Research On Resource Allocation Based On Virtual Network In Cloud Computing Environment

Cloud computing platform provides infrastructure resources to various cloudapplications through data centers, in which how to obtain the maximum resourceutilization under the premise of satisfying the quality of services for the users is achallenging problem. The unreasonable resource allocation approach would result inlow resource utilization and sometimes even difficult to meet the dynamic resourcerequirements of cloud services. Virtualization technology is an important means ofachieving the maximum resource utilization in cloud computing environment. However,due to its lack of effective management of the network resources, the main existingresource allocation approaches based on virtual machines (VMs) seem difficult to meetthe quality of services of the network-intensive applications.Virtual network introduces the concept of virtualization based on substrate network,and multiple virtual networks can be constructed on the shared substrate network aimsat increasing the resource utilization. How to maximize the resource utilization throughvirtual network technology in data center networks is an important challenge of resourceallocation problem based on virtual network. The resource allocation problem can bedivided into online and offline resource allocation problem according to whether therequests are known in advance. In the former, the requests are known in advance andwould not change, while in the latter, the requests are not known in advance and mayarrive dynamically and stay in the substrate network for an arbitrary period beforedeparting. For these two problems, academia and industry have launched a preliminaryexploration, and achieved some research results. However, the main existing heuristicapproaches have limitations in evaluating the quality of the solution, which leads to theappearance of resource fragments and inefficient resource allocation in online resourceallocation problem.Relying on the virtual network, this dissertation launches the research in view ofthese limitations, mainly includes the offline resource allocation and online resourceallocation. The main contributions of this dissertation are as follows:(1) Optimal resource allocation based on distributed constraint optimization modelExisting works lack of considering the quality of the solution in resource allocationproblem, which results in low resource utilization, then a resource allocation approachbased on distributed constraint optimization model (RADCO) is proposed for findingthe optimal mapping solution, in which the minimum of the resource allocated to theapplications is the solving objective. RADCO models the offline resource allocationproblem as the distributed constraint optimization problem (DCOP), in which therequest is modeled as the Agent, the agents that share the same substrate resources aremodeled as the constraints, and the virtual link of the request is modeled as the variables. Based on the depth-first search tree, RADCO finds the optimal solution while allowingthe agents to operate asynchronously. Theoretical proof has demonstrated that theoptimality of the solution. Finally, multiple simulation experiments have been used forevaluating the performance of it.(2) Topology awareness of virtual network for resource optimization allocationConsidering the variousness of the organization of the nodes and communicationmatrix would have effects on the resource utilization, and then the technology based ontopology awareness of virtual network is proposed for optimizing the resourceallocation. Firstly, for the generic cloud applications, a topology awareness of virtualnetwork mapping approach (TAMA) is proposed, it divides the mapping process asnode mapping stage and link mapping stage. In the node mapping stage, it considers notonly the required resource of the virtual nodes, but also the impacts that the nodemapping stage to the link mapping stage. In the link mapping stage, it adopts k-shortestpath algorithm to map the virtual links. Simulation results show that TAMA can largelyimprove the resource utilization. Secondly, for the data-intensive MapReduceapplications, a shortest path graph matching mapping approach (SPGM) is proposed.SPGM maps the virtual nodes and virtual links together, and can largely decrease themapping time of data-intensive MapReduce applications under the premise of enusringthe other performance metrics.(3) Multidimensional resource node ranking based resource allocation approachConsidering the existing approaches evaluate the resources of the nodes based onheuristic strategies would easily tend to favor some resources, and further results in theappearance of large amount of resource fragments. Thus, a multidimensional resourcenode ranking mapping approach (TK-Match) is proposed. Firstly, TK-Match ranks theresource of virtual nodes and substrate nodes according to the Top-k dominating model,aiming at balancing these different resources. The mapping process is divided into nodemapping stage and link mapping stage. In the former, it maps the nodes based on nodemapping tree (NMT), which can avoid the hops of some substrate paths are too largecaused by the complete separation of the two stages. In the latter, it adopts the k-shortestpath algorithm to map the links. Simulation results show that TK-Match can obtaingood performance in addressing the online resource allocation problem.(4) Dynamic resource allocation approach based on load sheddingConsidering the dynamic of the substrate network would lead to the pre-feasibleschemes disabled, a dynamic resource allocation approach based on load shedding(DRALS) is proposed. DRALS judges the virtual nodes and virtual links that needshedding according to load-shedding index of them, and reconfigures the sheddingvirtual nodes and links according to the greedy heuristic. Thus, when the substratechanges, DRALS can reconfigure the failed requests quickly, and effectively improvethe resource utilization and satisfy the QoS of the cloud users.