Performance Analysis And Prediction Of Large-scale Cloud Service Platforms

A large-scale cloud service platform can support business operation, innovation and evolution for modern service industries. And performance analysis is one of hot research topics in large-scale cloud service platforms. However, the traditional methods of perfor-mance analysis can not overcome challenges of the scale, architecture and new techniques for cloud computing. In this paper, we focus on the research of bursty workload generation, resource prediction under virtualized environment and performance models of parallel pro-cessing.Firstly, a bursty workload generation method named MBWG for cloud benchmark Cloudstone has been proposed based on2-state Markovian arrival processes. Using MBWG, performance analysts can construct MAP2model easily and generate bursty workloads with the Faban framework. The relationship between the parameters of MAP2and the intension of burstiness has been studied. By comparing the actual value of index of dispersion for counts estimated from system logs with the target value deduced from MAP2, we show that our method is more accurate than related work. Besides, the performance of Cloud-stone under bursty workloads has been studied in a Xen-based virtualized environment.Secondly, a fine-grained resource utilization prediction method has been suggested for large-scale cloud service platforms. The resource cost by virtualization technique has been considered during the prediction of pmf function of resource utilization. By analyzing the relationship between workload and scheduling policy of Xen VMM, we build a model to estimated the resource consumed by VMM. And we combined the queueing network and statistical learning methods to build a unified resource utilization model for non-saturated, semi-saturated and full-saturated workload. Then the model for bursty workload also be derived. The experiment results show the accuracy of our prediction models and explain the motivation of pmf function prediction.Thirdly, an efficient method named horizontal decomposition has been developed for Fork-Join queueing network models, which is used to analyze the performance of parallel processing systems in large-scale cloud service platforms. The main idea of our method is to approximate a non-product-form FJQN with some closed and open product-form networks. So the computational complexity can be dramatically reduced compared with the hierar-chical decomposition approach. And the algorithms for solving single-class and multi-class closed FJQNs have been developed respectively based on the horizontal decomposition.The evaluation results show that90percentile of relative errors of most service centers are less than15%except for the shared ones.Finally, we apply these performance models and prediction methods in a cloud service integration platform called JTangCSB. The automatic component installation method for JTangCSB can consolidate the components based on the results of the resource utilization prediction and correlation coefficient calculation, which can optimize the performance by minimizing the resource contention between components. And the intelligent container de-ployment method will model all possible container deployment topologies with FJQN. By resolving the FJQN models, JTangCSB can find the best topology that will avoid unneces-sary data transmission through Internet and reduce the total response time. We also use a real scenario to show the effectiveness of these two methods.