Strategies for User Allocation and Service Placement in Multi-Access Edge Computing

Abstract

The emergence of Multi-access Edge Computing (MEC) grants service providers the ability to deploy services at edge servers near base stations to mitigate the effects of high network latencies often encountered in cloud-based system deployments. As users move around, their application service invocations are routed to proximate MEC servers en route to curtail the high latencies of cloud communication networks. In contrast to cloud servers, edge servers have constraints on resources such as computation, storage, energy, etc. Embedded devices often function as edge servers which are quite less flexible and resource impaired when compared to their full-fledged cloud server counterparts when hosting services. Thus, placement and allocation of services on edge servers and binding user service requests to the service instances hosted on the edge pose a number of research challenges. Also, the movement of users in the edge environment leads to the challenge of migration of service data and placement of hosted services. To efficiently use the available edge server resources and handle the mobility of users, an edge user allocation policy is designed. An edge user allocation policy determines how to allocate service requests from mobile users to MEC servers. An efficient edge user allocation policy is quite challenging to design due to the influence of an extensive variety of factors like the mobility of users, considerations of optimal Quality-of-Service (QoS) and Quality-of-Experience (QoE), variable latencies, stochastic nature of user service requests, limited resources, device energy constraints and so forth. This thesis predominantly focuses on the user allocation and service placement problems in MEC with an objective to provide efficient and scalable solutions. Classical MEC policies that bind user service requests to edge servers, seldom take into account user preferences of QoS and the resulting QoE. In our first contribution, we propose a novel user-centric optimal allocation policy considering user QoS preferences, with an attempt to maximize overall QoE. Furthermore, traditional allocation and placement policies cater to service request allocation and placement without much consideration of workload fluctuations. To address such issues, the second contributory chapter of this thesis proposes a variation aware stochastic model for user service allocation. In addition, current state-ofthe-art techniques assume MEC resource utilization to be linearly dependent on the number of service request demands and usages, i.e. the combined resources utilized by a group of user services is the sum of service resource utilization per user. In our third contributory chapter, we propose a real-time on-device learnable Reinforcement Learning (RL) framework to design user allocation policies that accommodate the non-linear nature of resource utilization by services. We implemented our proposed approaches on real-world datasets and analyzed the performance of our proposed algorithms to demonstrate the efficiency of our proposals. We believe our work will open up a lot of new research directions and applications of learning based methods in the MEC context.