Research On Optimal Resource Allocation Approach Based On Service Selection For Cloud Application

Service-Based Software system(SBS),due to its advantage of loose coupling and dynamic reconfiguration,has become an important form for rapidly building large-scale and distributed applications in Internet.With the booming using of cloud computing techniques,more and more service providers have deployed their SBSs on cloud platforms.However,because of the pay-as-you-go model,there still exists an optimal resource allocation problem for deploying such applications,i.e.,how to determine the best resource configuration for each component service,called application's Resource Allocation Strategy(RAS)for short,thus meeting users' demand on Quality of Service(QoS)while minimizing the overall resource cost.Most of the current researches on optimal resource allocation for cloud application fall into two categories:approaches that take the application as a whole entity to determine its optimal resource configuration,and those that take the component services in an application as independent entities to locally optimize their resource configurations.However,for SBS based cloud application,these two types of approaches have the following shortcomings.First,QoS aggregate relationships between component services and application make the formers' resource configurations have direct influence on the latters' global optimization goals.Therefore,the SBS cannot be allocated resources as a whole.Also,the local optimization cannot ensure the global optimality of SBS's RAS.Second,since the approximately continuous cloud resources bring in a huge search space,it is unpractical to find the optimal RAS meeting global constraints.Third,the current approaches ignore the stateful component services whose constraints add the complexity of resource allocation.Forth,dynamic cloud environment often makes the RAS found at the initial phase not keep the global optimality at runtime,or even leads to the failure in achieving optimization purpose under new resource status by choosing a single resource configuration for each component.To solve the above issues,an approach based on service selection to optimize resource allocation for SBS in cloud is proposed in this thesis.This approach first divides the continuous cloud resources into relatively limited discrete resource configurations,which makes it is possible to find the optimal RAS in acceptable time cost.Then,each resource configuration,together with the corresponding component service's QoS values and resource cost,is encapsulated into a logic service so as to transform the resource allocation for SBS based cloud application into a logic service selection problem.Finally,aiming at different types of problems,the corresponding service selection algorithms are proposed to find the optimal RAS.The main distributions of this thesis are as follows.(1)A framework of resource allocation based on service selection for SBS based cloud application is proposed to achieve the global optimization goals.For different SBSs,goals and phrases,the fundamental resource allocation process is summarized as two steps:logic service generation and selection.Besides improving the efficiency of service selection,the logic service generation avoids the influence that component services' resource-performance models have on the structure of selection algorithms,thus making the framework a good commonality.(2)The ways of dividing continuous cloud resources are studied,and according to whether depending on execution history of component services,two resource partition strategies,respectively named equal-width partition and entropy based minimum description length partition,are proposed.The two strategies complement each other in terms of the quality of RAS and the time cost of logic service generation algorithms.(3)Aiming at the initial optimal resource allocation problem for SBS cloud application with no relation state,a formalized model based on service selection is built.During solving this NP-hard model,a hybrid genetic algorithm based on elitism and local search strategy is designed to improve convergence of searching solutions.The experiments under several typical cloud scenarios prove the effectiveness of proposed resource allocation approach and the efficiency of its solving algorithm.(4)Aiming at the initial optimal resource allocation problem for SBS cloud application with relation states,a formalized model is built based on service selection by introducing the notion of stateful component service group,and then solved using a discrete differential evolution algorithm.Using status-labeled vector as well as a new crossover strategy to cope with binding constraints of logic service,the algorithm avoids the increase of individual coding complexity.Besides,the improved selection operator can also ensure the solution's feasibility while speeding up the convergence.(5)Aiming at the initial optimal resource allocation problem for multi-SLA(Service Level Agreement)SBS cloud application,a multi-objective genetic algorithm is designed to solve the service selection based resource allocation model.To cope with the conflicts among objectives,such as resource costs,end-to-end response time and reliability under different SLAs,the algorithm is specially used to find the Pareto optimal logic service compositions.With the proposed fitness functions of feasible and infeasible individuals,solutions more equally distributed over the QoS objective space can be found,which contributes to choosing the satisfactory solution as the optimal RAS according to application providers' preferences.(6)To deal with the changing runtime environment,an adaptive optimization mechanism following the "Monitor-Analyze-Plan-Execute" model is designed to realize the dynamic resource allocation for SBS cloud application.The key is a service selection based adaptation policy model,where more than one logic services are chosen for each component service in the RAS,and switched periodically at runtime according to an optimal probability distribution,thus making it possible to achieve better QoS objectives as well as improve the availability of application instance in case of environmental degradation.