Research On Performance Management For Cloud Computing Platform

As cloud computing technology is increasingly applied to various fields of informationindustry, it becomes more and more important to monitor and manage the various heterogeneousresources in the cloud environment.Compared with other large-scale distributed networks, cloudcomputing is more complex with the characteristics of virtual, hierarchical and dynamic. This thesisfocuses on the performance management under cloud computing environment, and the maincontributions are described as follows:(1)The basic concept, characteristics and development of cloud computing are summarized,and then the existing cloud environment performance management techniques are summarized.(2)Aiming at the problem of low efficiency of communication and load imbalance ofmonitoring servers in traditional monitoring model, a dynamic hierarchical monitoring model isproposed, which sets up the mathematical model for dynamic partition in monitoring region, andensures the monitoring real-time and load balancing of monitoring centers as well as reduces thecommunication cost as far as possible. In addition, a self-adaptive genetic algorithm is proposed tosolve the mathematical model. The algorithm uses binary coding to represent the region partitionresults, designs the fitness function on the basis of the monitoring target, and dynamically adjuststhe probability of cross and mutation according to the fitness and evolution algebra. Finally, theperformance of proposed model is evaluated from the monitoring time, network traffic and the loadbalancing. Simulation results show that the monitoring model can reduce the monitoring traffic,shorten monitoring time and balance the load of the region centers comparing with the staticmonitoring model and the greedy monitoring model.(3)Since traditional prediction model cannot adapt to the complex changes in the performanceof the cloud platform and its prediction accuracy rate is not high, a resource performance predictionmodel is proposed, which is based on PSO-Elman. It uses Elman neural network to predict theperformance and determine the number of nodes in the network input layer according to thecorrelation of the sample data. Then, it uses PSO algorithm to train the Elman neural network.Meanwhile the concept of the degree of aggregation in PSO algorithm is introduced to adjust theparticle swarm diversity. In addition, a performance forecast model based on MapReduce aiming atthe cloud environment is proposed. Finally, this thesis simulates the model from the followingaspects: training speed and optimization ability of the neural network learning algorithm and the accuracy of prediction of the prediction model. The results show that the prediction model proposedin this paper can not only keep better accuracy in both the short-term and the long-term forecasts,but also improve the training speed of the neural network.