UNIVERSITY OF CALIFORNIA, SAN DIEGO Accurate Temperature Sensing and E�cient Dynamic Thermal Management in MPSoCs A dissertation submitted in partial satisfaction of the requirements for the degree Doctor of Philosophy in Computer Science (Computer Engineering) by Shervin Sharifi Committee in charge: Professor Tajana Simunic Rosing, Chair Professor Chung-Kuan Cheng Professor Tara Javidi Professor Ryan Kastner Professor Joseph Pasquale 2011 Copyright Shervin Sharifi, 2011 All rights reserved. The dissertation of Shervin Sharifi is approved, and it is acceptable in quality and form for publication on micro- film and electronically: Chair University of California, San Diego 2011 iii DEDICATION To those who have dedicated their lives to me, Mom and Dad. And to my dearest Avisha. iv TABLE OF CONTENTS Signature Page . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii Dedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv Table of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . x Vita . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xii Abstract of the Dissertation . . . . . . . . . . . . . . . . . . . . . . . . . . . xv Chapter 1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.1 Thesis Contributions . . . . . . . . . . . . . . . . . . . . 8 Chapter 2 Direct temperature sensing . . . . . . . . . . . . . . . . . . . . 12 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 12 2.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . 14 2.3 Analytical Model for Upper Bound of On-Chip Temper- ature Di↵erences . . . . . . . . . . . . . . . . . . . . . . 16 2.4 Thermal Sensor Placement . . . . . . . . . . . . . . . . . 21 2.5 Experimental Results . . . . . . . . . . . . . . . . . . . . 23 2.5.1 Maximum Temperature Di↵erence Model . . . . . 23 2.5.2 Sensor Placement . . . . . . . . . . . . . . . . . . 26 2.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . 29 Chapter 3 Indirect Temperature Sensing . . . . . . . . . . . . . . . . . . 33 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 33 3.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . 34 3.3 Components of Indirect Temperature Sensing . . . . . . . 36 3.3.1 KF-based Temperature Estimation . . . . . . . . 37 3.3.2 Reducing Computational Complexity . . . . . . . 40 3.3.3 Detecting Sensor Failure and Degradation . . . . 42 3.4 Experimental results . . . . . . . . . . . . . . . . . . . . 45 3.4.1 Indirect temperature sensing . . . . . . . . . . . 46 3.4.2 Detecting sensor failure and degradation . . . . . 51 3.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . 53 v Chapter 4 Tempo Temperature Prediction . . . . . . . . . . . . . . . . . 57 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 57 4.2 Related work . . . . . . . . . . . . . . . . . . . . . . . . 58 4.3 Temperature Prediction . . . . . . . . . . . . . . . . . . . 60 4.3.1 Theoretical Analysis of Tempo . . . . . . . . . . . 66 4.4 Experimental results . . . . . . . . . . . . . . . . . . . . 71 4.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . 74 Chapter 5 PROMETHEUS Framework for Temperature-aware Schedul- ing on Heterogeneous MPSoCs . . . . . . . . . . . . . . . . . 76 5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 76 5.2 Related work . . . . . . . . . . . . . . . . . . . . . . . . 78 5.3 PROMETHEUS Scheduling Framework . . . . . . . . . . 80 5.3.1 Power state assignment in TempoMP . . . . . . . 82 5.3.2 Power state assignment in TemPrompt . . . . . 87 5.3.3 Runtime task assignment to the cores . . . . . . . 89 5.4 Experimental results . . . . . . . . . . . . . . . . . . . . 90 5.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . 95 Chapter 6 Conclusion and Future Work . . . . . . . . . . . . . . . . . . . 97 6.1 Thesis Summary . . . . . . . . . . . . . . . . . . . . . . 98 6.1.1 Analytical Model for Upper Bound on On-chip Spatial Thermal Gradients . . . . . . . . . . . . . 98 6.1.2 Accurate Direct Temperature Sensing . . . . . . . 99 6.1.3 Accurate Indirect Temperature Sensing . . . . . . 100 6.1.4 Tempo Temperature Prediction . . . . . . . . . . 100 6.1.5 PROMETHEUS FrameworkforTemperature-aware Scheduling in Heterogeneous MPSoCs . . . . . . . 101 6.2 Future Research Directions . . . . . . . . . . . . . . . . . 102 6.2.1 Thermal Management in Heterogeneous MPSoCs with Special Purpose Cores . . . . . . . . . . . . 102 6.2.2 Thermal Management in Many-core MPSoCs . . . 103 Appendix A Compact Thermal Modeling . . . . . . . . . . . . . . . . . . . 105 A.1 Electrical Representation of Heat Transfer . . . . . . . . 105 A.2 Extracting the Parameters of the Thermal Network . . . 107 Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110 vi LIST OF FIGURES Figure 1.1: Scaling of (a) Transistor integration capacity (2) Frequency, Vdd and power [12] . . . . . . . . . . . . . . . . . . . . . . . . 2 Figure 1.2: Distibution of (a) power density vs. (b) temperature across a chip [67] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 Figure 2.1: Contour map of maximum temperature di↵erence to a point of interest . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 Figure 2.2: Algorithm for calculating the upper bounds . . . . . . . . . . . 20 Figure 2.3: Layout of SoC2 . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 Figure 2.4: Temperature di↵erence between points a and b . . . . . . . . . 25 Figure 2.5: Generating the maximum temperature di↵erence by construct- ing the proper power trace . . . . . . . . . . . . . . . . . . . . . 26 Figure 2.6: Temperature di↵erence in the example of figure 2.5 . . . . . . . 27 Figure 2.7: Using observability area vs. circular range . . . . . . . . . . . 28 Figure 3.1: Proposedtechnique. (a)O✏inesetup(b)Runtimetemperature estimation by KF . . . . . . . . . . . . . . . . . . . . . . . . . . 39 Figure 3.2: Comparison of sensor, actual and estimated temperatures . . . 52 Figure 3.3: Run time of the technique on (a) XScale R (b) SPARCR . . . 54 � � Figure 3.4: SPRT technique to detect sensor degradation . . . . . . . . . . 55 Figure 4.1: (a) Temperature of the core (b) Breakdown of temperature into components of equation (4.7) (c) Temperature of corresponding nodesinthermalinterfacematerial,heatspreaderandheatsink, all relative to ambient . . . . . . . . . . . . . . . . . . . . . . . 62 Figure 4.2: Gershgorin discs of matrix �t in complex plane for (a) a high s end package and (b) an embedded-type package . . . . . . . . 68 Figure 4.3: Characteristics of the MPSoC . . . . . . . . . . . . . . . . . . 72 Figure 4.4: Comparison of Tempo and BLP predictor [10] . . . . . . . . . 73 Figure 5.1: Scheduling system in PROMETHEUS . . . . . . . . . . . . . . 80 Figure 5.2: O✏ine stage of TempoMP . . . . . . . . . . . . . . . . . . . . . 82 Figure 5.3: A very simple example describing use of multi-parametric pro- gramming in power state assignment . . . . . . . . . . . . . . . 84 Figure 5.4: Comparison of Maximum Temperature . . . . . . . . . . . . . 91 Figure 5.5: Average lateness (seconds) . . . . . . . . . . . . . . . . . . . . . 92 Figure 5.6: Throughput (million instructions executed per second) . . . . . 93 Figure 5.7: Average power consumption . . . . . . . . . . . . . . . . . . . . 93 Figure 5.8: Average energy per billion instructions executed . . . . . . . . . 95 Figure 5.9: Average Energy Lateness Product (ELP) . . . . . . . . . . . . . 95 vii Figure A.1: An example of a chip and package, together with their corre- sponding thermal RC network . . . . . . . . . . . . . . . . . . 108 viii LIST OF TABLES Table 1.1: Examples of di↵erent classes of thermal management techniques 3 Table 2.1: Errors in temperature di↵erence simulations (oC) . . . . . . . . 28 Table 2.2: Number of sensors needed by our technique and range-based methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 Table 2.3: Error statistics for limited number of sensors . . . . . . . . . . . 30 Table 2.4: Error statistics for di↵erent time steps . . . . . . . . . . . . . . . 31 Table 3.1: E↵ectofnumberofmatchedmomentsontemperatureestimation Error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 Table 3.2: E↵ects of sensor degradation and failure . . . . . . . . . . . . . 49 Table A.1: Duality between Thermal and Electrical Quantities . . . . . . . 107 ix ACKNOWLEDGMENTS The work presented in this thesis would not have been possible without the help and support of wonderful people that I have had the privilege of interacting with during my years at UC San Diego. First and foremost, I would like to thank my advisor, Professor Tajana SimunicRosingforherguidance, supportandthemanylessonsIhavelearnedfrom her. I especially thank her for her understanding, support and patience during the di�cult times I had. Without her help, I would not have been able to overcome many technical and non-technical challenges. In addition, I really appreciate the supportive and friendly environment she has created in our research group. I really appreciate the e↵ort and the time my thesis committee members, Professor Chung-Kuan Cheng, Professor Tara Javidi, Professor Ryan Kastner and Professor Joseph Pasquale have taken to review my manuscript and conduct my defense. I have also been very lucky to have wonderful colleagues and friends in our department and in our research group. I wish to thank them all for providing a friendly and joyful environment and for their valuable comments and discussions. Special thanks to Raid Ayoub, Yen-Kuan Wu, Ayse Coskun, Gaurav Dhiman, EdoardoRegini, PritiAghera, NimaNikzad, BryanKim, YasharAsgarieh, Richard Strong, Aruna Ravinagarajan, Jamie Bradley Steck and Giacomo Marchetti. I had the privilege to collaborate with Raid Ayoub, Ayse Coskun, Dilip Krishnaswamy and Chun-Chen Liu. I have been very fortunate to have a wonderful group of friends in San Diego. I sincerely thank them for their friendship, support and being my family when I was far from my own family. I would like to especially thank Kambiz Samadi, Amirali Shayan, Ahsan Samiee, Kiarash Kiantaj, Hamed Movahedpour, Ehsan Ardestanizadeh, Haleh Azartash, Behrokh Farzad and Setareh Setayesh among the others. We have shared many memorable moments which I will never forget. My utmost and deepest gratitude, a↵ection and love belong to my family, especially my parents for their unconditional and endless love and support. I am x