Performance Analysis Of IaaS Cloud Services Under Batch Task Arrivals

Cloud computing is a new computing paradigm which is developed in a fast speed. More and more people request guaranteed quality of service. Performance analysis is necessary for users to measure the quality of cloud services. Cloud data center consists of many physical machines which hold hundreds of virtual machines. We analyze the performance of cloud data centers by using scalable model due to their large scale and complexity. The foundation of the scalable model is modeling for the physical machine which is the unit in the cloud. It is an effective method to get performance measures including the number of tasks in the system and the job immediate service probability and so on by describing the physical machine which contains hundreds of virtual machine using proper queuing system. The analysis of realistic Data Trace indicated that25%Cloud tasks are batch tasks and the service time should be modeled with a general probability distribution. Therefore, it is necessary to propose a model for Clouds under batch tasks.The studies on batch arrival queuing system almost focused on Mx|G|1System which has one service machine and a small number of them paid attention to Mx|G|m System which has several service machines. Stochastic models have been proposed for evaluating the performance of cloud computing centers by assuming that a job consists of only one task. There is no accurate monolithic model of a physical machine under burst arrivals. This thesis aims to evaluate the physical machine system performance by applying analytical modeling technique.The main contributions are given as follows.1. This thesis proposes an approximate analytical approach-NMC[x] model to evaluate the performance of a pool of active physical machines on IaaS clouds under batch task arrivals, by using a Mx/G/m/m+r queue. The batch task arrivals are modeled with Poisson distribution, task service time is modeled with a general probability and the number of the virtual machine is m with the queue size is r. I first introduce the system, related assumption and the parameter definitions. Then, I present the one-step transition probabilities of the embedded semi-Markov process.2. Based on the NMC[x] model, I used an approximation technique to get performance measures. Due to NMC[x] doesn’t hold the PASTA property. So we use an approximation technique with an embedded semi-Markov process to get performance measures. Finally, I got the number of tasks in the system and the job immediate service probability. Verified by numerical results, this model not only has higher accuracy rating but also can use for a large cloud system.