Research On Key Theories And Methods Of Kubernetes Resource Prediction

Due to its flexible,efficient and fast features,containers have gradually replaced the original virtual machines and are recognized by the industry as the first choice for individual or enterprise users.Currently,more and more applications are deployed in containers.Accurately predict container resources and loads can implement resource flexible scheduling,enabling container clusters to respond in advance to the resource usage of applications deployed on the cluster,thereby enabling accurate and dynamic scheduling and allocation of resources,which not only ensures the application 's Service level agreements can also improve the resource utilization of container clouds.However,because the container load will fluctuate sharply in a small time frame,it is still a huge challenge to accurately predict the container load.In addition,there are many container indicators,and it is difficult to predict in advance which indicators are useful.In order to meet these challenges,this article focuses on the problem of resource prediction in Kubernetes from two aspects.On the one hand,in order to enable the Kubernetes cluster to respond to the resource requirements of the applications deployed on the cluster in advance,and to dynamically schedule and allocate resources according to the resource usage in the future,an ES-TCN combined prediction model was proposed.The combined forecasting model combines the advantages of the cubic exponential smoothing method and the time convolutional network to solve the problem of Kubernetes resource usage forecasting.Using the TPOT idea to optimize the parameter combination of the time convolution network model,while ensuring the performance of the model,it greatly reduces the training time of the model.At the same time,according to the characteristics of Kubernetes resource usage,a weighted allocation strategy is designed to assign different weights to the cubic exponential smoothing method and the time convolution network,thereby improving the prediction performance of the ES-TCN combined prediction model.On the other hand,in response to the problem that the load of containers in Kubernetes will fluctuate sharply in a short period of time and there are many container indicators,and it is difficult to predict which indicators are useful in advance,a Kubernetes container resource load prediction model is designed.The model mainly includes two modules: index selection module and load prediction module.First use Lasso regression method to select key indicators,and then combine the advantages of gray model and long-term and short-term memory network to propose a new mixed prediction model(GM-LSTM).GM-LSTM mixed prediction model first builds a gray prediction model.For each selected The key indicators are predicted,and then the predicted values of the key indicators obtained through the gray prediction model are used as the input of the long and short-term memory network model to obtain the final load prediction value.Finally,we use the data sets of five services Apache,Mysql,Nginx,Redis and MongoDB collected from the real online Kubernetes cluster to evaluate the proposed prediction model and the existing prediction model.The experimental results confirm that the ES-TCN combined prediction model has better prediction effect on the Kubernetes resource usage(CPU and memory)on the five data sets than the two single models;the GM-LSTM hybrid prediction model can reduce the training index Maintain the prediction accuracy,and the prediction effect of Kubernetes container resource load on the five data sets is better than the existing ARIMA model,LSTM model,and GM-SVR hybrid prediction model in most cases.