1 / 31

Energy Scale-Down in System Design: Optimizations for Reducing Power

This talk focuses on energy scale-down for mobile devices and explores optimizations to reduce power consumption in display, processor, and wireless components.

evalerie
Download Presentation

Energy Scale-Down in System Design: Optimizations for Reducing Power

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Energy Scale-Down in System Design:Optimizations for Reducing Power Parthasarathy (Partha) Ranganathan (with Bob Mayo) Hewlett Packard Labs July 3, 2003 Partha Ranganathan E-scale project, HP Labs

  2. Broader context • Energy scale-down one component of broader power management work • This talk will focus on scale-down for mobile devices Power and energy management Enterprise systems Power costs & cooling Mobile systems Battery life display wireless CPU Partha Ranganathan E-scale project, HP Labs

  3. Energy Scale-Down: Motivation Mismatched system energy efficiency & desired functionality • Tethered system (performance) hangover… • Increased performance at any cost, target worst-case benchmark • Non-peak benchmarks consume more energy than needed • Optimizations where energy costs outweigh small performance benefits • User preference for convergence of diverse mobile devices • Combination of diff. needs => general-purpose designs (e.g. phone/PDA) • Individual tasks consume more energy than needed Do you need the full display to say three words: “you have mail”? Do you need your wireless to respond within 100ms for email? Do you need a 466 MHz processor for idle mode? for MS Word? Solution: energy scale-down design adaptivity to optimize energy efficiency based on task requirements Partha Ranganathan E-scale project, HP Labs

  4. Talk Roadmap • Motivation • Quantifying energy costs of inefficiencies • Scale-down optimizations to reduce energy Display scale-down Processor scale-down Wireless scale-down Ongoing work and summary Partha Ranganathan E-scale project, HP Labs

  5. Quantifying Energy Costs of Inefficiencies • Mismatched system energy efficiency and task functionality • What is the “optimal” energy needed for a task? But, optimal energy consumption of task a challenging problem • Past work “lower is better”, but no limits • Hard-to-define target – fidelity, performance, costs, engineering • Our approach: use surrogate lower-bounds • Special-purpose devices optimized for particular task • Representative successful tradeoffs in functionality and battery-life Partha Ranganathan E-scale project, HP Labs

  6. Experimental methodology • Energy comparison for a spectrum of mobile devices • First such study to perform a consistent comparison • Devices: • Laptop (Armada M300), PDA (iPAQ 3630) • Cell phone (Nokia 8260), Pager (Blackberry W1000), High-end MP3 (Nomad jukebox), low-end MP3 (ipaq PA1), voice-recorder (VoiceItVT90) • Benchmarks representative of typical mobile workloads • Email, text messaging, phone calls, web browsing • MP3 play-back, text notes, audio notes, games, idle mode • Benchmarks structured to have core functionality consistent • Measurement – data acquisition of current/voltage • Total energy for task • Temporal power signatures Partha Ranganathan E-scale project, HP Labs

  7. Energy Scale-down Results: Energy Comparison for Email • Email: • Laptop: 165X • Handheld: 15X • Cell phone: 6X • RIM pager: 92 mW • Radio wakeup 100ms (iPAQ) 1.2 sec (cell) 5 sec (RIM) Partha Ranganathan E-scale project, HP Labs

  8. Wide variation in power 950% to 22,000% for similar task functions iPAQ 5X-10X higher energy Laptop 10X-100X higher energy Variations related to better task-specific component matching Significant potential from addressing energy inefficiencies Overall results Partha Ranganathan E-scale project, HP Labs

  9. e Energy scale-down • Addressing general-purpose energy inefficiencies • Energy scale-down • Design and use adaptivity in hardware and software to scale-down energy based on task requirements • An informal taxonomy • Scale-down mechanism • Gradation-based: same component, multiple modes • Examples : v/f scaling, gating, memory states, disk states, OLED-based displays, protocol-level wireless optimizations, fidelity optimizations • Plurality-based: “the kitchen-sink approach!” • Examples: hierarchy of displays, plurality networks, heterogeneous chip multiprocessing • Scale-down impact: user-directed versus user-transparent Partha Ranganathan E-scale project, HP Labs

  10. Talk Roadmap • Motivation • Quantifying energy costs of inefficiencies • Scale-down optimizations to reduce energy Display scale-down Ongoing work and summary Partha Ranganathan E-scale project, HP Labs

  11. Display scale-down [Mobisys2003] • Displays consume significant power in mobile systems • 50% on laptops[7], 61% on handhelds[1] Previous approaches: • Turning off the entire display • Using lower quality or smaller sized displays Our approach: energy-adaptive display • Power consumption based on content being displayed • Understand user requirements • Design and evaluate example Partha Ranganathan E-scale project, HP Labs

  12. Characterizing user requirements • User study: understand usage behavior of 17 Windows users • Display capacities are not fully utilized • On average, ~60% of screen area used (window-of-focus) • Even smaller for some users • Other functions of display are not used always (color, res., …) Partha Ranganathan E-scale project, HP Labs

  13. Display property vs. usage mismatches • Mismatches occur because of user/application-specific window usage • Small: system-related messages and low-content windows • Large: development, web, and emails • But display power is constant all the time • Can we provide a means for energy to scale-down with lower usage? Partha Ranganathan E-scale project, HP Labs

  14. Laptops PDAs,Handhelds 3G Phones,Automotive DigitalCamera&Camcorders Energy-adaptive display systems • Hardware support for power control at finer granularity • Leverage emerging OLED technologies • Pixel power based on pixel value (brightness, color) • Currently in cell phones, expected in handhelds/laptops 2004-5 Partha Ranganathan E-scale project, HP Labs

  15. Software support: energy-aware user interfaces (DarkWindows) Approximate user interest to window of focus Automatic power-aware adaptation of background brightness/color Energy-aware user interfaces Partha Ranganathan E-scale project, HP Labs

  16. Xvnc X protocol VNC protocol Applications VNC Server VNC Viewer TrackFocus Window Change pixel values in framebuffer Xvnc VNC Viewer Original Framebuffer Modified Framebuffer Evaluation methodology • Prototype user interface using VNC under Linux • OLED power model for representative user trace Display Power = Pcontroller + Pdriver + Panel Power Panel Power = Pixel Array Power = ∑ Pred x pixelR + Pgreen x pixelG + Pblue x pixelB Pred = 4.3 µW, Pgreen = 2.3 µW, Pblue = 4.3 µW Partha Ranganathan E-scale project, HP Labs

  17. Benefits from energy adaptivity Partha Ranganathan E-scale project, HP Labs

  18. Power benefits from different interfaces Benefits from both hardware and software Broad acceptance of user interfaces in user study Power savings Partha Ranganathan E-scale project, HP Labs

  19. Energy savings function of user preference Power savings: sensitivity experiments Partha Ranganathan E-scale project, HP Labs

  20. Other energy-adaptive designs • Hardware adaptability • Emissive displays • Hybrid technologies • Multi-display configuration • Other output modes Software adaptability • “Flashlight” or “headlight” cursor • “Sticky lamps” on desktop • Application-specific dimming Partha Ranganathan E-scale project, HP Labs

  21. Display scale-down: Summary • Display component a large fraction of total power • First detailed user study on screen usage behavior • Only fraction of screen area used • Many properties of display (color, resolution) often not used • Energy-adaptive display design • Hardware support for fine-grained power control • Software support for energy-aware user interfaces • Significant power benefits with low user intrusiveness Partha Ranganathan E-scale project, HP Labs

  22. Talk Roadmap • Motivation • Quantifying energy costs of inefficiencies • Scale-down optimizations to reduce energy Display scale-down Processor scale-down Ongoing work and summary Partha Ranganathan E-scale project, HP Labs

  23. Processor Scale-down [MICRO2003] • Motivation: CPU power important component of total power • Previous approaches • Voltage and frequency scaling limited by feature size • Architectural adaptation limited to dynamic power • Our Solution: Heterogeneous Multi-core Single-ISA Architecture • Have multiple heterogeneous cores on the same die • Match workload to core with best energy efficiency • Power down the unused cores Partha Ranganathan E-scale project, HP Labs

  24. Characterizing workload behavior Methodology • Simulation study of 14 SPEC2000 benchmarks • Five-core CPU (MIPS R4K, EV4, EV5, EV6, EV8-) Mismatch between energy efficiency and workload requirement Core difference varies based on workload or workload phases (IPS) Varying core energy efficiencies for the same workload (IPS/W) Partha Ranganathan E-scale project, HP Labs

  25. Oracle-choose best energy efficiency 39% average energy savings with 3% performance loss 2X-4X benefits in half the benchmarks Oracle-choose best energy-delay 75% average energy savings with 24% performance loss 2X-11X benefits in all benchmarks Significantly better than voltage/frequency scaling Realistic heuristics within 90% of oracle switching Power benefits Partha Ranganathan E-scale project, HP Labs

  26. CPU scale-down: Summary • Using scale-down to address processor power • Simulation study characterizing energy efficiency mismatch • Heterogeneous single-ISA CMP architecture • Significant power benefits • Better than voltage/frequency scaling • Ongoing work • Other heuristics • Other architectures • Less diversity, energy-accentuated diversity • Implications on performance • Area vs. throughput Partha Ranganathan E-scale project, HP Labs

  27. Talk Roadmap • Motivation • Quantifying energy costs of inefficiencies • Scale-down optimizations to reduce energy Display scale-down Processor scale-down Other work and summary Partha Ranganathan E-scale project, HP Labs

  28. Listen Interval Other work: Wireless scale-down • Motivation: wireless component of power • Many workloads spend most power “listening” • E.g., email, phone calls, SMS messages, conferencing • Idle power 89% of total wireless power • Our approach: scale-down for idle-mode power management • Expose application requirements to physical layer • Change “listen interval” parameters for 802.11 • Power benefits • Changing power interval to 1sec: 20% power benefits • Changing listen interval to 1min: 90% power benefits Partha Ranganathan E-scale project, HP Labs

  29. Other work: Enterprise scale-down Energy scale-down adaptivity to optimize energy efficiency based on task requirements • Inefficiencies from designing for peak-performance needs • Inefficiencies from designing for peak-tolerance needs • Inefficiencies from aggregation of components • Inefficiencies from modularity of functions • Inefficiencies from not addressing total costs of ownership • Inefficiencies from inadequate automation • Preliminary results promising Partha Ranganathan E-scale project, HP Labs

  30. Summary • Energy and power important considerations for future systems • Significant mismatches in energy efficiency and task functions • Quantification energy costs of inefficiencies • First study to perform consistent comparison of spectrum of devices • Special-purpose devices 5X-100X better than general-purpose devices • Good surrogate-bounds and best-practices for energy optimizations • Scale-down: adaptivity to optimize efficiency based on requirements • Energy-adaptive displays: energy benefits with acceptable user interfaces • Heterogeneous CMPs: energy benefits with acceptable performance • Wireless scale-down: energy benefits with acceptable response delays • Critical to integrate energy scale-down in future designs Partha Ranganathan E-scale project, HP Labs

  31. More information • Relevant Papers • Energy consumption in mobile systems: why future systems need requirements-aware energy scale-down, Mayo and Ranganathan, HP Tech report, HPL TR2003-167 [Under review, IEEE Computer] • Energy-adaptive display system designs for future mobile environments, Iyer, Luo, Mayo and Ranganathan, Mobisys 2003 • Single-ISA Heterogeneous Multi-Core Architectures: The Potential for Power Reduction, Kumar, Farkas, Jouppi, Ranganathan, Tullsen, MICRO 2003, CAL2003 • Idle-Mode Power Management for Personal Wireless Devices, Abou-ghazala, Mayo and Ranganathan, HP Technical report HPL2003-102 Contact • http://web.hpl.hp.com/reserach/lss/projects/smartpower/ • Email: partha.ranganathan@hp.com Partha Ranganathan E-scale project, HP Labs

More Related