Resource Cost Oriented Decision Making Method And Its Application For Cloud Application Performance Optimization

Widespread popularization of Internet and emergence of web computing,mobility computing and diversity of computing have taken lots of challenges to software systems,such as the increasing cost on management and maintenance and more flexibility to handle the changes of the situated environments,etc.The adaptive methods of service system under the cloud environment have become an important research direction in the field of service computing and cloud applications.The utility computing characteristic and pay-as-you-go mode of cloud computing request that service system not only meets the demand of applications for the allocation of resources in the circumstance of keeping the minimum cost of resources,but also dynamically adjusts resources adaptively when it deviates from the expected behaviors so as to continue to provide the services in line with the user expectations.However,there are a lot of deficiencies in the existing adaptive methods,such as the lack of taking into account the overall cost of both the users and service providers from the system point of view,etc.Therefore,there is an urgent requirment to research the decision making problem for cloud application performance optimization systematacially from the view of overall efficiency optimization at present,which takes into account both the user benefits and the cost of cloud service providers,then practically promote the cloud service adaptive system to apply widely and efficiently.To solve the decision making problem for cloud application performance optimization,this thesis puts forward a decision making method of improved genetic algorithm based on the running self-optimize implementation framework of cloud service system which is designed and implemented by the research group.Firstly,to establish two critical factors of the resource cost oriented decision making model for cloud application performance optimization,including resource and quality of component service relationship model and cloud application load prediction model.The thesis puts forward a dynamic service performance modeling method named CSDM that combines collaborative filtering recommendation and support vector regression,which is suitable for nonlinear relationship model,and uses the method to establish resource and quality of component service relationship model,then uses the model to forecast respond time,throughput and reliability of component service under a certain resource state;The thesis uses the deep belief network method to establish load prediction model named DLM,then uses the model to forecast the next cycle load of cloud application.Secondly,to research resource cost oriented decision making method for cloud application performance optimization,the thesis defines optimization objective function and constraint conditions in the decision making problem for cloud application performance optimization,then converts the problem to the corresponding mathematical model,that is to say,to establish resource cost oriented decision making model for cloud application performance optimization;the thesis compares the effect of solving the the decision making problem for cloud application performance optimization with the standard genetic algorithm,the differential search algorithm and the improved genetic algorithm through solving the decision making model using decision algorithm,then puts forward a decision making method for cloud application performance optimization based on improved genetic algorithm.Finally,to introduce the application of resource cost oriented decision making method for cloud application performance optimization.To apply the research result of this thesis to the adaptive optimization experiment platform,then to run a service-based software system called scenery spot voice tour guide cloud service system,and carries on the contrast experiment with what doesn't deploy the research result of this thesis to the adaptive system based on the adaptive system accumulation logs.Test results show that for the load type of periodic variation of load and little variation of adjacent load,the cost performance of the adaptive system with applying the research result of this thesis is higher than the cost performance of the adaptive system without applying the research result of this thesis.