| Edge computing is a promising solution to support high-quality time-sensitive applications.However,due to users’ highly dynamic mobility,service migration has become a focusing challenge in edge computing systems.Therefore,in this thesis,we investigate the dynamic service migration strategy(i.e.,whether,when,and where to migrate the services)to seamlessly serve mobile users in edge computing systems.Firstly,the service migration is formulated as an optimization problem to minimize the system service delay including computation delay,communication delay,and migration delay in the long term.Then,considering the complex network scenarios,frequent user movement,and large-scale decision space caused by users,the formulated problem is difficult to solve in real time.To this end,we propose a Mobility-Aware Service Migration scheme,named MSM.It utilizes a data-driven user grouping mechanism and deep reinforcement learning(DRL)approaches,to make real-time service migration decisions in complex network scenarios.Specifically,the MSM scheme includes the following two parts:1)The design of association patterns based user grouping mechanism.The problem can be formulated as a 0-1 high-dimensional nonlinear integer programming when considering the service migration for multiple users.Moreover,in the real world,the number of users is increasing day by day.The complexity of multi-user service migration with individual users as scheduling granularity increases explosively with the increase of network scale,and the burden of the system increases sharply.To address the challenge,we first collect the fine-grained and long-term user Wi-Fi traces in a large-scale Wi-Fi system.Then various metrics are comprehensively used to analyze the user association patterns,combined with user behaviors and network conditions.After that,we propose a grouping management mechanism for users based on the patterns and the mobility characteristics mentioned above,where the user services with potential migration value are scheduled according to groups.Then,the service migration problem is simplified to a group service migration problem,and the burden of system scheduling is reduced.2)The design of a DRL-based group service migration strategy.Future information such as network environment and user trace can’t be obtained in advance,which makes it difficult to optimize the group service migration in the long term.To address the challenge,MSM introduces a reinforcement learning framework,which explores the correlation among user movement,network,and optimal service migration location through the interaction between agent and environment.In addition,to deal with the large-scale discrete action space caused by service migration,MSM expands the agent observation information to 2D space and adjusts the DRL network structure.In this way,MSM can effectively learn the optimal migration policy and reduce the system service delay.Finally,extensive data-driven experiments show that the MSM scheme reduces the system service delay by at least 14.6 % compared with benchmarks.Furthermore,with the decrease in computing resources or the increase in unit migration delay,MSM achieves more performance improvements. |
