Research On Edge Computing Network And QoE-Aware Resource Scheduling

The explosive growth of IoT applications has brought tremendous traffic pressure to the network,making traditional cloud computing architectures a major challenge.Edge computing,which is an emerging network infrastructure,extends cloud computing services to the edge of the network,reduces communication latency,easies the burden on the core network,and becomes an effective solution for IoT applications.General edge computing devices are developed on standard servers with powerful computing capabilities.However,with the development of single-board computers and lightweight virtualization technologies,more and more researches show that low-cost single-board computer clusters can be used to build cost-effective edge computing network.Single-board computers have the advantages of light weight,low cost,and low power consumption,but resource capacity limitations can affect the quality of service for applications.Efficient resource pool management is an effective way to solve this problem.On the one hand,reasonable resource scheduling can meet the needs of the application,on the other hand,optimize the use of resources and improve the efficiency of the edge computing network.This paper mainly studies the edge computing network and resource scheduling.The specific contents are as follows:(1)Inspired by the hyper-convergence infrastructure in data center,a lightweight edge computing network architecture with functional convergence as an important feature is proposed,and its corresponding networking scheme and implementation techniques of each level are given.The architecture is designed to use lightweight virtualization technology to develop edge nodes that combine network,computing and storage functions on a low-cost and low-power single-board computer to replace network devices and servers in traditional networking solutions.We use fiber optics to interconnect multiple edge nodes and provide network access via wireless.We also use SDN controllers for centralized network management and automated scaling.We built a test platform,tested the network performance of a variety of single-board platforms,and evaluated their performance when developing edge nodes with functional convergence.The test results can provide a reference for designing the lightweight edge computing networks in a variety of application scenarios.The experimental results comprehensively verify the feasibility of the architecture.When the single-board computer runs the network virtualization technology,it still maintains low delay and jitter under the condition of high link load and workload.In the case of tens of kilometers of fiber transmission,it also can achieve millisecond single-hop delay.(2)Based on the proposed lightweight edge computing network convergence architecture,a QoE-aware resource scheduling scheme is designed and implemented.The core decision module of the solution accepts real-time collected information of applications QoE and cluster resource usage,performs flexible resource reconfiguration and task scheduling based on evolutionary computation algorithm,and ensures the service quality of the applications while optimizing resource usage.We use the video archiving service based on web camera as an experimental case and build a test platform.The test results show that the proposed scheme can automatically increase the resource amount of the container instance when the demand is greater than the configured amount to ensure the service quality of applications,and reduce the resource amount when the demand is much smaller than the configuration amount so as to improve resource utilization.While changing the amount of virtual instance resources,the scheme performs global resource rescheduling to balance the CPU and memory resource load of each node in the cluster.In order to obtain better scheduling results,we tested genetic,particle swarm optimization,ant colony evolutionary algorithms and Kubemetes default scheduling algorithms in different experimental scales.The results show that the ant colony algorithm is better than the other two evolutionary computation algorithms in both convergence speed and optimization results.Although the Kubemetes default scheduling algorithm can achieve similar CPU and memory load balancing,the resource distance equilibrium is far less than the evolutionary computation algorithm.