| In the era of big data,mobile applications emerge in an endless stream with increasingly rich functions,and complex computing tasks lead to a shortage of mobile device resources.The latency of offloading computing tasks to the cloud server is high,which cannot meet the needs of mobile devices.Mobile edge computing is a potential solution,where servers deployed at the edge of network can provide computing services to users with lower latency.But limited computing resources and highly dynamic mobile devices bring challenges to the efficient scheduling of computing tasks.To this end,this paper aims to maximize the use of edge resources to reduce user computing costs,explore the impact of user dynamics on the system through data-driven research,and study the collaborative scheduling of user tasks in multi-node edge systems.The main work and innovations of this paper are as follows:(1)For the task dynamic scheduling problem in the multi-server and multi-user edge computing system,this paper firstly transforms it into an optimization problem that minimizes the computing cost of user tasks through computational cost modeling.Then based on data-driven exploration of user characteristics in real network systems: imbalance distribution and real-time movement,which brings many challenges to optimization problem solving.First,the imbalance distribution of users leads to imbalance server load.For this reason,the task transfer and server grouping mechanism is introduced in this paper.At this time,the optimization goals include task offloading decision and server grouping decision.Secondly,the realtime movement of users leads to dynamic changes in the edge scene,which requires that the problem can be solved in real time.Therefore,this paper decomposes the original problem into task offloading control problem and server grouping problem to reduce the difficulty of solving.(2)For the task offloading control problem,this paper first proves that it is a convex optimization problem,but in dynamic edge scenarios,the future number of users and the user task arrival rate are unknown.Therefore,this paper designs a user quantity prediction algorithm based on long-shortterm memory network and a user task arrival rate estimation algorithm based on Kalman filter.After eliminating the unknown parameters,the problem is solved by gradient descent method to obtain the task offloading decision.For the server grouping problem,this paper designs a dynamic grouping algorithm of the Deep Double-Q network based on Dueling architecture to group edge servers,and achieve load balancing through task transfer within the group.In order to avoid the problem that the algorithm is difficult to converge due to the large action space during grouping,the algorithm divides the grouping action into multiple steps.The final simulation results show that the algorithm proposed in this paper can effectively reduce the task computing cost and improve the system resource utilization. |
