Power Aware Optimization And Scheduling In Data Centers

The ever-increasing cloud-based services,big data analytics,e-commerce,and Internet traffic make the energy consumption of data centers grow faster and faster.High energy efficiency(EE)becomes one of the major criteria of data center design and operation.Performance per Watt is the basic metric for measuring the server's energy efficiency.Because servers have different EE at different levels of utilization,energy proportionality(EP)and energy proportional computing gradually become the hottest research topic in both industry and academia.For an ideally energy proportional server(EP=1.0),its power consumption is proportional to its utilization.For example,if a server consumes 100 watts when its utilization is 100%,then it should consume 30 watts when its utilization is 30%,and it should consume nearly zero watts when it's idle.Unfortunately,current power aware scheduling approaches pack as many jobs as possible to a subset of servers until they are saturated such that the remaining servers become idle or should be powered off.In a highly heterogeneous data centers with multiple generations of servers,such kind of approaches not only mask the energy proportionality of individual server but also deteriorate the overall energy efficiency of the data center.Therefore,good understanding of server's energy efficiency and energy proportionality can help optimize workload placement and job scheduling as well as improve the data center's energy efficiency and Qo S.Thus,in this thesis we combine the server's EE and EP to analyze the data center's overall energy efficiency.Since the resource utilization of data center is greatly affected by the workload characteristics,we propose task scheduling approach for data centers based on EE and EP.Furthermore,we propose e Scope,an energy efficiency simulator for energy efficiency evaluation to alleviate the impact of heterogeneity on data center caused by hardware iteration and replacement.The main contributions of this thesis are as follows:(1)We thoroughly analyze the published results of SPECpower benchmark from 2007 to 2017,and propose the PEEP metric,i.e.,the ratio of peak energy efficiency over energy efficiency at peak utilization.The SPECpower results are reorganized by their hardware availability year.Then the relationship between EP stagnation in recent years and microarchitecture is explored and validated.We also find that the power consumption when the server is idle and the number of identical nodes are the key factors that affect the server's EP.In addition,we find that the peak EE is from 100% utilization to other utilization such as 60%-70% and the impact of hardware configuration on EE contributes the most to the improvement of EP.(2)We propose software-defined architecture for EE,and conduct extensive experiments to induce the relationship between EE and hardware configuration,and research on large memory system.The aforementioned analysis of SPECpower results shows that the hardware configuration has impact on the server's EE.In order to explore the relationship between server hardware configuration and EE,we setup SPECpower environment on four 2U servers and conduct experiments to explore how the CPU frequency scaling and memory installation affect server's EE.In addition,as DRAM plays a more important role in modern servers and processors,we conduct STREAM benchmark testing on three 2U servers and conducts experiments to get the EE features of large memory systems when running memory-intensive applications.(3)We propose energy-aware batch-online hybrid task scheduling approach and e Scope,the real server based energy efficiency simulator for large scale data center energy efficiency simulation and evaluation.Due to the reconfigurable hardware configuration and workload complexity and variety,we design an energy efficiency-aware data center batch-online hybrid task scheduling strategy to make full use of data center resources,and improve overall data center EE and reduce energy consumption.By using e Scope,the proportion of different server and utilization rates of servers can be configured so that the completed workload of data center can be maximized when a certain power capping is specified.The research outputs of this thesis,i.e.,the SPECpower analysis results from the perspective of EP and EE,the relationship between server configuration and EE explored from extensive experiments,the proposed energy-aware batch-online hybrid task scheduling approach,and the e Scope simulator,can help system design,workload placement and task scheduling for large scale data centers.