UNIVERSITY OF CALIFORNIA, IRVINE Cooperative CPU-GPU Dynamic Power Management Methodologies for Energy-efficient Mobile Gaming DISSERTATION submitted in partial satisfaction of the requirements for the degree of DOCTOR OF PHILOSOPHY in Computer Science by Jurn-Gyu Park Dissertation Committee: Professor Nikil Dutt, Chair Professor Alex Nicolau Professor Mohammad Al Faruque 2017 Portion of Chapter 2 c 2014 IEEE All other materials c 2017 Jurn-Gyu Park DEDICATION To my God The LORD is my shepherd, I shall not be in want. He makes me lie down in green pastures, he leads me beside quiet waters, he restores my soul. He guides me in paths of righteousness for his name's sake. Even though I walk through the valley of the shadow of death, I will fear no evil, for you are with me; your rod and your staff, they comfort me. You prepare a table before me in the presence of my enemies. You anoint my head with oil; my cup overflows. Surely goodness and love will follow me all the days of my life, and I will dwell in the house of the LORD forever. - Psalms 23 - To my Wife A wife of noble character who can find? She is worth far more than rubies. - Proverbs 31:10 - ii TABLE OF CONTENTS Page LIST OF FIGURES vi LIST OF TABLES viii ACKNOWLEDGMENTS ix CURRICULUM VITAE x ABSTRACT OF THE DISSERTATION xii 1 Introduction 1 1.1 Smartphone Systems . .1 1.2 Motivation and Challenges . .4 1.2.1 Existing Approaches . .5 1.3 Thesis Overview . .6 1.3.1 Graphics Workload Characterization (GWC) . .6 1.3.2 Cooperative Frequency-Capping Governor (Co-Cap) . .7 1.3.3 Hierarchical-FSM based Dynamic Behavior Modeling Governor (HiCAP)8 1.3.4 Machine Learning enhanced Modeling Governor (ML-Gov) . .9 1.4 Thesis Contributions . .9 1.4.1 Design and Implementation of micro-benchmarks . .9 1.4.2 Simple but highly effective CPU-GPU Frequency Capping . 10 1.4.3 Hierarchical FSM-based Dynamic Behavior Modeling . 10 1.4.4 Machine Learning enhanced Simple and Accurate Prediction Models . 11 2 Graphics Workload Characterization for DVFS Design 12 2.1 Introduction . 12 2.1.1 Contributions . 14 2.2 Motivation and Related Work . 15 2.2.1 Motivation . 15 2.2.2 Related Work . 16 2.3 Graphics Workload Characterization . 18 2.3.1 Mobile Graphics Pipeline . 18 2.3.2 Workload Characterization and Micro-benchmarks . 21 2.4 Experimental Setup and Results . 23 iii 2.4.1 Experimental Setup and Methodology . 23 2.4.2 Experimental Results . 25 2.5 Opportunities for DVFS design . 34 2.5.1 GPU DVFS . 35 2.5.2 CPU DVFS . 36 2.5.3 Integrated DVFS . 37 2.6 Summary . 37 3 Cooperative CPU-GPU Frequency Capping 39 3.1 Introduction . 39 3.2 Related Work . 42 3.3 Motivation . 44 3.4 Co-Cap Methodology . 46 3.4.1 Training Phase . 47 3.4.2 Deployment Phase . 57 3.5 Evaluation of Co-Cap . 58 3.5.1 Experimental Setup . 58 3.5.2 Experimental Results . 63 3.5.3 Analysis and Discussion . 66 3.6 Conclusion . 69 4 Hierarchical FSM-based Integrated CPU-GPU Frequency Capping 70 4.1 Introduction . 70 4.2 Related Work . 73 4.3 Approach . 74 4.3.1 Preliminaries . 74 4.3.2 HFSM-based Dynamic Behavior Model . 76 4.3.3 Frequency-Capping . 79 4.4 Experimental Results . 81 4.4.1 Experimental Setup . 81 4.4.2 Automatic Measurement Tool . 82 4.4.3 Results and Analysis . 84 4.5 Conclusion . 88 5 A Machine Learning Enhanced Integrated Governor 89 5.1 Introduction . 89 5.2 Motivation and Related Work . 91 5.2.1 Motivation . 92 5.2.2 Related Work . 94 5.3 ML-Gov Methodology . 96 5.3.1 Learning Phase . 96 5.3.2 Prediction Phase . 104 5.4 Experimental Results . 108 5.4.1 Experimental Setup . 108 5.4.2 Results and Analysis . 111 iv 5.4.3 Discussion . 114 5.5 Conclusion . 115 6 Conclusion and Future Directions 117 6.1 Summary . 117 6.2 Contributions . 118 6.3 Future Directions . 120 Bibliography 121 v LIST OF FIGURES Page 1.1 Trends of Smartphone Applications . .1 1.2 Change of High-performance Mobile Integrated GPUs . .2 1.3 Change of Apple Smartphone Platforms [29] . .2 1.4 Correlation between Performance and Power [7] (The performance metric is Drystone MIPS and power consumption unit is mW) . .3 1.5 Comparison of CPU and GPU power consumption [5]: x-axis: time in second, y-axis: Power (mW) . .3 1.6 State-of-the-art Approaches . .5 1.7 Thesis Overview . .7 2.1 Mobile GPU Power Consumption [5] . 13 2.2 Motivating Example for GPU Workload Characterization. 15 2.3 Abstract Mobile Graphics Pipeline. 18 2.4 Logical Graphics Rendering Pipeline . 20 2.5 Power Measurement and Adreno Profiler Setup . 24 2.6 Results of mb-TexM at Different Frequencies with Workload Variation . 26 2.7 Results of mb-VerM at Different Frequencies with Workload Variation . 29 2.8 Workload Factor result with Enabled CPU DVFS on mb-VerM . 30 2.9 Results of mb-App at Different Frequencies with Workload Variation . 31 2.10 FPS of mb-VerSh and mb-FragSh ........................ 32 2.11 Average Power of mb-VerSh and mb-FragSh .................. 33 2.12 Energy per Frame of mb-VerSh and mb-FragSh ................ 33 3.1 System Comparison . 40 3.2 Mobile Platform Trends. 40 3.3 Different types of CPU/GPU Workload. 41 3.4 Motivating Examples . 45 3.5 Co-Cap Overview. 47 3.6 Co-Cap Training Phase. 47 3.7 A Sample of the Training Set. 49 3.8 Effects of CPU (or GPU) maximum frequency capping on FPS and Power. 49 3.9 The Proposed Methodology for Data Collection and LUT Building. 51 3.10 LUTs after the Building Step . 52 3.11 Fine-grained Refinement Step. 53 3.12 The Methodology of Fine-grained Refinement Step. 53 vi 3.13 Detailed Fine-grained Refinement Step. 54 3.14 Final LUTs after the Refinement Steps (∗frequencies are refined during the refinement step 2) . 56 3.15 The Evaluation of the Fine-grained Refinement Steps . 56 3.16 Co-Cap Deployment Phase. 57 3.17 The Training Sets. 59 3.18 The Deployment Sets. 60 3.19 FPS, Power and EpF Results of Different Types of Graphics Workloads. 62 3.20 The Average Results of the Training Set (High-variation). 64 3.21 The Detailed Results of the Training Set (High-variation). 64 3.22 Average Results of the Deployment Sets . 65 3.23 The results of the Deployment Set (MB). 66 3.24 The results of the Deployment Set (RG). 66 4.1 CPU-GPU Mobile Governors . 71 4.2 HFSMs in Game Design. 71 4.3 Sample Mobile Game (ShootEmDown) . 72 4.4 A set of benchmarks. ..