摘要： 容器云是5G边缘计算的重要支撑技术，5G的大带宽、低时延和大连接三大特性给边缘计算带来较大的资源压力，容器云编排器Kubernetes仅采集Node剩余CPU和内存两大资源指标，并运用统一的权重值计算Node优先级作为调度依据，该机制无法适应边缘计算场景下精细化的资源调度需求。面向5G边缘计算的资源调度场景，通过扩展Kubernetes资源调度评价指标，并增加带宽、磁盘两种评价指标进行节点的过滤和选择，提出一种基于资源利用率进行指标权重自学习的调度机制WSLB。根据运行过程中的资源利用率动态计算该应用的资源权重集合，使其能够随着应用流量的大小进行自适应动态调整，利用动态学习得到的资源权重集合来计算候选Node的优先级，并选择优先级最高的Node进行部署。实验结果表明，与Kubernetes原生调度策略相比，WSLB考虑了边缘应用的带宽、磁盘需求，避免了将应用部署到带宽、磁盘资源已饱和的Node，在大负荷与异构请求场景下可使集群资源的均衡度提升10%，资源综合利用率提升2%。

关键词: 5G网络, 边缘计算, 资源调度, 权重自学习, Kubernetes调度策略

Abstract: Container cloud is a key supporting technology for 5G edge computing,but edge computing faces great resource pressure imposed by the large bandwidth,low latency,and massive connections of 5G.The scheduler of container cloud,Kubernetes,only collects remaining CPU and memory of nodes,and uses a fixed weight to calculate the priority of nodes as the basis of scheduling.The mechanism cannot meet the demand for refined resource scheduling in edge computing scenarios.To addresses the resource scheduling needs of 5G edge computing,this paper expands the resource scheduling evaluation indicators of Kubernetes,adds bandwidth and disk evaluation indicators to filter and select nodes,and on this basis proposes a scheduling mechanism named WSLB,which realizes weight self-learning based on resource occupancy.WSLB dynamically calculates the resource weight set of the application according to its resource utilization during the running process to enable the weight set to dynamically and adaptively adjust itself based on the size of application traffic.The resource weight set obtained from dynamic learning is used to calculate the priority of candidate nodes,and the node with the highest priority is selected for deployment.Experimental results show that compared with the native scheduling strategy of Kubernetes,WSLB fully considers the bandwidth and disk requirements of edge applications,and avoids deploying applications to nodes where resources are all occupied.In the heavy load and heterogeneous request scenario,the balance of cluster resources under the WSLB mechanism is increased by 10%,the comprehensive utilization rate of resources increased by 2%.

Key words: 5G network, edge computing, resource scheduling, weight self-learning, Kubernetes scheduling strategy

中图分类号: 

• TP391