Energy-saving Scheduling Of Servers With Multi-sleep Modes For Cloud Data Center

With the rapid increasing demand of cloud services,the cloud service providers have deployed a large number of data centers.Meanwhile,the energy consumption of data center has increased dramatically,and a large part is wasted.On the one hand,the servers may be in low utilization due tounreasonable scheduling of tasks.On the other hand,servers are always over-provisioned to meet the peak demand of requests,leaving a lot of servers in idle states when there are few requests.In order to solve the problems,studies are focusing on two studies: request scheduling and energy management.In request scheduling,most researches focus on improving the current utilization of resource through transforming the problem into a bin-packing problem.They ignoredthat the time demand of the requests have effect on resource utilization.In energy management,most researches always switch the idle server into a single ideal low-power sleeping state,but ignoring the power and delay of during wake-up and sleep-down transitions.Most modern servers have multi-sleep states,and more energy will be saved if properly used.Taking into account of the power and delay during the transitions,we first givethe energy model of the data center based on ACPI multi-sleep model.And then the problem is formulated as a constraint optimization with objectiveand constraints.After that,a time-based request scheduling and multi-sleep energy management algorithm is proposed to solve this problem.Different from existing work that mainly use resources matching policies,we propose a time-matching-first algorithm.We analyze the effects of request lifetime on the total energy.The longedthe request is,the longer the servers will run.The total energy will be reduced if the tasks with similar end time onto the same server.In other word,it will take more running time when a request later than the latest end time of requests is allocated onto the server,lowering the utilization.After analyzing the different end time matching schemes,we select the best onefor our scheduling algorithm.Based on requesttracesfrom real data center,we compare our algorithm with resource matching algorithms.Experiment results show that our algorithm is more energy-efficient than resource matching algorithm.In this thesis,we propos two multi-sleep energy management algorithms for online and offline scenarios,respectively.For offline condition,the incoming requests during the scheduling period are known in advance.When the computing capacity of the servers isinsufficient,we backtrack to adjust the server states to increase the number of active servers.We analyze the extra energy consumption of backtrack and select the most energysaving way of adjusting.For online condition,we only know the information of history requests.We analyze the influence of different sleep modes on the processing capacity of the data center.After that,we predict the loads in the durations corresponding to the wakeup delays of the sleep modes,and then adjust the corresponding sleep servers to deal with the short term and long term volatility of load.Meanwhile,we also leave some servers in idle and delay the sleep-down operations for the servers,so as to ensure the processing capacity while saving as much energy as possible.Compared with single sleep-mode and non-prediction energy management algorithm,our algorithm can effectively reduce the total energy for data center.