Researches On DVFS-based Task Scheduling Problem In Data Centers

Data centers run lots of tasks to provide users services every day.Recent years the high energy consumption problem of the data centers is an important issue since the size of data centers grow fast.One serious fact is that data center electricity consump-tion occupies 1.3%of total consumption in the world.Energy efficient task scheduling in data centers becomes urgent in such circumstances.Dynamic Voltage Frequency Scaling(DVFS)technology is a method to make processors work at suitable frequency,makes it possible to achieve power saving via adjusting the frequency dynamically.Basically all hardware manufacturers support DVFS technology so far.Generally,users apply DVFS via software modules such as cpufreq,which is integrated into Linux kernel,in order to eliminate the difference of different hardware manufacturers.The usage of this manner is uniform and simple.However,lower frequency will reduce the computing capacity of the servers,which leads to using more servers for a given total workload in a data center.It is a challenging problem to make a tradeoff between the number of servers and the frequency of each server in this settings.This paper give a definition of DVFS-based task scheduling problem and develop a solution,which aims at making a tradeoff between the number of servers and the frequency of each server.Furthermore,we propose new DVFS-based performance and power consumption models through analyzing a large amount of data from real experiments.Based on these models,we redefine the problem of DVFS-based task scheduling with the target of minimizing power consumption.Moreover,we prove the problem is NP-Hard and propose algorithms to solve it.Our work is presented as follows.Firstly,we design a simulation scenario,where tasks are scheduled on hundreds of simulated servers.The power consumption data of these simulated servers come from reality.We get the conclusion that the least amount of servers will not get the most power consumption savings,which is not accordant to usual knowledge via several experiments.Then we find out that,the traditional DVFS-based power consumption is not ac-curate after we analyzed and summarized many previous related works.We propose a new DVFS-based power consumption model according to extensive real experiments.The new model take into account both the processors' frequency and utilization,and the experiments results show our model is more accurate.Based on the new model,we redefine the problem of DVFS-based task scheduling and propose our methods of solutions.In the solution part,we present a power-aware algorithm named Power-aware Threshold Unit(PTU)algorithm to schedule the online tasks in a data center dynam-ically.The main idea of PTU algorithm comes from threshold idea,and the fact that data centers are rack-sized both.The PTU algorithm holds a uniform maximum server-s' frequency threshold in a multi-server sized unit(several servers,such as rack).That is,all servers in this unit could not exceed the frequency threshold.The algorithm in-crease the uniform threshold of this unit in cases that all servers all can not deal with this task.We carry out extensive real experiments on a cluster which consists of several multi-core servers.To the best of our knowledge,we are the first to propose the new power con-sumption characterization method.The experiment results in real scenario show that the new DVFS-based power consumption model is very accurate.Especially,the PTU algorithm,which is implemented in the real scenario,can get the nearest power con-sumption savings to the most power consumption savings PBF(Power-aware Best Fit)algorithm amongst traditional algorithms,and has the potential to deploy in real data centers since it is designed based on the fact that data centers are rack-sized,in order to get more power consumption savings.