The Research Of Resource Allocation Strategies Oriented Cloud Testing Service

Software testing is facing many new challenges in a cloud computing environment. The thesis found that it needed reliable resource allocation strategies as a guarantee in cloud testing service after doing some cloud testing service surveys. Currently, the study of cloud resource allocation strategies is with some theoretical researches and practical operations, but it is lack of the overall distribution architecture which stands for the view of test users and cloud service providers demands.Based on the referential cloud service models, the research designed a cloud testing service model. The model was divided into two parts. The resource dispenser 1 provided resource availability allocation strategy for users. The resource dispenser 2 provided resource efficiency allocation strategy for cloud service providers on condition that testing virtual machines met the availability of resource allocation. Then the thesis used a predictive allocation method for resource allocation strategies and implemented algorithms for the two components.The resource dispenser 1 predicted the utilization rate of CPU and available size of memory by BP neural network, so as to provide a strategy to allocate virtual machine resources. In prediction process, in order to improve the accuracy of forecasting resources, the thesis introduced an improved particle swarm intelligence algorithm to initialize weights and thresholds of BP neural network. At last, comparison experiments have proved the effectiveness of the algorithm. On the foundation of multiple testing virtual machines allocated by resource dispenser 1 meeting the availability of resource allocation, the resource dispenser 2 using generic algorithm minimized the cloud servers' memory and completed the settlement between testing virtual machines and cloud servers. It was another strategy to provide a virtual machine resource allocation. In allocation process, the traditional genetic algorithm is easy to obtain infeasible values. To solve this problem, combining with one-point crossover repair, rotation variation and external penalty function theory, this thesis introduced an improved genetic algorithm. Lastly, the experiments have proved the feasibility of the improved algorithm's optimization capability.