Research Of Scheduling Policies In Cloud Computing

The continuous development of cloud computing results in great diversification of user requirements and huge complexity of applications, which puts forward higher requests for resource allocation, load balancing and scheduling management. The traditional scheduling algorithms in gird computing environments mainly focuses on the improvement of execution efficiency, while for cloud computing environments, the commercialization determines that more factors such as resource and power consumption are needed to develop models and do research on scheduling policies. Moreover, it is necessary to set up a virtualization management platform with graphical user interface for efficient administration of cloud computing environments.In cloud computing environments, demands from different users and often handled on virtual machines (VMs) which are deployed over plenty of hosts. Huge amount of electrical power is consumed by these hosts and auxiliary infrastructures that support them. However, demands are usually time-variant and of some seasonal pattern. It is possible to reduce power consumption by forecasting varying demands periodically and allocating VMs accordingly. In this paper, we propose a power-saving approach based on demand forecast for allocation of VMs. First of all, we forecast demands of next period with Holt-Winters'exponential smoothing method. Second, a modified knapsack algorithm is used to find the appropriate allocation between VMs and hosts. Third, a self-optimizing module updates the values of parameters in Holt-Winters'model and determines the reasonable forecast frequency. We carry out a set of experiments whose results indicate that our approach can reduce the frequency of switching on/off hosts. In comparison with other approaches, this method leads to considerable power saving for cloud computing environments.To meet conditions in application environment such as limited resources, random requests, etc., we reconstruct model by adding limitations and propose a VM allocation algorithm based on fixed job scheduling model. It satisfies the demand of FCFS and minimizes the amount of unallocated VMs. We add time dimension to describe start time, end time and duration of requests. At last, we illustrate our approach's feasibility and study the reasonable trigger mechanism by simulation experiments.In the end, we implement a prototype system for virtualization management on the basis of scheduling models into which we embed above two VM allocation algorithms. The implementation details of system architecture, modules, use cases, etc. is given by different UML diagrams.