Research And Implementation On Energy-aware Load Balancing Strategies In Data Centers

With the continuous development of applications such as Cloud Computing and Big Data,as the supporting infrastructure,data centers are rapidly expanding in terms of number and scale,and the energy consumption problems brought about are attracting more and more attentions.The over-provisioning of server resources is one of the common problems in today's data centers.Redundant servers not only waste energy and increase operating costs,but also reduce resource utilizations.How to dynamically adjust the scale of the servers according to workload changes to reach energy proportionality of the data center under the premise of maintaining the high availability of services is a hot issue researched today.To solve the above issue,energy-aware load balancing algorithms are proposed to save energy consumption of data centers,with high availability requirements of services taken into account.The algorithms dynamically change the server scale according to the workload demand,and redundant servers are switched to energy-saving modes to reduce energy consumption.The work of this thesis is as follows.(1)The energy consumption model of data center is established upon theoretical analysis and experimental measurements.According to the operating states of the server and the transition between states,the energy consumption of the server is divided into working energy,sleeping energy,and transitioning energy,and each part is defined in mathematical forms.On the basis of the proposed models,the energy-saving optimization problem in homogeneous data centers is defined.By sharing the load evenly among active servers,the energy optimization problem is transformed into the scheduling problem of server resources,which could be solved by an offline algorithm with complete knowledge of load information.The algorithm adopts dynamic programming technology and produces an optimal solution to the problem.(2)In order to improve the provisioning efficiency of server resources,an adaptive workload prediction algorithm is proposed,which gives data centers some insights into near-future load and the ability of proactive provisioning.The algorithm selects a suitable prediction model in real-time manners according to workload changes,and it outperforms traditional methods in terms of prediction accuracy where single models are adopted throughout the predicting procedure.(3)The most important work in this thesis is to design and implement energy-aware load balancing algorithms.A reactive algorithm without adopting prediction is firstly proposed,which considers service availability reduction during bursty load periods and server state switching process.Then,a proactive algorithm with load prediction is proposed.Both of them are developed for homogeneous data centers.For heterogeneous data centers,the previous two algorithms are extended to a distributed algorithm by adding multi-level loading balancing fabrics into the architecture of existing data centers.Simulation results indicate that energy-aware load balancing algorithms presented in this thesis not only have good energy-saving effect,but also maintain high service availability in both homogeneous and heterogeneous data centers.