| Nowadays,microservices and container technology have developed rapidly.More and more companies choose microservices software architecture to split the system into multiple services and deploy each service in form of container on Kubernetes.In order to ensure service quality and deal with workload fluctuations,it is necessary to scale and adjust container resources in a timely manner through accurate workload prediction.In recent years,many researches on workload prediction mainly focused on the machine dimension with physical machines and virtual machines as the core,and there are few researches on container dimension.Compared with machine workload,container workload has the characteristics of rapid change,large fluctuation range,obvious workload characteristics,and potential workload correlation between containers.In view of these characteristics,this thesis deeply studies container-oriented workload prediction,scaling and scheduling technology,and designs a resource management system for the Kubernetes container platform.This thesis first analyzes and summarizes the domestic and foreign research results of time series forecasting in the field of workload forecasting.Aiming at the problem of large differences in container workload patterns,an online workload prediction strategy with adaptive workload category is proposed:perform classification based on characteristics of workload,and an attention-based LSTM AutoEncoder algorithm is proposed to train prediction models for each workload category of containers;further,build an online ensemble predictor with model weights updated in real time.In order to mine and utilize the potential correlation between containers to improve the prediction accuracy,a workload prediction algorithm based on workload correlation between containers is proposed:the workload correlation between containers is learned through GRU and self-attention mechanism,the workload sequences is modeled as a graph,and a graph neural network is used to perform the workload prediction.Then,an automatic scaling and scheduling scheme based on predictive workload and workload balancing is proposed,in which the scaling strategy based on workload correlation and workload balancing can improve service quality,and the scaling strategy based on cluster workload and instance dispersion can save resource usage.Based on the above theoretical research results,this thesis designs and implements each module of the Kubernetes resource management system,and compares it with the native Kubernetes platform.The experimental results show that the workload prediction technology and scaling scheduling strategy proposed in this thesis can effectively improve the quality of service,reduce the cost of system resources,and provide help for the system administrator to operate and maintain the cluster more conveniently. |
