Computation Offloading And Resource Allocation In Mobile Edge Networks

With the rapid development of wireless communication technology,mobile terminal devices,and data traffic are growing rapidly,which has led to a series of delaysensitive and compute-intensive applications,such as virtual reality,ultra-high definition video,and so on.However,due to the long service response delay,traditional centralized cloud computing can not meet the real-time processing requirements of applications.Mobile Edge Computing(MEC)deploys computing,storage,and other resources at the edge of the wireless network to provide users with adjacent services and improve the user experience.In mobile edge computing,the problem of limited local computing resources and battery capacity is solved by offloading computationintensive and delay-sensitive applications to edge servers through computation offloading.The inference task of neural networks is a typical computation-intensive and delay-sensitive application.How to perform high-performance neural network inference tasks in edge networks is a straightforward problem to study.Therefore,in this paper,we mainly study the computational offloading and resource allocation for neural network inference tasks in the edge network,and the main contributions include the following three aspects.1.Aiming at the problem that the terminal equipment can not meet the real-time processing requirements of the neural network inference task and the battery capacity of the terminal equipment is limited,this paper proposes a neural network inference task offloading algorithm based on deep reinforcement learning.Firstly,the structure of the neural network is analyzed and modeled as a directed acyclic graph considering the dependency between the nodes of the neural network,and then the reasoning task is split into multiple subtasks,which can be computed locally or offloaded to nodes with computing resources,thus making full use of the computing resources of terminal devices and edge nodes.Then,the delay and reliability models are established,and the optimization problem of minimizing the inference energy consumption is set.Then,the unloading process is modeled as a Markov decision process,and the unloading strategy is obtained by deep reinforcement learning to minimize the energy consumption of reasoning.Simulation results show that,when the delay requirement is 700 ms,compared with the random algorithm,particle swarm algorithm,and genetic algorithm,the inference energy consumption is reduced by 45.0%,36.3%and 35.8%,respectively.Under different delay requirements and environmental conditions,the proposed algorithm shows good performance.2.Aiming at the problems that the computing resources of edge servers are limited,the terminal devices are difficult to meet the real-time demand of reasoning,and there is resource competition among users,this paper proposes a joint decision-making algorithm of task offloading and resource allocation based on model division.Firstly,the neural network model is divided into the communication model and a delay model,and the optimization problem of minimizing the delay is set.Then,a resource allocation algorithm based on convex optimization is proposed to allocate the resources of partition nodes and edge nodes.Finally,the simulation results show that the proposed algorithm can effectively reduce the inference delay compared with the comparison algorithm.Taking the average inference delay of the system as an example,compared with the local inference,the average delay is reduced by about 41%.3.Aiming at the problem that terminal devices are difficult to meet the requirements of low latency and high accuracy of reasoning tasks,and the limited resources in the edge network,this paper proposes a task offloading and resource allocation algorithm based on early exit for multi-user edge scenarios.First,we model the neural network inference process based on the early exit,and then we set the optimization problem to maximize the inference accuracy and make a joint decision on the confidence threshold,partition point,and resource allocation.Simulation experiments show that the proposed algorithm can improve the task completion rate compared with benchmarks,and the accuracy loss is 4%compared with not using the early exit.