Research On Kubernetes Scheduling Strategy Based On Workload Similarity

In recent years,with the continuous expansion of the cloud computing industry and the update and iteration of virtualization technology,container technology has gradually occupied a place in the infrastructure construction of cloud platforms.Compared with traditional virtual machines,container technology is more lightweight and flexible,and greatly reduces the performance loss during virtualization.As the de facto standard of container orchestration system,Kubernetes’ scheduling algorithm has always been the research focus of scholars in related fields.In the current research on the Kubernetes scheduling algorithm,an important input parameter during scheduling is the amount of resources requested by the user for the container,which has an important impact on the decision-making of the scheduling algorithm.However,in the production environment,a common phenomenon is that there is a big difference between the resources requested by the container and the resources actually used.At this time,the scheduling algorithm often cannot obtain the scheduling effect consistent with its optimization goal.In order to solve the problem of the discrepancy between the resources requested and the resources actually used during container scheduling,this paper proposes a container load prediction method based on time series similarity.This method combines the time series clustering algorithm and the classification algorithm to find the containers in the cluster that may be similar to the load of the container to be scheduled,and uses the time series prediction algorithm to take the load of these existing containers as input to obtain the load of the container to be scheduled Estimated load,so as to give the scheduler a reference value close to the actual resource usage.In order to solve the problem that the models of time series clustering and time series prediction cannot be reused,this paper proposes a composite time series feature extraction model,which combines the advantages of the two models of autoencoder and BiLSTM,so that the model can be used for time series prediction at the same time,The extracted intermediate features can also be used for time series clustering,saving model training time and computing resources.In addition,this paper also built an experimental Kubernetes cluster,designed and implemented a custom scheduler based on predicted load,and optimized node load balance and Pod stability.In the load prediction experiment based on the public container load data set,the average absolute percentage error between the actual container load and the predicted load is 5.97%,which is much smaller than the difference between the container application load,which verifies the feasibility of the method.In the scheduling experiment based on the selfbuilt Kubernetes cluster,compared with the default scheduler,the scheduler designed in this paper reduces the CPU resource balance of the cluster from 0.5 to 0.23,and the memory resource balance drops from 0.4 to 0.25,which significantly improves the load balancing and operation stability of the cluster.