Research On Inter-datacenter Application Oriented Resource Management Technology

Since the 21st century,geo-distributed public cloud platforms have become the most popular technology.Deploying inter-datacenter applications on public cloud platforms is currently the preferred solution for major Internet companies due to its low infrastructure investment,high performance,and portability.Due to the high real-time requirements,large volume of inter-datacenter traffic,and the competition between similar applications of inter-datacenter application,difficult to guarantee the real-time requirement,high transmission costs and difficult to make application deployment resource allocation strategy are still challenging problems.In this thesis,we carry out specific research work form the following aspects: application real-time requirement oriented inter-datacenter network traffic scheduling technology,low transmission cost inter-datacenter network bandwidth management technology and competition environment inter-datacenter application multi-resource management technology,and carries out specific researches.The innovation achieved are listed as follows:Firstly,this thesis proposes a fair inter-datacenter transmission mechanism with deadline guarantee,called TINA.TINA leverages a stochastic traffic scheduling method,which solves the service interference problem by considering fairness comparing with deterministic traffic scheduling algorithms.TINA consists of two parts: an access control mechanism and a deadline guaranteed priority assignment mechanism.The former is responsible for determining whether each transfer is allowed to be transmitted,and the latter is used to assign an appropriate priority to transfers that are allowed to be transmitted.Experimental results show that TINA improves fairness and deadline guarantee rate while reducing computational overhead without sacrificing network throughput,link utilization,utility and request acceptance rate.Then,this thesis proposes a low-cost information-agnostic inter-datacenter network transmission mechanism,called RIFT,Through long request dividing mechanism,a long-term transmission request into several short-term ones,so as to solve the bandwidth over-provisioning problem caused by traffic information-agnosticism and multitier pricing scheme.Through continuous short-term transmission request traffic volume estimation,and determine the most appropriate deadline and service level of each one of them,a trade-off between transmission usage cost and deadline guaranteed is achieved.Furthermore,this thesis analyzes the countermeasures cloud service providers may take and the optimality of the algorithm.The experimental results show that RIFT can reduce cloud tenant's inter-datacenter transmission cost by up to 59%.Finally,this thesis studies the competitive inter-datacenter applications multiple resource allocation method with incomplete information.Leveraging a inter-datacenter application competition model and a machine learning based resource allocation method,the resource allocation is solved.This method first decomposes the competitive interdatacenter applications multiple resource allocation problem into two sub-problems:multiple datacenters budget allocation problem and single datacenter resource provisioning problem,and then models them as a colonel game problem and a linear optimization problem respectively.Finally,the optimal solution is derived by finding the optimal solution of both subproblems.The experimental results show that the content provider with less budget should concentrate its budget on a few specific regions.In summary,the method proposed in this thesis can provide deadline guaranteed inter-datacenter transmission service for inter-datacenter applications,reduce transmission cost,and improve the economic benefits of inter-datacenter applications in a competitive environment.The method in this paper is simple,efficient,scalable,easy to be deployed,and has broad application prospects.In addition,the overall consideration of resource management of inter-datacenter network and the intra-datacenter network is insufficient,which is also a key topic for in-depth research in the future.