Desing And Implementation Of Container Management Elastic Scaling Mechanism Based On Load Forecast

With the rise of artificial intelligence and big data services,the underlying supporting role of cloud computing platforms has become more and more important.In the cloud platform,the container technology represented by Docker has been widely used due to its light weight and convenience.The reasonable allocation of resources is the key to ensuring container service quality and controlling facility costs,and memory resources are particularly important in the era of data analysis.In the face of complex and changeable cloud business requirements,how to ensure that container services have good performance while saving cluster memory resources as much as possible is a problem worthy of attention and research.Then,dynamically scaling resources according to the current memory requirements of the container service is an important method to achieve effective use of resources and avoid waste.Resource dynamic scaling is mainly divided into two directions:horizontal scaling and vertical scaling.Horizontal scaling achieves the effect of elastic scaling by adjusting the number of computing nodes.However,in some big data jobs with a long single response time,or a single instance node cannot be replicated and in the case of decomposition,it does not play a very good role.Vertical scaling is to adjust the resource quota of a single computing node to achieve the effect of elastic scaling,refine the granularity of elasticity,and optimize the resource allocation structure in depth.At present,vertical elastic scaling mainly uses threshold-based reactive scaling and fixed-size scaling units,which have problems such as hysteresis and poor accuracy.In view of the above problems,this article optimizes and improves,and finally designs and implements a container memory vertical scaling based on load prediction.The main work of the paper is as follows:(1)This paper proposes a hybrid prediction algorithm based on the combination of memory resources and service performance,oriented towards service performance levels,based on historical time series service quality monitoring data and current container resource usage,predict the memory resource demand in the next period and plan the scaled resource size,making proactive actions in advance vertical expansion and contraction to solve the problem of scaling lag.Use garbage collection time as a measure of service performance to solve the service quality detection problem of large-scale operations in the scaling process.Reinforcement learning algorithm is used to optimize the elasticity coefficient,so that the forecasting model can adapt to different elastic demands and make optimal decisions in an unfamiliar environment.(2)This paper proposes a dynamic scaling interval strategy based on threshold judgment.When the memory pressure reaches the threshold and lasts for a certain period of time,the elastic operation is executed immediately without waiting for the default elastic time.In this way,the elastic frequency can be flexibly controlled to avoid the deterioration of service quality due to sudden increase in load but untimely elastic scaling.A filling strategy for underestimating forecasts is proposed.According to the sudden situation of memory consumption in the time window,the share is dynamically filled to prevent the memory usage from exceeding the limit and affecting the quality of service.(3)This article completes the overall design of the container vertical scaling mechanism,and completes the specific implementation of the resource monitoring module,load forecasting module,and elastic scaling module.The scaling mechanism is tested and analyzed.The experimental results show that the container elastic scaling mechanism can well complete the resource monitoring function and realize the vertical elastic scaling of the memory according to the monitoring data.It can effectively reduce the waste of memory resources while ensuring that the container service quality is not affected.