UNIVERSITY OF CALIFORNIA, SAN DIEGO Building Aggressively Duty-Cycled Platforms to Achieve Energy Efficiency A dissertation submitted in partial satisfaction of the requirements for the degree Doctor of Philosophy in Computer Science (Computer Engineering) by Yuvraj Agarwal Committee in charge: Rajesh Gupta, Chair Paramvir Bahl William Hodgkiss Stefan Savage Alex Snoeren Geoff Voelker 2009 Copyright Yuvraj Agarwal, 2009 All rights reserved. The dissertation of Yuvraj Agarwal is approved, and it is acceptable in quality and form for publication on mi- crofilm and electronically: Chair University of California, San Diego 2009 iii DEDICATION To Dadi, Shyam Babaji, Papa and Ma. iv TABLE OF CONTENTS Signature Page .................................. iii Dedication ..................................... iv Table of Contents ................................. v List of Figures .................................. ix List of Tables ................................... xi Acknowledgements ................................ xii Vita and Publications .............................. xv Abstract of the Dissertation ........................... xvi Chapter 1 Introduction ............................ 1 1.1 Reducing the Energy Consumption of Computing Devices 3 1.2 Using Collaboration to Aggressively Duty-Cycle Platforms 5 1.3 Contributions ........................ 7 1.4 Organization ......................... 8 Chapter 2 Background and Related Work .................. 9 2.1 Mobile Platforms ...................... 9 2.2 Power and Energy ...................... 12 2.2.1 Measuring Power and Energy Consumption .... 13 2.3 Related Work ........................ 14 2.3.1 Power Management in Mobile Devices ....... 14 2.3.2 Power Management in Laptops and Desktop PCs 17 Chapter 3 Radio Collaboration - Cellular and LAN Data Radios ..... 20 3.1 Overview of a VoIP Deployment .............. 23 3.2 Alternatives to VoIP over Wi-Fi Radios .......... 24 3.2.1 Cellular Data vs. Wi-Fi .............. 25 3.2.2 Smartphone Power Measurements ......... 28 3.3 Cell2Notify Architecture .................. 30 3.3.1 Cell2Notify Protocol ................ 32 3.3.2 Connectivity Scenarios ............... 33 3.3.3 Modifications to the VoIP Server ......... 35 3.3.4 Modifications to the Smartphone ......... 36 3.3.5 Other Applications ................. 38 3.3.6 Alternatives to Cell2Notify ............. 38 v 3.4 Implementation ....................... 39 3.4.1 Prototype Cell2Notify Server ............ 40 3.4.2 Prototype Cell2Notify Client ............ 43 3.5 Evaluation .......................... 45 3.5.1 Reduction in Energy Consumption ........ 45 3.5.2 End-to-End Latency ................ 50 3.6 Discussion .......................... 53 3.6.1 Is Caller-ID Spoofing Legal? ............ 54 3.6.2 Handling Spoofed Caller-IDs ............ 54 3.6.3 Concerns of Cellular Operators .......... 55 3.6.4 Deploying Cell2Notify ............... 55 3.7 Related Work – Paging and Wakeup ............ 56 3.8 Summary .......................... 58 Chapter 4 Building a Switching Hierarchy using Collaborative Data Radios 59 4.1 CoolSpots Architecture ................... 60 4.2 Switching Policies ...................... 63 4.2.1 Switching Framework ................ 64 4.2.2 Baseline Policies ................... 65 4.2.3 Bandwidth Policy .................. 66 4.2.4 Cap-Static Policy .................. 66 4.2.5 Cap-Dynamic Policy ................ 67 4.3 Benchmarks ......................... 68 4.3.1 Baseline Benchmarks ................ 69 4.3.2 Streaming Benchmarks ............... 70 4.3.3 Web Traffic Benchmarks .............. 71 4.4 Experimental Setup ..................... 71 4.4.1 Hardware Specifications .............. 73 4.4.2 Energy Measurement ................ 73 4.4.3 Location Configuration ............... 74 4.5 Evaluation .......................... 75 4.5.1 Characterizing Radio Switching .......... 76 4.5.2 Energy Savings for Individual Benchmarks .... 78 4.5.3 Effect of Radio Ranges and Location ....... 79 4.5.4 Discussion ...................... 80 4.6 Summary .......................... 81 Chapter 5 Deploying a Collaborative Radio Infrastructure ........ 83 5.1 SwitchR Architecture .................... 84 5.1.1 Separating the Wi-Fi AP and the Bluetooth Gate- way .......................... 85 5.1.2 Handling Multiple Clients ............. 86 5.2 Switching Mechanism .................... 87 vi 5.2.1 Switching from Wi-Fi to BT ............ 87 5.2.2 Switching from BT to Wi-Fi ............ 88 5.2.3 Handling Mobility .................. 88 5.2.4 Baseline Switching Analysis ............ 89 5.3 Switching Policies ...................... 89 5.3.1 Baseline Policies ................... 91 5.3.2 Cap-Dynamic Policy ................ 91 5.3.3 Multi-Client Policy ................. 92 5.4 Benchmarks ......................... 94 5.4.1 Idle and Transfer .................. 94 5.4.2 Streaming ...................... 95 5.4.3 Web Traffic ..................... 95 5.5 Experimental Setup ..................... 96 5.5.1 Energy Measurement ................ 97 5.5.2 Experimental Design ................ 98 5.6 Evaluation .......................... 99 5.6.1 Media Streaming Applications ...........101 5.7 Related Work – Radio Collaboration ...........105 5.8 Summary ..........................106 Chapter 6 Processor Collaboration - Energy Saving for PCs .......108 6.1 Somniloquy Architecture ..................110 6.1.1 Supporting Stateless Applications: Wakeup Filters 113 6.1.2 Supporting Stateful Applications: Stubs . 114 6.1.3 Quantifying Energy Savings ............117 6.2 Prototype Implementation .................118 6.2.1 Hardware and Software Overview .........118 6.2.2 Three different prototypes .............121 6.2.3 Applications Without Stubs ............123 6.2.4 Applications Using Stubs ..............123 6.3 Evaluation ..........................125 6.3.1 Microbenchmarks – Power, Latency ........126 6.3.2 Somniloquy in Operation ..............128 6.3.3 Application Performance ..............130 6.3.4 Energy Savings using Somniloquy .........134 6.4 Discussion ..........................138 6.4.1 Handling Security Implications ...........138 6.4.2 Alternative Design of Somniloquy .........139 6.5 Related Work – Processor Collaboration .........140 6.6 Summary ..........................143 vii Chapter 7 Conclusions ............................145 7.1 Contributions ........................146 7.1.1 Improving Duty-Cycling using Collaboration . 146 7.1.2 Deployable Prototypes ...............147 7.1.3 Evaluation Results, Platforms and Power Mea- surements ......................148 7.2 Future Work .........................149 7.3 Deploying Somniloquy in Enterprises ...........149 7.3.1 Extending Processor Collaboration ........149 7.3.2 Detailed Energy Accounting by In-situ Measure- ments ........................150 Bibliography ...................................152 viii LIST OF FIGURES Figure 1.1: Power consumption of a Intel-PXA255 processor using slow- down and shutdown. ........................ 5 Figure 2.1: Power consumption of the Stargate[21] mobile platform ..... 10 Figure 2.2: Power consumption of the HTC Tornado smartphone platform . 11 Figure 2.3: Power measurement methodology. ................ 13 Figure 3.1: A typical enterprise VoIP deployment. Incoming calls to the SIP server can be received over the IP network or over the PSTN line. An Analog Telephony Adapter (ATA) acts as a bridge between PTSN and IP networks. ................. 23 Figure 3.2: Power measurements of 1xEVDO, GPRS/EDGE and Wi-Fi in- terfaces for different scenarios. The “Connected and Active” measurements show the power when transmitting 32 Kbps of VoIP traffic over UDP. Note that when active, VoIP over Wi-Fi consumes the least amount of battery power. .......... 26 Figure 3.3: Smartphone power measurement setup .............. 28 Figure 3.4: Steps of the Cell2Notify protocol. ................ 31 Figure 3.5: Our prototype implementation of Cell2Notify. We implement the Cell2Notify server as a combination of a commonly available SIP proxy and an Internet- based VoIP gateway. We emulate a smartphone using a combination of a cellphone that communi- cates with a Wi-Fi equipped laptop using Bluetooth. ...... 41 Figure 3.6: Call logs of three different users. ................. 47 Figure 3.7: Energy consumption using two cards, with and without Cell2Notify for three different users. As expected Cell2Notify saves more energy for lighter usage patterns. ................. 48 Figure 3.8: Energy consumption of a Cingular 2125 with and without Cell2Notify for three users. We assume that the user does not use the smart- phone for any other purpose, but only for making and receiving VoIP calls. ............................. 49 Figure 3.9: Breakdown of various steps of the Cell2Notify protocol in call- setup latency. The right bar shows the expected latency with our proposed optimizations. Even without optimizations, the extra delay is around ten seconds, which is less than two rings. 51 Figure 4.1: Multiple Bluetooth-enabled CoolSpots, inside of a traditional Wi-Fi HotSpot, allow mobile devices to connect other devices through the backbone network. CoolSpots are connected to the backbone network either directly (wired) or through the Wi-Fi network (wireless). ......................... 60 ix Figure 4.2: CoolSpots enables enables radio collaboration on top of individ- ual radio power management techniques such as Bluetooth sniff mode and Wi-Fi PSM. ....................... 61 Figure 4.3: Experimental Setup. The Test Machine (TM) and the Base Sta- tion (BS) are on a cart which can be moved around to