Research On Resource Scheduling Strategy Based On Kubernetes Cluster

With the popularity of cloud-native related technologies,Kubernetes,as an open source platform and excellent performance container orchestration technology,is widely used in many cloud platforms.However,there are still some shortcomings: due to the lack of prediction capability of resource usage,the default horizontal scaling policy has delayed response when dealing with unexpected traffic,which leads to service response timeout and poor anti-shake capability;the default scheduling policy does not consider the resource utilization rate of each node and the total cluster resources,making it difficult for the cluster to maintain long-term load balancing and stability.First,to address the lack of resource prediction capability of container cloud clusters,we propose a prediction model combining the Variational Modal Decomposition based on Whale optimization(WVMD),Bidirectional Convolutional Gated Recurrent Unit(BCGRU)and Attention Mechanism(ATT).The model can better learn the container cloud resource load characteristics which are highly nonlinear and unstable,thus improving the prediction accuracy.Second,to address the shortcomings of the default horizontal scaling strategy,we propose the Pod Predict Based Horizon Pod Autoscaling(PPB-HPA)by using a combined model to predict Pod resource usage.By predicting the resource usage of Pods in advance to increase the judgment index for horizontal scaling,and then using the two-stage judgment method to accurately determine the scaling timing,thus avoiding waste of resource and reducing operation and maintenance costs.Then,to address the shortcomings of the default scheduling strategy,we propose the Server Predict Based Load Balance Priority(SPBLBP)by using a combinatorial model to predict node resource usage.The total cluster resources and node resource usage as well as the future resource utilization are considered comprehensively,so that the scheduling link has the ability to sense ahead and thus improve the cluster load balancing degree.Finally,experiments are designed to validate the WVM-BCGRU-ATT prediction model,the PPB-HPA scaling strategy and the SPBLBP scheduling strategy.The experimental results show that on the Ali cloud Cluster-trace-v2018 public dataset,the root mean square error of the WVM-BCGRU-ATT prediction model decreases by 20%,14%,7% and 3%,respectively,compared with the GRU,CNN-Bi LSTM,GRU-LSTM and WVMD-BCGRU models,which have higher prediction accuracy;in the Kubernetes experimental environment,the PPB-HPA scaling strategy has better pre-scaling performance and better anti-shake ability,while the SPBLBP scheduling strategy can The PPB-HPA scaling strategy has better pre-scaling performance and anti-shake ability than other models,while the SPBLBP scheduling strategy can improve the overall balanced efficiency of cluster resource usage.