Abstract
Driven by the visions of smart devices, IoT (Internet of Things), and 5G communications, recent years have witnessed the emergence of Edge Computing which pushes computing functionalities from the centralized Cloud towards the network edge. Mobile Edge Computing (MEC) is one of the major Edge Computing paradigms that place computation resources at cellular base stations or wireless access points to support computation-intensive and latency-critical applications at mobile devices. The existing literature has studied various topics in MEC systems, e.g., computation offloading, computing/radio resource scheduling, user mobility management, energy saving, security, and privacy protection, to deliver high-quality edge service to end-users. This dissertation focus on an equally important design topic, service placement/provisioning, which is still largely under-investigated in the literature. We propose novel implementation scenarios and design optimization and learning techniques for delivering better computing service on MEC platforms.
Service placement refers to configuring the service platform and storing related libraries at edge servers such that the corresponding service requests can be executed. This problem is a unique issue in MEC systems because the limited computing capacity of edge servers cannot support all possible service applications. Therefore, which services to place on edge servers have to be judiciously decided to optimize the performance of MEC systems. We formally define the service placement problem and consider a number of key issues including the service heterogeneity, uncertainty of system dynamics, spatial and temporal demand coupling, and decentralized coordination in edge systems. An optimization based solution, called Collaborative Service Placement (CSP), is studied and its goal is to reduce the reliance on cloud computing and improve the resource efficiency in MEC systems by promoting collaboration among edge servers. CSP is an efficient distributed algorithm that facilitates the decision-making of service placement in distributed MEC systems.
A service provisioning problem is further investigated for MEC systems. It is in essence a service placement problem from the perspective of Application Service Providers (ASP). The crux in service provisioning is a resource rental problem --- an ASP decides the amount of computing resource to rent on edge servers for delivering its application services. Such a framework enables the ASP to better realize the value of its service and also allows edge systems to monetize their computing resources. Oftentimes, the benefit of deploying edge service depends on users' service demand patterns. A major issue here is that users' service demand pattern is unknown to ASPs a priori and varies temporally and spatially across geographically distributed edge sites. To address this problem, we design a novel multi-armed bandit framework, called contextual combinatorial multi-armed bandit, for learning the users' demand patterns and optimize ASP's resource rental decisions on-the-fly based on learned knowledge.
Besides the uncertainty in user service demand, the service provisioning problem also exhibits complex correlations with multifarious factors in MEC systems, ranging from user behavior to computation offloading, which are difficult to be fully captured by mathematical modeling and also put off traditional machine learning techniques due to the induction of high-dimension state space. The recent success of deep learning (DL) underpins new tools for addressing our problem. While previous works provide valuable insights on applying DL techniques, e.g., distributed DL, deep reinforcement learning (DRL), and multi-agent DL, in MEC systems, these techniques cannot solely handle the distributed and heterogeneous nature of EC systems. To address these limitations, we propose a novel framework based on multi-agent DRL, distributed neural network orchestration (N2O), and knowledge distilling. The multi-agent DRL enables edge servers to learn deep neural networks that shelve distinct features learned from local edge sites and hence caters to the heterogeneity of EC systems. N2O coordinates edge servers in a fully distributed manner toward a common goal of maximizing ASP's reward. It requires only local communications during execution and provides provable performance guarantees. The knowledge distilling is further utilized to distill the N2O policy for reducing the communication overhead and stabilizing the decision-making.