Modeling And Algorithm Design For Green Scheduling In Cloud Data Centers

The continuous advancement of cloud computing technology brings along the booming development of data centers,reflected not only in the number of growth,but also in the expansion of scale.At the same time,the problems of high energy consumption,high cost and high carbon emissions are becoming prominent,restricting the development of cloud computing.Headed by Google,the industry are beginning to build green cloud data centers,involving the key technologies including: energy monitoring,power-saving deployment and green scheduling.Therefore,this thesis will study the existing problems in these aspects,including the following four parts:First,power metering models and methods for cloud servers.From the user's perspective,cloud server can be classified into two types: physical machine and virtual machine.As basic scheduling unit of data center,server power monitoring in real time is of great significance.In this thesis,we first estimate the energy consumption of physical server from model and method point of view,paving the way for the subsequent research regarding energy scheduling algorithms of data center.The virtual machines running on the same server contribute differently to the total energy consumption.In order to realize charging and management in a pay-by-demand way like water,electricity,gas,it is also of great significance to measure the energy consumption of each virtual machine.For physical machine power metering,we adopt regression tree method.It divides the dataset of vectors that are composed of collected physical machine resources and power into several subsets,and then makes linear regression for these subsets one by one.It proved that regression tree is more accurate than existing methods.Based on this model,a fair dividing method is proposed to estimate the energy consumption of each virtual machine running onside.To be more objective,novel evaluation methods of accuracy and stability are proposed.The new computational method of accuracy solved the problem that the accuracy is too high due to the too large proportion of basic server power in traditional method.Second,multi-sleep modes based energy-saving scheduling for a datacenter.Make schedule of the servers between active state and different sleep states,such that the total energy consumption of the data center can be minimized and Qo S constraint can be met.Currently,most servers have ACPI power management function,so the servers can be switched to different sleep modes to save energy.The servers in a deeper sleep state usually consume less energy,but taking more time to be waken up.For the problem of power-saving scheduling of servers with multi-sleep modes,we first formulate it into a large-scale integer linear programming problem.Due to its too large problem size,it is divided into sub-problems of the same type,which can be solved using optimization tools like Cplex.Since there is correlation between adjacent sub-problems,the Backtrack-andUpdate method is proposed to adjust the servers' states,so as to ensure that the servers in transition states during the breakpoint will not be violated while the latter sub-problem has feasible solution.In order to reduce the energy consumption of active servers,integer linear programming modeling method is used again to make schedule of the frequencies of active servers in each time slot.Experiment results show that the proposed multi-sleep mode scheduling algorithm is more energy-efficient than the existing methods.Third,DVFS based energy-saving scheduling for a data center.Make schedule of tasks,frequencies and On/Off switchings of the servers,such that the total energy consumption of the data center can be minimized.Suppose each task has its own arrival time,execution time and deadline.Within the deadlines of the tasks,a server can choose to lower down its frequency to save energy using DVFS,or to adjust the begin time of the tasks to widen their time gaps in between to save energy by On/Off switchings.However,the decisions about task dispatching,frequency adjusting and On/Off switchings are restricting and affecting each other.For this problem,a three level frequency adjustment algorithm is proposed: 1)Based on the optimal frequency level,the tasks are dispatched in a dense way into different servers according to largest length first strategy.2)If the time gap between adjacent tasks is large enough and using On/Off switching is more energyefficient,the tasks next to this gap will be adjusted to the most proper frequency level.3)Then,the problem of task scheduling on each server can be transformed into an integer linear programming problem,which can be solved using Cplex.Thus,the start time and execution frequency of each task can be determined.Experiment results show that our method can save more energy than the existing work.Forth,green scheduling for distributed data centers.Make schedule of different types of energy and the tasks with multi-dimensional resource demand into the servers in different data centers,such that the total energy cost and carbon emissions can be minimized,respectively.We assume there are five energy sources for each data center: the purchased energy from the power grid,wind and solar energy from the renewable power plant,and the self-generated wind and solar energy by each data center.In order to make full use of the time-varying and location-varying electricity prices and renewable energy,ESDs(Energy Storage Devices)are introduced to store different types of energy.The energy of each data center can also be sold back to the power grid.For the problems of minimizing cost and carbon emissions,we dispatch as many tasks as possible to the data centers with higher priorities of lowest computing price priority(LCPP)and most green priority(MGP),respectively.The tasks dispatched to the same data center will then be packed into as few servers as possible by using resource vector matching method(RVM).At this time,the energy scheduling problem is transformed into a large-scale mixed integer linear programming problem,which can be solved by the branch-and-bound method.Experiment results show that the green scheduling framework and algorithm proposed in this thesis can further reduce the total energy cost and carbon emissions for data centers.