A Survey on Energy-Efficient Data Management
Total Page:16
File Type:pdf, Size:1020Kb
A Survey on Energy-Efficient Data Management Jun Wang, Ling Feng Wenwei Xue, Zhanjiang Song Dept. of CS&T, Tsinghua Univ., Beijing, China Nokia Research Center, Beijing, China [email protected] [email protected] [email protected] [email protected] ABSTRACT was about 1.5% of the total U.S. electricity con- Energy management has now become a critical and ur- sumption in 2006, and this energy consumption is gent issue in green computing. A lot of efforts have been expected to double by 2011 if continuously powering made on energy-efficiency computing at various levels computer servers and data centers using the same from individual hardware components, system software, methods [11].” Xu et al. showed that electricity con- to applications. In this paper, we describe the energy- sumed by computer servers and cooling systems in a efficiency computing problem, as well as possible strate- typical data center contributes to around 20 percent gies to tackle the problem. We survey some recently of the total ownership cost, equivalent to one-third developed energy-saving data management techniques. of the total maintenance cost [42]. When a data Benchmarks and power models are described in the end center reaches its maximum provisioned power, it for the evaluation of energy-efficiency solutions. has to be replaced or augmented at a great expense [33]. In the very near future, energy efficiency is ex- pected to be one of the key purchasing arguments 1. INTRODUCTION in the society. Now and in the future, green computing will be a Nowadays, power and energy have started to seve- key challenge for both information technology and rely constrain the design of components, systems, business. Green computing aims at environmentally computing clusters, data centers, and applications. sustainable computing and responsible use of com- Better equipment design and better energy man- puters and related resources. Murugesan et al. de- agement policies are desirable to address these con- fined the field of green computing as “the study and cerns. practice of designing, manufacturing, using, and dis- More and more computer designers and users are posing of computers, servers, and associated subsys- concerned about energy-efficient computing. In this tems such as monitors, printers, storage devices, survey article, we overview these great efforts, with and networking and communications systems effi- an emphasis on energy-efficient data management, ciently and effectively with minimal or no impact which was ignored and starts to get attention very on the environment [28].” recently. With the limited primary sources of energy and The remainder of this paper is organized as fol- rapid climbing of energy demanded by computing, lows. In Section 2, the energy-efficiency computing the commitment to reduce power consumption and problem and some guidelines to tackle the problem environmental impact becomes increasingly impor- are discussed. In Section 3, we overview work done tant. Energy efficiency thus constitutes a focal point on energy-efficient data management. Evaluation for green computing. techniques on energy performance including bench- In fact, most people in the world today are aware marks and power models are described in Section 4. of the energy problem at a high level: even if our pri- We conclude the paper in section 5. mary sources of energy are running out, the demand for energy in both commercial and domestic envi- 2. FUNDAMENTALS OF ENERGY EF- ronments is increasing, and the side effect of con- FICIENT COMPUTING sistent energy use influences negatively our global In this section, after a brief description of the environment. Based on the report of the US Envi- energy-efficiency problem, we outline some feasible ronmental Protection Agency, “the servers and data ways to tackle the problem. centers in USA alone consumed about 61 billion kilowatt-hours (kWh) at a cost of $4.5 billion, which 2.1 Energy-Efficiency Problem SIGMOD Record, June 2011 (Vol. 40, No. 2) 17 investigate where the power is spent and how to op- Energy consumption can be generally defined as: timize the power usage. Within a computer system, there are generally five energy consumers, namely, Energy = AvgP ower T ime × processor, disk, memory, I/O devices, and chipset. where Energy and AvgP ower are measured in Achieving energy-efficiency requires improvements Joule and W att, respectively, and 1 Joule =1W att in the energy usage profile of every system compo- 1 Second. × nent. Energy efficiency is equivalent to the ratio of per- 2) Adopting Power-Manageable Hardware Com- formance, measured as the rate of work done, to ponents. Adopting power-manageable hardware com- the power used [37] and the performance can be represented by response time or throughput of the ponents could help improve energy-efficiency. For computing system. example, the voltage of hardware components can be increased or decreased through dynamic voltage Workdone Workdone Energy Efficiency = = scaling (DVS), which is a power management tech- Energy Power Time × nique in computer architecture, depending upon cir- P erformance = (1) cumstances. Dynamic voltage scaling to decrease Power voltage is known as undervolting, and this situation The main approach towards energy-efficiency is can conserve power [12]. In addition, employing efficient power management. According to equa- small form factor disk drives, solid state disk drives, tion (1), there are two ways to enhance energy- large memory configurations, low power processors efficient computing: either improving the perfor- and memories could decrease power consumption mance with the same power, or reducing power con- [31]. HP and IDC also estimated that about 69 sumption without sacrificing too much performance. percent energy reduction can be achieved within a For energy-efficient systems, while maximal perfor- three-year period for IT organizations that migrate mance for some tasks (or the whole workload) is to blade self-contained architecture, where blades still desirable in some cases, the systems must also can span from servers and storage devices to work- ensure the energy usage is minimized. Preferably, a stations and virtual desktops [20, 1]. computing system consumes the minimum amount 3) Building Power Models for Computing Sys- of energy to perform a task at the maximal perfor- tems. Also, one needs to know how a computing mance level [10]. system is constructed and how an energy-efficient Note that the relationship between performance system operates. It is important to construct a and energy efficiency is not mutually exclusive. A power model that allows the system to know how maximal performance could also be achieved by de- the power is consumed, and how the system can activating some resources or lowering certain in- manipulate and tune that power [10]. dividual performance without affecting the work- 4) Understanding and Measuring System Perfor- load’s best possible completion time or throughput mance. To counter for performance with the least in order to optimize energy usage. Brown et al. power consumption, computing systems must have treated energy efficiency as an optimization prob- ways to timely understand and measure system per- lem [10]. To minimize the total energy, an energy- formance related to task execution under different efficient system must adjust the system’s hardware dynamic workloads. resources dynamically, so that only what is needed 5) Constructing Energy Optimizers. The system to execute tasks is made available. Rivoire et al. must accommodate an energy optimizer component, pointed out two major complementary ways to solve which is responsible for an energy-efficient hardware the energy-efficiency problem: either building en- configuration throughout the system operation at ergy efficiency into the initial design of computer all times. The optimization approaches may be components and systems, or adaptively managing based on either heuristic or analytical techniques, the power consumption of systems or groups of sys- as indicated by Brown et al. in [10]. tems in response to changing conditions related to 6) Reducing Peak Power. Barroso et al. ex- the workload or environment [36]. plained that current desktop and server processors can consume less than one-third of their peak power 2.2 Solution Guidelines at very low activity modes, which can thus save To deliver effective solutions to the energy-efficiency around 70 percent of peak power [7]. Tsirogiannis et problem, the following six considerations can be al. indicated that almost 50 percent of peak power taken as the solution design guidelines. is actually consumed at idle [37]. 1) Comprehensive Examination of System Com- ponents. To save power consumption, we shall first 18 SIGMOD Record, June 2011 (Vol. 40, No. 2) 3. ENERGY-EFFICIENT DATA MANAGE- commercial DBMS, respectively. MENT Over the past few decades, performance remains 3.1.2 Software-Based Approaches as the main goal of database management systems Hardware-based approaches constitute only part (DBMSs). In light of an increasing concern about of solutions. Considering hardware heterogeneity energy, energy management starts to draw atten- and limited power knobs that most hardware of- tion of the database community. The 2008 Clare- fers today, data management software shall play an mont report on database research emphasized ex- effective role in energy optimization as well. Phys- plicitly the importance of power-aware DBMSs that ical data independence and query optimization of limit energy costs without sacrificing scalability [4]. DBMSs