Research On Performance Interference Prediction Model And Schedule Of Virtual Machine

As a new computing model, the cloud computing provides configurable and shared infrastructure to the users and change the ways to access service, the means to the spending. Virtualization technology is the key technology of could computing, because of its characters of consolidation, migration, isolation and other advantages. It plays a key role to the performance improvement and resource optimization of cloud data centers and brings fresh methods to energy management. However, when we use virtualization technology to distribute physical machine resource into multiple virtual machines (VMs), it will cause performance interference among the VMs which co-located in a same physical machine contending for the shared physical resources. Therefore, predicting the performance interference of VM in real-time is the point to meet the quality of service. The performance of VM and the energy efficiency of data center can not achieve the optimal simultaneously, because if improving the resource utilization to reduce the energy of data centers, high resource utilization will reduce the performance of VM because of resource contending. How to schedule VMs to the physical machines that let the interference smaller and energy efficiency better is the hot issue of data center.In this paper, we studied the issues of performance interference predicting and VM scheduling which has the target of improve energy efficiency and reduce VMs interference. Combined with previous researches, the works of this paper are follows:(1) This thesis has made experiments to simulate the performance interference among VMs co-located on one machine, by which we study VM real-time characteristics and collect runtime performance parameters under interference, so as to model the performance interference of VMs.(2) The values collected between different resource-intensive applications are significantly different. Thus the performance interference predicting model which is built by all kinds of resource-intensive application has bad predicting result. In this thesis, we firstly build application classified model, and classify the applications into CPU-intensive, I/O-intensive, Memory-intensive and Mixed. Then we build the models for each kind of resource intensive applications. Finally, comparing two models, we find the classified model achieves a better predicting result than the unclassified.(3) About the problem of interference and energy efficiency cannot achieve optimal at the same time, this paper established a comprehensive weight function of VM interference and physical machine energy efficiency, and designed a performance-energy-balancing VM scheduling policy. Intended to allocate VM to the physical machine which have best energy efficiency and have little interference. By adjusting the parameters of the comprehensive weight function, we can also implement the performance priority or energy efficiency priority scheduling. Finally, compared with other scheduling, this one achieved better results.