges), we hinge on stochastic Lyapunov optimization to solve them in a per-slot basis, with low complexity solutions for the instantaneous optimization problems. In particular, the proposed strategies do not require any prior knowledge of the statistics of time varying data arrivals and wireless channels. Using stochastic Lyapunov optimization, theoretical guarantees on system stability and long-term constraints are provided, as well as guarantees on the asymptotic optimality of the solutions.
In the second part of the work, we extend this analysis to a holistic view, in which the whole network energy consumption (comprising end devices, access point, and edge server) is taken into account. We will show how it is possible to achieve globally green solutions, without dramatically degrading the performance with respect to strategies that only optimize single agents' objectives.
In the last part, we focus on the specific case where the edge server runs a machine learning algorithm, thus introducing a new aspect in our trade-off: the performance/accuracy of the learning tasks. We will show how, diving into the specific requirements of the applications in terms of accuracy can help in achieving better trade-offs. Then, hinging on stochastic Lyapunov optimization, we will devise minimum energy strategies with delay and accuracy guarantees, in different scenarios and for different learning applications. Changing perspective, we will also show how to optimize accuracy under energy and delay constraints.
Several numerical results based on computer simulations will be presented for each part, to corroborate our theoretical findings and to show performance also when some theoretical assumptions are relaxed. Finally, a simple practical implementation of a Lyapunov-based resource allocation algorithm will be presented, showing the performance of the proposed algorithms in a real scenario.