Kubernetes-based Resource Scheduling Optimization Method

Cloud computing is the unified management and scheduling of computing resources,storage resources,software resources,etc.through the Internet,and provides these resource services to end users on demand.Due to the characteristics of high flexibility,low resource consumption and short deployment time,container technology is widely used,but the development time of container technology is relatively short and the maturity is not high.How to manage the huge number of containers in the cluster and allocate cluster resources reasonably and efficiently under the premise of ensuring the security and stable operation of the cloud computing environment has become one of the hot issues of current research in the field.Among the many container orchestration systems,Kubernetes is widely used in the industry for its powerful scalability.However,Kubernetes ignores the consideration of resource metrics such as network bandwidth and disk,and cannot reasonably schedule resources for task requests with different resource types,while the Kubernetes native scheduling policy uses a single scoring function,which cannot meet personalized resource requirements.Therefore,the dissertation proposes a corresponding improvement scheme for the resource scheduling policy of Kubernetes,and its main work is as follows.(1)The dissertation proposes a decomposition-based ARIMA-LSTM resource forecasting model.Resource forecasting is the feature extraction of historical data to predict resource usage in future time.The dissertation treats the resource usage in the cluster as a continuous time series,decomposing it into a trend component and a periodic component for the time series characteristics,based on the respective characteristics of the decomposition terms.Resource forecasting is the analysis and processing of historical data to extract the potential characteristics of the data in the time dimension of the before and after correlation,and then model it to predict the resource load data of the resource in the future period.The ARIMA forecasting model is constructed for the period part and the LSTM forecasting model is used for the trend part,while probability estimates are made for the errors of the obtained forecasting results and the forecasting values are adjusted so as to forecast the cluster resource situation.The experimental results show that the decomposition-based ARIMA-LSTM resource prediction model proposed in the dissertationis more accurate than other single models.(2)The dissertation proposes a resource scheduling strategy based on Lyapunov optimization.Firstly,Docker technology and Kubernetes are studied,then the analysis focuses on the Kubernetes resource scheduler and its algorithmic process,and the current resource scheduling strategy of Kubernetes is optimized and improved.The dissertation takes into account the actual usage of CPU,memory,disk and network bandwidth resource metrics,pre-classifies Pod task requests by resource type and predicts the maximum number of tasks that can be scheduled for that resource type.A new evaluation function is obtained by combining the Pod’s own weight and the Node’s proprietary weight,and a node selection method based on Lyapunov optimization is proposed to enable the Pod to be deployed to the Node with the highest fitness.Through experiments,it is shown that the new scheduling policy enables efficient and fair allocation of cluster resources,and better balance of cluster resource utilization.(3)Based on the resource prediction model,the overall Kubernetes cloud platform is optimized and designed to add a monitoring and integration module,a resource prediction module and a resource scheduling module to the cloud platform.On the basis of the prediction module,it provides data reference for the resource scheduling module,and at the same time combines the resource scheduling strategy based on Lyapunov optimization to achieve dynamic scheduling of resources,so that the cluster resources can be reasonably utilized and allocated.