Research On Incremental Learning Based Resource Management Technology For Data Center

Low resource utilization is the urgent problem that the data center needs to solve at present.On the one hand,tenants apply for resources according to the peak demand when deploying applications in the data center.The resources they used are much smaller than the resources requested,resulting in a waste of resources.On the other hand,the utilization of different resources within a node is not balanced.Data center resource management improves the physical resource utilization and reduces energy consumption,becoming research hotspot of the cloud computing field.The dynamic nature of the cloud environment poses a severe challenge to the resource management: First,the dynamic change of the resource demand of the application requires resource management to predict the resource demand accurately to realize the elastic expansion.Load changes will make the node overload or imbalance of different resources,which requires resource management technology achieve the accurate adjustment.At present,most of the related research on resource management neglects the mutual influence of different resources and lacks incremental updating of model,which cannot cope with the dynamic change of cloud environment.To this end,focusing on the goal of improving the resources utilization,this dissertation deeply studies the resource demand prediction technology and resource dynamic adjustment technology.In large-scale multi-tenant cloud environment,tenants apply for resources according to the peak demand when deploying applications in the data center,resulting in a waste of resources.Resource demand prediction is an indispensable step in resource allocation and elastic expansion in cloud environment.To this end,this paper proposes a TrendMatching Resources Coupled Prediction method TMRCP.First,to cope with the diversity of the cloud environment,we propose a Resources Utilization Trend Matching algorithm(RUTM),which defines a new similarity measure for multi-dimensional sequences and takes the correlation among resources into consideration.Second,we propose a dynamic prediction window adjustment algorithm that selects appropriate prediction length for different resource utilization trends to overcome the disadvantage of fixed window.Third,in response to the sudden changes,we put forward a mixed synthesis algorithm to improve the robustness of the method.Experiments on Google's cluster usage trace show that the Mean Absolute Percentage Error of TMRCP is 4.7%,20% better than the state-of-the-art.In addition,the TMRCP is still accurate in multi-step-ahead prediction.In the data center,resources overloaded results a bad performance,and the utilization of different resources within a node is not balanced.To this end,this paper proposes an Overhead-Aware Dynamic Adjustment Method for Multiple Resources ODAMR,to achieve resource balance between multiple nodes.First,we propose the definitions of the ideal utilization and the state of the physical nodes,and calculate the urgency of the overloaded physical nodes and the imbalance in the use of the normal nodes.Second,we calculate the improvement of node resource allocation for each possible adjustment scheme.Finally,we design a migration scoring algorithm to solve the two-objective combination optimization problem,weighing the effect and cost of the migration.Experiments based on the CloudSim platform show that ODAMR can effectively match applications and allocate resources to ensure system load balancing.Compared to existing methods,service level agreement(SLA)violation of the ODAMR declines by 10%.Based on above theoretical research results,an Incremental Learning Based Resource Management System for Data Center is designed and implemented on Storm.We develop resources demand prediction module and dynamic adjustment module.The resource demand prediction module predicts the resource requirement of the application.The resource dynamic adjustment module allocates resources according to the resources requirement.Experiments show that ILRM can effectively manage resources.Compared to existing methods,resource utilization increases by 15%.