상세정보

Recently, the dramatic enhancement in performance of laptops and mobile devices have led to the explosive increase in their sales, which have outstripped that of desktop PCs. With the explosive growth in sales, power management of those devices is gaining ever increasing importance, as they always lack from limited battery life. In addition, thermal management has also become more im-portant, since it is difficult to aid from hardware cooling solutions, due to the limited area. To efficiently alleviate these problems, we propose three schemes: 1) Display power management scheme, 2) CPU power management scheme, and 3) Thermal management scheme.
Firstly, we tackle the essential weakness of the existing DPM (display power management) schemes that turns on or off the display based on user presence detection; regardless of the user presence, the display only needs to be turned on when the user is gazing at the display. Based on this proposition, we propose the user-aware display power management that uses a camera to determine if the user is actually gazing at the display. Our scheme captures pictures and detects user’s frontal face from the obtained pictures. When a frontal face is detected, our scheme determines that the user is gazing at the display. In our evaluation results, we show that our proposed scheme turns off the display for 52% of the total time, by detecting the user intent. Our scheme reduces average system-wide energy by 5%, compared to the most widely used timeout-based DPM, with ac-ceptable amount (less than once an hour) of user irritation.
Secondly, we investigate the CPU power management, which has become essential in recent mobile devices, where multi-core processors are adopted to handle CPU-intensive applications. Although there have been many studies to reduce the power consumption of mobile devices, there have not been sufficient efforts to consider the characteristics of the mobile applications. In this thesis, we introduce the limitations of the conventional power management schemes used in the mobile devices. Then, we propose an application-aware power man-agement scheme for multi-core based mobile devices that categorizes most exist-ing applications to either performance-sensitive application or QoS-sensitive ap-plication. We evaluate our scheme compared to conventional schemes, hotplug and ondemand, based on real measurements. In case of performance sensitive applications, our scheme saves energy (by average of 12% and 13% compared to hotplug and ondemand, respectively) and enhances the performance (by av-erage of 4.3% and 16% compared to hotplug and ondemand, respectively) at the same time for most applications. In case of QoS sensitive applications, our scheme saves 9.8% energy compared to hotplug and consumes similar energy compared to ondemand without a noticeable QoS degradation.
Lastly, we look into the thermal problems of recent mobile devices, which are caused by the adaptation of high performance processors. As the higher per-formance inevitably leads to power density increase, they eventually results in thermal problems. In order to alleviate the thermal problems, recent mobile de-vices scales down the voltage and frequency, when there are thermal emergen-cies. Unfortunately, this operation leads to an excessive number of voltage and frequency scaling, which not only worsen the performance but also the power efficiency. In this thesis, we describe how the conventional thermal management worsen power efficiency. To overcome the weakness of the conventional ther-mal management, we propose a temperature-aware DVFS (dynamic voltage and frequency scaling) scheme for Android-based mobile devices to optimize power or performance depending on the user-selectable option. We evaluate our scheme in the off-the-shelf mobile device. Our evaluation results show that by using the power optimizing option, our scheme saves energy consumption by 12.7%, on average, without any degradation in performance. By using the per-formance optimizing option, our scheme enhances the performance by 6.3%, on average, by using the performance optimizing scheme, still reducing the energy consumption by 6.7%.
There are no doubts that power management issues will become more severe as the industry continues introducing higher performance laptops and mobile devices, while there is no remarkable improvement in the battery. Furthermore, thermal management issues will also become more severe as the power density of the processors continues to increase with the advance in process technology. In this thesis, we propose applicable power and thermal management schemes that efficiently address the power and thermal problems.

Table of Contents

Abstract 	 i
Table of Contents 	 iv
List of Figures 	 viii
List of Tables 	 xi
1. Introduction 	 1
1.1 Power Management 	 1
1.1.1 Display Power Management 	 2
1.1.2 CPU Power Management 	 3
1.2 Thermal Management 	 6
2. Related Work 	 9
2.1 Display Power Management 	 9
2.1.1 Brightness and Color based Power Management 	 9
2.1.2 Power-off based Power Management 	 10
2.2 CPU Power Management 	 13
2.2.1 DVFS-based Power Management 	 13
2.2.2 CPU Shutdown-based Power Management	 14
2.2.3 Mobile Device Power Management	 15
2.3 Thermal Management 	 17
3. User-aware Display Power Management 	 19
3.1 Display Power Management that Detects User Intent 	 19
3.1.1 User State Definition 	 19
3.1.2 User State Transition 	 21
3.1.3 User Detection Algorithm 	 22
3.2 Evaluation Methodology 	 24
3.2.1 Determining Tlatency 	 24
3.2.2 Energy Consumption Model 	 26
3.2.3 Analytic Evaluation of User Irritation 	 28
3.2.4 User Log Collection Method 	 28
3.2.5 Evaluation Environment 	 31
3.3 Evaluation Results 	 32
3.3.1 User State Analysis 	 32
3.3.2 Energy Consumption 	 34
3.3.3 User Irritation 	 36
4. Application-aware CPU Power Management	 38
4.1 Motivation 	 38
4.1.1 Utilization Threshold 	 39
4.1.2 Sampling Period 	 41
4.2 Application-aware CPU Power Management 	 43
4.2.1 Application-aware DVFS 	 43
4.2.1.1 Performance-sensitive Application 	 45
4.2.1.2 QoS-sensitive Application 	 48
4.2.1.3 Indeterminate Application 	 50
4.2.1.4 Insignificant Application 	 51
4.2.1.5 Execution of Multiple Applications 	 51
4.2.2 Load Balancing-aware CPU Shutdown 	 54
4.3 Evaluation Methodology 	 57
4.3.1 Evaluation Methodology for DVFS Policy 	 58
4.3.2 Evaluation Methodology for CPU Shutdown Scheme 	 61
4.4 Evaluation Results 	 64
4.4.1 DVFS Scheme 	 64
4.4.1.1 Performance-sensitive Application 	 65
4.4.1.2 QoS-sensitive Application 	 68
4.4.1.3 Evaluation of Performance-sensitive applications on Oth-er Devices 	 71
4.4.2 CPU Shutdown Scheme 	 76
5. Power-effective Thermal Management 	 80
5.1 Stabilizing CPU Frequency and Voltage for Temperature-aware DVFS 	 80
5.1.1 Scheme Overview 	 80
5.1.2 Frequency Determination Routine 	 82
5.1.2.1 Power Optimizing Option 	 82
5.1.2.2 Performance Optimizing Option 	 86
5.1.3 Actual Frequency Scaling and Exit Cases 	 86
5.2 Evaluation Methodology 	 89
5.2.1 Sampling Period and Temperature Margin for Performance Optimizing 	 89
5.2.2 Benchmark 	 92
5.2.3 Evaluation Environment 	 94
5.3 Evaluation Results 	 95
5.3.1 Analysis of the Conventional DTM 	 95
5.3.2 Evaluation Results of Our Proposed Scheme: Power Optimiz-ing Option 	 97
5.3.3 Evaluation Results of Our Proposed Scheme: Performance Op-timizing Option 	 100
6. Conclusion and Future Work 	 105
References 	 108
Summary (in Korean) 	 115
Acknowledgement 	 118