Research On Container Elastic Scaling Strategy Based On Load Forecasting And Reinforcement Learning

Elastic scaling is an important feature of container cloud platforms that aims to improve the ability of applications to cope with dynamic changes in load.As traditional reactive scaling strategies suffer from elastic lag as well as configuration complexity,smarter proactive elastic scaling has attracted the attention of many scholars.By proactively predicting future resource demands and making preemptive arrangements for possible scenarios,proactive elastic scaling strategies can effectively avoid Service Level Agreement(SLA)violations and improve resource utilization.However,accurately predicting future loads and making reasonable elastic scaling actions are difficult problems in the task of proactive elastic scaling.To address this problem,this thesis first conducts a load prediction model study,based on which,a hybrid elastic scaling strategy based on load prediction and reinforcement learning is proposed.The main research contents and innovation points are as follows:(1)For the problem that it is difficult to predict accurately due to the diverse patterns of temporal changes in cloud load sequences,a load prediction model based on multi-scale dilated convolution and LSTM is proposed.The model uses parallel and stacked dilated convolution residual networks to form multi-scale feature extraction blocks to extract timing features from continuous and discontinuous input time steps,thus learning short-term load fluctuations,persistent changes and periodic information simultaneously,and then the LSTM further learns long-term timing dependencies from the extracted features.Load prediction experiments are conducted on several cloud load datasets,and the obtained results show that the prediction accuracy of the model proposed in this thesis is better than comparative models.(2)For the problem that it is difficult to make reasonable elastic scaling actions due to the complexity of elastic scaling in cloud environments,a hybrid elastic scaling strategy based on load prediction and reinforcement learning is proposed.The proactive scaling module of this strategy makes full use of the current load state and prediction information of the application,and the reinforcement learning agent make more reasonable proactive scaling decisions for the future.Meanwhile,in order to cope with extreme traffic peak scenarios,a reactive scaling module combined with SLA is designed to be able to respond and adjust in time according to real-time quality of service.Finally,in order to coordinate the above two scaling strategies and combine the advantages of both,a hybrid scaling controller is designed to balance proactive scaling decisions and effective scaling under extreme peak traffic.(3)In order to verify the effectiveness of the proposed elastic scaling strategy,an elastic scaling system is implemented on the Kubernetes platform.The monitoring module of the system is based on the Prometheus solution,which can comprehensively collect the scaling metrics data and store them persistently;the reinforcement learning module is based on the Open AI Gym tool,which can conveniently abstract the Kubernetes environment and build the reinforcement learning model;the scaling control module is based on a custom controller,which can effectively extend the scaling strategies control logic.In order to simulate different load scenarios,two kinds of request loads with different characteristics are selected for elastic scaling comparison experiments.The experimental results show that the scaling strategies proposed in this thesis can reduce SLA violations and improve resource utilization more effectively than the existing methods.