Research On Energy And Cost-aware Resource Scheduling In Cloud Data Centers

In recent years, with the rapid development of cloud computing, mobile Internet and big data, it has been in a high-speed construction period for the next-generation cloud data centers around the world. Cloud data center which is the most important carrier of Internet services has to scale up dramatically to catch the growing resource demand of Internet business. However, a large amount of IT infrastructure leads to a serious problem of high energy consumption and operating costs.In this paper, we focus on developing energy and cost-aware scheduling algorithms in cloud data center. We consider the service types, system architecture and multi-regional deployment of cloud data centers. We exploit the characteristics of all the aspects above and try to take advantage of them in our scheduling. The algorithms solve three different problems in cloud data centers:the first is how to reduce the energy consumption of running parallel tasks; the second is how to minimize the number of active physical machines while holding a large amount of virtual machines; the third is how to reduce the energy cost of geographic data center. The main contributions of this thesis are as follows:(1) To schedule parallel tasks with minimum energy consumption, we give a formulation for the problem of parallel task scheduling in cloud data center based on dynamic voltage and frequency scaling (DVFS) technique. By exploiting the characteristics of the parallel tasks and the DVFS technique, we develop a mechanism called Partial Optimal Scaling (POS) and give a theoretical proof for that the mechanism can achieve the optimal energy consumption for parallel task scheduling. Then, we propose a new energy-aware parallel task scheduling algorithm based on the POS mechanism. The experimental results show that our algorithm can substantially reduce the energy consumption while guaranteeing the performance of parallel tasks.(2) Due to the low resource utilization and a great number of idle servers, a large amount of energy is wasted in data center. For this problem, we propose a system framework for virtual machine consolidation to improve the server resource utilization and minimize the number of active servers. In the system, we develop a resource demand prediction mechanism and a scheduling algorithm based on resource prediction. The experimental results show that our prediction mechanism can accurately predict the demand resource and that our algorithm can substantially improve the resource utilization and reduce the number of active servers.(3) To solve the problem of high energy costs in data centers, we develop workload scheduling models and energy cost models based on the location-varying and time-varying diversities of electricity price in multi-regional data centers. We also formulate the schedule problem into Integer Linear Programing problem and propose an energy-cost-aware scheduling algorithm. The experimental results show that our algorithm can substantially reduce the energy cost of data centers and achieve a near-optimal result.