云计算中基于负载预测的高效能资源配置和管理策略研究

The need for higher computing capability promotes the rapid development of large-scale data center, which causes serious energy consumption problems. Hence, a further study on energy saving strategies of data center is of great significance. The project aims at solving the problems of low resource utilization and high energy consumption of cloud platforms caused by the varying workload. In order to explore the problems, the project is trying to integrate the theories of multi-subject such as queuing theory, operational research, neural network, and optimization theory, and research energy saving strategies in three aspects: platform configuration, server management, and VM migration. Considering the gap between the workloads in peak hours and off-peak hours, try to configure cloud platforms combining the fixed resources and IaaS resource renting. Adopt the M/M/c+D queuing model with impatient waiting to model the cloud multi-server system, build the energy optimization model and solve it combining the partial derivatives and gird bisection. Aiming at the service latency caused by server sleeping mechanism, this project tries to design a forward looking resources management strategy. A workload predictive model is tried to elaborated based on BP neural network. By predicting the changing trend of workload, the resources can be reconfigured in prior. Besides, VM migration scheme is predesigned to achieve the objectives of reducing energy consumption and improving resource utilization without degrading the quality of service of cloud computing. This project has the potential to improve the high energy consumption problem of private cloud, and can promote the further development of cloud computing.

对计算能力的需求，促进了大规模数据中心的飞速发展，同时也造成了严重的能耗危机。深入研究云数据中心的节能问题具有非常重要的意义。本项目申请针对云计算负载动态性造成的资源低利用率和高能耗的问题，从云平台配置、资源管理、虚拟机迁移三个层次研究云计算的节能策略。针对云服务负载高峰和低谷时段的差异性，拟采用固定配置+资源动态租用的模式对云平台进行配置。采用M/M/c+D不耐烦顾客排队模型对多服务器系统进行建模，建立能耗最优化模型，并结合偏导极值法和网格二分法求解最优值；针对已有服务器休眠机制导致的服务延迟问题，拟建立系统负载的BP神经网预测模型，提前预测负载的变化趋势，计算满足未来负载需求且能耗最优的资源配置方案，并对服务器进行休眠或者预唤醒。同时针对预测的系统状态对虚机进行动态迁移以及重调度，达到降低能耗和提高资源利用率的目的。本项目有潜力改善企业私有云的高能耗问题，促进云计算的进一步发展。

项目执行三年来，紧跟云计算的最新动向和发展趋势，在云计算资源管理及利润优化方面进行了系统深入的研究工作。从云平台的资源部署、资源租用与定价、虚机管理及任务调度、应用数据查询优化等几个方面展开研究。.1、从云服务供应商利益角度出发，研究在云服务商在资源约束下合理选择资源配置类型，以及资源配置规模问题。采用多服务器排队模型，结合边际效益递减原则，建立资金约束下利润最优化模型。.2、用云用户角度出发，综合考虑云计算的特征，引入云经纪人角色，并针对云服务商的资源特性及云用户的任务类型，利用排队论建模，提出平台自适应的配置及定价方案。.3、从计算机任务调度角度出发，设计满足资源约束和任务响应时间约束的双目标调度算法，提高云平台虚拟集群性能，最大化服务商利益。实验表明，提出的算法在makespan上性能更好，并且在平均响应时间方面也具有良好的性能。.4、在云计算大数据处理中，针对数据的不确定性，首先提出了一种基于强支配的改进p-skyline（MPS）查询算法。然后基于一个新提出的排序准则，构造了一个top k MPS（TkMPS）查询，并采用重用技术来提高查询性能。提出的算法可大大降低计算查询的开销。.本项目从云服务各个层面提出解决方案，实现云服务利益最大化。为云计算发展及运营提供管理及技术支撑。