PaperlayerView exact revision on arXiv

Source-linked AI summary

A Taxonomy and Survey of Energy-Efficient Data Centers and Cloud Computing Systems

Anton Beloglazov, Rajkumar Buyya, Young Choon Lee, Albert Zomaya

arXiv:1007.0066v2cs.DC

TL;DR

Rising computing-system energy use creates financial, thermal, and environmental concerns, motivating energy-efficient design. This chapter synthesizes and classifies research across hardware, operating systems, virtualization, and data centers, identifying advances and future directions. Its supported conclusion is a taxonomy and survey that organizes approaches and open challenges for energy-efficient computing.

  • Problem

    Increasing computing-system energy consumption raises operating costs, thermal concerns, and CO2 emissions, creating a need to understand energy-efficient design across system levels.

  • Method

    The chapter surveys and classifies energy-efficient resource-management research across hardware and firmware, operating systems, virtualization, and data centers.

  • Results

    The chapter presents a taxonomy of power and energy management and maps research approaches to hardware, operating-system, virtualization, and data-center levels.

  • Takeaways & Limitations

    The synthesis identifies advancements and open research directions for improving energy-efficient computing.

  • Takeaways & Limitations

    The surveyed hosting-center approach manages only CPU resources, uses outdated APM rather than ACPI, omits thermal and switching-latency factors, and becomes relatively expensive under major workload changes.

Abstract

from arXiv · show

Traditionally, the development of computing systems has been focused on performance improvements driven by the demand of applications from consumer, scientific and business domains. However, the ever increasing energy consumption of computing systems has started to limit further performance growth due to overwhelming electricity bills and carbon dioxide footprints. Therefore, the goal of the computer system design has been shifted to power and energy efficiency. To identify open challenges in the area and facilitate future advancements it is essential to synthesize and classify the research on power and energy-efficient design conducted to date. In this work we discuss causes and problems of high power / energy consumption, and present a taxonomy of energy-efficient design of computing systems covering the hardware, operating system, virtualization and data center levels. We survey various key works in the area and map them to our taxonomy to guide future design and development efforts. This chapter is concluded with a discussion of advancements identified in energy-efficient computing and our vision on future research directions.

1 Introduction

Computing systems have improved performance while total power use, operating costs, and CO2 emissions continue to rise. The chapter surveys energy-efficient design across hardware, operating systems, virtualization, and data centers to classify advances and open challenges.

  • Motivation: Performance growth has not reduced total computing-system power draw, and continued trends could make lifetime server energy costs exceed hardware costs.The introduction attributes performance growth to more efficient designs and increasing component density.
  • Motivation: High energy consumption increases resource-provider ownership costs through electricity bills and contributes to carbon dioxide emissions.Energy efficiency therefore affects both infrastructure economics and environmental impact.
  • Background: Green computing practices emerged alongside initiatives such as Energy Star and later certification and industry programs promoting energy-efficient IT equipment.The introduction traces these developments from 1992 onward.
  • Scope and contribution: The chapter classifies research on energy-efficient computing at the hardware and firmware, operating-system, virtualization, and data-center levels.It also discusses common characteristics, open challenges, and future research directions.

2 Power and Energy Models

The chapter distinguishes power as the rate of performing work from energy as work accumulated over time, then relates consumption to components, utilization, and management techniques. It presents models and mechanisms for estimating and reducing system power.

  • Basic concepts: Power is the rate of work, whereas energy is the total work performed over a period; reducing power does not necessarily reduce energy.Lower CPU performance can lengthen execution enough to leave consumed energy unchanged.
  • Power sources: Dynamic CMOS power depends on switching activity, capacitance, supply voltage, and clock frequency, motivating DVFS through coordinated voltage and frequency reduction.DVFS down-scales CPU performance when the processor is not fully utilized.
  • System components: Memory, power-supply losses, and storage are important consumption sources alongside the CPU, with memory power increasing as server memory capacity grows.A server with eight 1 GB DIMMs consumes about 80 W, while power supplies are most efficient at 50-75% load.
  • Power modeling: A linear CPU-utilization model and an empirical model estimate server power from utilization and idle or full-load values, requiring calibration for each machine class.The empirical model includes calibration parameter r obtained experimentally.
  • Power modeling: Experiments on several thousand nodes yielded prediction errors below 5% for the linear model and 1% for the empirical model.The presented empirical results use calibration parameter r = 1.4.

3 Problems of High Power and Energy Consumption

High power and energy consumption create infrastructure, operating-cost, thermal, performance, and environmental problems. Low average utilization forces over-provisioning for rare peaks, while cooling and conversion losses consume substantial energy.

  • Responses: Hardware improvements and software approaches both contribute to energy efficiency and should be treated as complementary.The chapter also identifies user awareness as a factor in energy-efficiency discussions.
  • High energy consumption: Peak power determines infrastructure requirements, whereas average power over time determines energy consumption and electricity costs.The chapter notes that cooling and UPS/PDU conversions can consume 50% of data-center energy.
  • High power consumption: Low average resource utilization makes data centers inefficient because infrastructure must support peak loads that rarely occur.Over-provisioning includes cooling, power delivery, generators, UPS, and related capacity.
  • High power consumption: For each watt consumed by computing resources, cooling requires an additional 0.5 to 1 W, and overheating can reduce component lifetime and increase errors.At LLNL, dissipating 1 W from an HPC system requires 0.7 W of additional cooling power.
  • Environmental impact: Growing ICT energy use contributes to CO2 emissions, with the ICT industry estimated at about 2% of global emissions in 2007.The EPA projection cited in the chapter rises from 42.8 MMTCO2 in 2007 to 67.9 MMTCO2 in 2011.

4 Taxonomy of Power / Energy Management in Computing Systems

The taxonomy organizes power management by static versus dynamic techniques and spans hardware, software, virtualization, and data-center contexts. It emphasizes runtime adaptation, resource utilization, and quality-of-service constraints.

  • Static and dynamic management: Static Power Management optimizes power at design time across circuit, logic, architecture, and system levels.These methods target switching activity and the mapping of high-level specifications onto implementations.
  • Static and dynamic management: Dynamic Power Management adapts system behavior at runtime according to resource requirements or other changing system-state characteristics.Its use assumes variable workloads and some ability to predict workload behavior.
  • System levels: The survey covers power-management research for devices, operating systems, virtualized and non-virtualized servers, clusters, data centers, and Clouds.The taxonomy is developed across multiple system scales and management contexts.
  • Hardware and firmware: DVFS provides a broad CPU dynamic-power range by scaling performance according to current requirements while attempting to limit performance degradation.The chapter notes that coordinated scaling has also been applied across multiple servers.
  • Virtualization and Cloud: Virtualization can place several virtual machines on one physical server, reducing hardware in use and improving resource utilization.Cloud computing builds on virtualization to provision resources on demand using a pay-as-you-go model.
  • Virtualization and Cloud: Cloud energy management must balance aggressive consolidation with reliable QoS expressed through service-level requirements such as throughput, response time, and latency.Workload variability can prevent some VMs from meeting their service requirements.

5 Hardware and Firmware Level

Hardware and firmware power management is organized around deactivating components or scaling their performance, with decisions constrained by workload variability and transition overheads.

  • Power-management taxonomy: Dynamic power management at hardware and firmware levels comprises Dynamic Component Deactivation and Dynamic Performance Scaling.DCD disables inactive components, whereas DPS adjusts performance in proportion to power consumption.
  • Dynamic Component Deactivation: DCD transitions are worthwhile only when inactivity lasts long enough to repay transition delay, reinitialization energy, and saved-power requirements.Unknown future workloads require prediction from historical data or system models.
  • Dynamic Component Deactivation: DCD prediction methods include static, adaptive predictive, and stochastic approaches for handling workload variability.Adaptive techniques adjust previously fixed parameters according to past prediction quality and are generally more efficient than static methods when workload type is unknown.
  • Dynamic Performance Scaling: DPS reduces CPU power by jointly adjusting clock frequency and supply voltage, but lower frequency also reduces instruction throughput and can extend CPU-bound execution time.Frequency scaling alone is rarely worthwhile for conserving switching power; voltage scaling is needed for greater savings.
  • Dynamic Performance Scaling: DVFS algorithms are commonly classified as interval-based, intertask, or intratask according to how they use CPU activity and workload structure.Interval-based methods predict near-future utilization from prior CPU-activity intervals and adjust voltage and frequency accordingly.
  • Firmware interfaces: ACPI provides platform-independent interfaces and runtime power states while leaving particular power-management policies to software developers.Its principal states include CPU C-states and implementation-dependent P-states representing DVFS settings.

6 Operating System Level

Operating-system energy management coordinates resource control, application behavior, and system-wide policies across devices and workloads. The surveyed systems target battery lifetime, energy allocation, performance preservation, and server power reduction, while exposing scalability and transition-time constraints.

  • Taxonomy: The OS-level taxonomy classifies power-efficient resource-management research by system characteristics summarized in Table 2.The reviewed scope includes CPU, memory, disk storage, and network interfaces, with goals spanning energy minimization and performance preservation.
  • CPU management: The ondemand governor monitors CPU utilization multiple times per second and selects frequency-voltage pairs to keep the CPU approximately 80% busy.It aims to minimize performance loss from reduced frequency and supports multiprocessor, multicore, and multithreaded architectures.
  • CPU management: Centralized utilization sampling can become significant overhead as CPU count increases, motivating parallel sampling and dedicated governor work queues.These improvements were under investigation for multiprocessor systems.
  • Application-aware management: ECOsystem treats energy as a first-class OS resource, using battery-lifetime targets and application priorities to allocate energy among competing applications.Its modified Linux-kernel implementation was experimentally evaluated for target battery lifetime and proportional energy distribution.
  • Cross-layer adaptation: GRACE coordinates global, per-application, and internal adaptation across system layers to respond to application resource demand.With CPU and network adaptations, adding application adaptation reached 32% energy savings and averaged 22%.
  • Server management: PowerNap targets idle-server power through fast transitions between sleep and fully active states, reporting a 74% reduction in average server power when requirements are satisfied.The approach requires transition times below 10 ms, while modern servers were reported at 300 ms.

7 Virtualization Level

Virtualization enables power management above physical hardware by coordinating hypervisor-level controls, guest-OS policies, VM migration, and resource consolidation. The surveyed framework also enforces power limits across shared VM and device resources.

  • Virtualization fundamentals: Virtualization abstracts operating systems and applications from hardware by partitioning physical resources into isolated virtual machines.The virtualization layer lies between hardware and the OS and is managed by a virtual machine monitor.
  • Power-management strategies: A VMM can manage power system-wide through DVFS or deactivation, or coordinate guest-OS power calls with hardware state changes and system-wide limits.The second approach incorporates OS-specific policies and application-level knowledge.
  • Virtualization technologies: VM migration transfers workloads between physical hosts and supports power-aware consolidation and load balancing.Live migration aims to avoid user-visible suspension during transfer.
  • Virtualization technologies: Xen, VMware, and KVM support hardware power states, while VMware additionally uses live migration and DPM to switch off spare servers.Xen supports ACPI P- and C-states; VMware DPM consolidates VMs to keep the minimum number of servers powered on.
  • Coordinated VM management: Stoess et al. address distributed energy accounting in virtualized systems with a host resource manager and an energy-aware OS.The host manager enforces power limits across VM instances and device types, including limits imposed by batteries, generators, thermal constraints, or cooling capacity.
  • Evaluation and future work: The prototype enforced power limits for both energy-aware and energy-unaware guest operating systems.Future work identified multiple-power-state devices, hardware-assisted virtualization, and multicore architectures.

8 Data Center Level

Data-center power management primarily addresses low resource utilization by consolidating workloads, powering down idle resources, and coordinating performance and energy controls. The taxonomy covers diverse consolidation, DVFS, migration, and power-switching approaches.

  • Workload consolidation: Workload consolidation allocates requests or virtual machines to minimal physical resources and powers down or sleeps idle resources.The allocation problem includes placing new requests and managing the resulting performance.
  • Surveyed techniques: The reviewed data-center methods combine DVFS, workload or VM consolidation, server power switching, migration, and geographical distribution.Their stated goals include minimizing energy or power while satisfying performance requirements and power budgets.
  • Surveyed objectives: The taxonomy includes approaches that allocate power budgets to minimize mean response time, maximize utilization, or maximize profit while considering CO2 emissions.These objectives appear across the reviewed server-farm and cloud-computing studies.

8.1 Implications of Cloud Computing

Cloud computing offers economical resource use for consumers and providers, while energy minimization is central to provider profitability. More effective provisioning can reduce energy consumption by avoiding under- and over-provisioning.

  • Clouds provide scalable, reliable, and economical services through pay-per-use consumption and improved utilization of poorly utilized provider resources.Providers seek to maximize profit, making energy consumption an important management concern.
  • 13% energy reduction can be obtained by more effectively handling resource provisioning and avoiding under- or over-provisioning.

8.2 Non-Virtualized Systems

Non-virtualized systems reduce energy through workload consolidation, power-state control, DVFS, and workload-aware scheduling. The surveyed approaches show substantial savings but commonly rely on restrictive assumptions and omit important resources or transition costs.

  • 8.2.1 Load Management for Power and Performance in Clusters: Load concentration switches idle homogeneous cluster nodes off while using throughput and execution time as QoS constraints.The approach periodically monitors demand and controls node activation, but static demand estimation can mispredict changing workloads.
  • 8.2.2 Managing Energy and Server Resources in Hosting Centers: Muse dynamically scales power demand in hosting centers to balance service quality and price while reducing operating costs, emissions, thermal vulnerability, and over-provisioning.Its economic framework balances resource cost against estimated service utility and supports dynamic matching of load and power.
  • 8.2.2 Managing Energy and Server Resources in Hosting Centers: Muse is limited to CPU management, uses outdated APM rather than ACPI, omits thermal effects and switching latency, and becomes expensive during major workload changes.
  • 8.2.3 Energy-Efficient Server Clusters: VOVO, IVS, and coordinated policies combine node switching with CPU voltage scaling under fixed single-application cluster assumptions.
  • 8.2.3 Energy-Efficient Server Clusters: Up to 29% energy savings are reported for IVS, up to 42% for VOVO, and 18% higher savings for VOVO-CVS than VOVO alone.
  • 8.2.4 Energy-Aware Consolidation for Cloud Computing: The consolidation heuristic uses Euclidean distance to allocate applications and achieves energy use about 5.4% above optimal in experiments.It supports heterogeneous environments but assumes known, constant resource requirements and omits migration and server switching costs.
  • 8.2.5 Optimal Power Allocation in Server Farms: Power allocation can improve server-farm performance by up to 5× and 1.4× on average, with workload-dependent benefits from running more servers at lower performance.
  • 8.2.6 Environment-Conscious Scheduling of HPC Applications: Energy-centric scheduling policies reduce energy costs by 33% on average, while a multi-objective policy reduces carbon emissions and maintains high provider profit.

8.3 Virtualized Systems

Virtualized systems coordinate hardware scaling, VM consolidation, guest policies, and resource allocation to improve energy efficiency while preserving QoS or yield. The surveyed methods remain constrained by workload assumptions, optimization complexity, incomplete resource models, and scalability or SLA limitations.

  • 8.3.1 VirtualPower: Coordinated Power Management: Virtualization supports fault and performance isolation, VM migration, and hardware and software heterogeneity in power-efficient data-center management.
  • 8.3.1 VirtualPower: Coordinated Power Management: Soft resource scaling emulates hardware scaling by reducing a VM's resource-use time when hardware scaling is unavailable or yields little power benefit.
  • 8.3.1 VirtualPower: Coordinated Power Management: VirtualPower coordinates local guest-VM policies with global resource management, but the surveyed approach does not ensure QoS or variable application SLAs.
  • 8.3.2 Coordinated Multi-level Power Management: Coordinating multiple power-management policies reduces power across workloads, although savings vary with workload and coordination benefits remain qualitatively similar.
  • 8.3.3 Power and Performance Management via Lookahead Control: Lookahead control predicts requests with a Kalman filter and optimizes VM reallocations while accounting for switching risks, transient consumption, and opportunity costs.
  • 8.3.3 Power and Performance Management via Lookahead Control: The lookahead controller reaches 30 minutes of execution time on 15 hosts, limiting suitability for large-scale real-world systems.
  • 8.3.4 Resource Allocation using Virtual Clusters: Virtual-cluster allocation models address parallel and dynamic workloads, but large-scale heuristics are absent and CPU-only optimization assumes resource needs known in advance.
  • 8.3.5 Resource Allocation in Multi-Application Virtualized Data Centers: Linear programming with heuristics allocates CPU and RAM proportionally to application priorities and reports satisfaction of defined SLAs across web, database, and office workloads.The approach requires machine learning to obtain application utility functions.

8.3.6 Shares and Utilities based Power Consolidation

The surveyed approaches consolidate or place virtual machines to reduce power while preserving utility, performance, quality of service, or migration costs. They span static allocation, dynamic controllers, bin-packing formulations, and multi-level cloud provisioning.

  • Shares and Utilities based Power Consolidation: PowerExpandMinMax combines VM resource bounds and utility-aware allocation, outperforming other policies across varied VM, server, and power-cost inputs.A real-testbed experiment improved overall data-center utility by 47%.
  • pMapper: Power and Migration Cost Aware Application Placement: pMapper formulates dynamic VM placement as continuous optimization that minimizes power while maximizing performance under SLA constraints.Its algorithms adapt bin packing to variable server sizes and power costs, while pMaP explicitly trades off power reduction against migration cost.
  • pMapper: Power and Migration Cost Aware Application Placement: The mPP and mPPH algorithms first determine target server utilization and then pack applications, with mPPH tracking prior placement to reduce migration cost.The two phases separately address cost minimization and application placement.
  • Resource Pool Management: Reactive Versus Proactive: A workload-placement controller using historical traces and a reactive migration controller using real-time utilization jointly manage VM consolidation under quality-of-service requirements.The reactive controller responds to overload and underload fluctuations at small time scales.
  • Resource Pool Management: Reactive Versus Proactive: Using both controllers produced the best results for 138 SAP applications, with the best policy requiring 10-20% more CPU capacity than the ideal case.The policy invokes workload placement every 4 hours and runs migration control in parallel for overloads and underloads.
  • Cloud Resource Provisioning: GreenCloud provisions cloud resources through broker-mediated VM allocation using CPU and deadline requirements, and δ-Advanced-DVS achieved the best profit per unit of consumed power in simulation.Adaptive-DVS was limited by its simplified queueing model.

9 Conclusions and Future Directions

Energy efficiency is a central requirement for modern computing systems, and the chapter synthesizes research across hardware, operating systems, virtualization, and data centers. It identifies requirements and future directions for more comprehensive Cloud resource management.

  • Conclusions: Energy-efficient computing research is organized across hardware, operating system, virtualization, and data-center levels because modern systems incur substantial power costs and CO2 emissions.The chapter frames energy efficiency as an important design requirement for Clouds and data centers.
  • Conclusions: DVFS and ACPI enable software control of CPU power consumption by continuously matching frequency and voltage to performance requirements.This control trades performance against power consumption.
  • Conclusions: Data-center power management is complicated by VM migration delays, performance overhead, and optimization that often considers CPU rather than multiple resources.The chapter also notes that migration is not always applied to improve runtime allocation.
  • Conclusions: Generic Cloud solutions should support virtualization, adaptive VM migration, multiple applications and SLAs, QoS guarantees, mixed workloads, and scalable optimization.The listed requirements also include decentralization and fault tolerance.
  • Future directions: Future research should optimize multiple resources, including CPU, memory, disk storage, and network interfaces.The chapter specifically identifies network interconnects as significant data-center energy consumers.
  • Future directions: Future directions include network-topology optimization, more efficient power supplies, lower transition and migration overheads, geographically distributed workload placement, and finer-grained user control.Geographic distribution could relocate workloads to locations with cheaper energy or cooling.

Appendix A. Operating system level research works.

The surveyed operating-system research applies power management through DVFS, adaptation, resource accounting, and low-power operating-system mechanisms. The appendix compares approaches by their control strategy, resource scope, application assumptions, and objectives.

  • Operating-system methods: The Ondemand governor monitors CPU utilization multiple times per second and selects a frequency-voltage pair for current performance requirements.Its goal is to limit performance loss while adjusting power dynamically.
  • Operating-system methods: ECOsystem introduces currentcy, distributing an energy-use budget among applications according to their priorities and resource consumption.Applications are expected to cooperate with the OS through a power-based API.
  • Operating-system methods: Nemesis notifies applications when energy consumption exceeds a threshold, requiring them to adapt their behavior to OS signals.This approach depends on application cooperation.
  • Operating-system methods: Operating-system approaches include DVFS, resource throttling, application adaptation, and low-power states for minimizing energy while preserving performance or battery lifetime.The surveyed applications include real-time multimedia and other real-time systems.
  • Operating-system methods: Some systems select among alternative DVFS algorithms automatically according to system characteristics, while others coordinate adaptation across global, application, and internal levels.The surveyed work includes Linux-based and distributed-file-system adaptations.
  • Operating-system methods: Power-saving systems also exploit short idle periods through fast transitions to system-wide low-power states while attempting to minimize performance loss.PowerNap is listed as an example of this approach.

Appendix B. Data center level research works.

Data-center research combines workload consolidation, power-state control, DVFS, resource allocation, and VM management to reduce energy while maintaining performance objectives. The surveyed approaches range from threshold policies and heuristics to predictive, hierarchical, and SLA-aware controllers.

  • Server consolidation and switching: Energy-efficient server-cluster approaches estimate required CPU frequency, determine the number of active nodes, and switch nodes according to workload thresholds.These methods combine DVFS with server power switching.
  • Server consolidation and switching: Energy-aware consolidation treats allocation as multi-dimensional bin packing, turning on a server and reallocating requests when existing capacity is insufficient.The objective is to reduce energy consumption while satisfying performance requirements.
  • Power allocation: Queueing models can determine power allocation that minimizes mean response time under factors such as arrival rate, frequency relationships, and peak power budgets.The approach applies to heterogeneous server farms.
  • Geographic scheduling: Geographically distributed Cloud scheduling can combine DVFS with workload placement to minimize energy and carbon emissions while maximizing provider profit.The surveyed approach uses five heuristics for HPC applications.
  • Coordinated management: Coordinated power-management systems combine local and global policies to manage guest VMs, physical machines, and rack- or blade-level requirements.Some systems dynamically re-provision power across machines to meet a power budget.
  • Predictive and adaptive control: Predictive and adaptive controllers use simulation-based learning, limited look-ahead control, Kalman filtering, proactive placement, and reactive migration.These methods target power reduction while limiting performance loss.

About the Authors

The authors are affiliated with research institutions in Australia and work across cloud computing, distributed systems, scheduling, and high-performance computing. Their backgrounds include research, publication, and commercialization experience.

  • Author biographies: Anton Beloglazov is a PhD candidate at the CLOUDS Laboratory at the University of Melbourne researching energy- and performance-efficient computing.He previously studied informatics and computer science at Novosibirsk State Technical University.
  • Author biographies: Rajkumar Buyya is a professor and CLOUDS Laboratory director at the University of Melbourne who has authored more than 300 research papers and four textbooks.He is also the founding CEO of Manjrasoft Pty Ltd.
  • Author biographies: Young Choon Lee researches scheduling in heterogeneous computing systems, nature-inspired techniques, and parallel and distributed algorithms.He received a PhD in problem-centric scheduling from the University of Sydney.
  • Author biographies: Albert Y. Zomaya is a chair professor of high-performance computing and networking at the University of Sydney with more than 350 publications.He has also authored or co-authored seven books and edited multiple volumes.
Loading 1007.0066v2…