Research On The Strategy Of Resource Scheduling Based On Workload Prediction In Kubernetes Cluster

Kubernetes is an open-source container management system based on Docker container technology.With its powerful container orchestration capabilities,Kubernetes has now become the preferred solution in the container management field.However,the default resource scheduling scheme of Kubernetes cannot satisfy the needs of resource scheduling in complex scenarios.This article considers Kubernetes as the research object,introduces the core concepts and resource scheduling strategy of Kubernetes.A resource scheduling scheme based on load prediction is developed,the scheme adopts a predictive scheduling mechanism,and schedules resources in advance based on the predicted results.It can reduce application request response time and improve system service quality.This dissertation analyzes the load characteristics of applications in Kubernetes clusters.Based on empirical mode decomposition(EMD)and temporal convolutional network(TCN),a combined load prediction model EMD-TCN is established.The EMD algorithm is used to decompose the load data.then the TCN model is used to predict the decomposed data,the final prediction results can be gained by superimposing the prediction results of each component.The Kubernetes cluster allocates resources to the applications in a timely and accurate manner based on the predicted value,avoiding the scheduling failure caused by the sudden change in load.In the process of pod scheduling,a node selection algorithm is proposed.According to the consumption of resources by different types of pods,it selects appropriate target nodes for scheduling,thereby eliminating the bottleneck of a single resource on the node,and improving the load balancing degree of the node.Finally,the Kubernetes cluster is built and a series of experiments are performed.The experimental results show that the EMD-TCN model proposed in this paper can achieve better prediction results;The resource scheduling strategy based on load prediction can expand the pod before the peak load is reached and can reduce application request response time;The node selection algorithm proposed in this study can improve the load balancing of the node and eliminate the bottleneck of a single resource.