Virtualized Resource Management In Data Center

With the development of big data and cloud computing technology,data center has become an indispensable infrastructure for making our daily life convenient.Since the data size is large and the cost of computing is amazing in biomedical field,intelli-gent transportation,personalized recommendation and so on,these computing depends on clusters in data center with powerful computing and mass storage resources.Im-proving the efficiency of data center not only provides better support for applications,but also saves huge costs for service providers.Therefore,how to improve the resource utilization for existing data center becomes an increasingly important research topic.In recent years,with the development of virtualization technology,the research work of data center has entered a new stage.The data center is no longer an infras-tructure that just put servers and other related components together,but separating the upper layer logical resources from the underlying infrastructure,which provides vir-tualized resources for tenants as cloud service.In this paper,we discuss the resource management problems separately for public and private cloud in modern data center.For shared virtual machines in public cloud,existing deployment algorithm for tenant request has not considered the software middleboxes in multi-tenant data cen-ter when provide bandwidth guarantee for tenants.In this paper,we address a virtual machine deployment algorithm called MISSILE for tenant requests with software mid-dleboxes.Our deployment algorithm is proposed to offer predictable network perfor-mance for each accepted tenant,and maximize the virtual machines' utilization in data center.For big data processing framework deployed in private cloud,existing works fo-cus on designing scheduling algorithms in cluster scheduler.However,the improp-er request from the upper-layer framework may adversely affect the performance of scheduling algorithms in cluster scheduler.In fact,users are always with no idea of their jobs'resource demand and the requirement may change in different stages.Although there are dynamic resource allocation strategies for some data processing framework,these strategies may come with additional overhead.In this paper,we design a dynamic resource sharing strategy based on the existing dynamic resource al-location strategy.In our experiment,we show that our work can effectively reduce the average job completion time for shared cluster.