V-EDGE: FAST SELF-CONSTRUCTIVE POWER MODELING OF SMARTPHONES BASED ON BATTERY VOLTAGE DYNAMICS

Transcription

1 USENIX NSDI 2013 V-EDGE: FAST SELF-CONSTRUCTIVE POWER MODELING OF SMARTPHONES BASED ON BATTERY VOLTAGE DYNAMICS Fengyuan Xu, Yunxin Liu, Qun Li, and Yongguang Zhang

2 Introducing Power Model 1 Power model - relationship between power draw & system activities It is foundation to power management & optimization software energy mapping %

3 Introducing Power Model 2 Power model - relationship between power draw & system activities It is foundation to power management & optimization 30% CPU usage 150/255 backlight level software energy F_cpu(30%) mapping + F_bl(150) + F_... (x) = 1400 mw %

4 Introducing Power Model 3 Power model - relationship between power draw & system activities It is foundation to power management & optimization software energy mapping %

5 How to Build Power Models? 4 power consumption input regression output models controlled resource usage Resource usage Sampled CPU utilizations Sampled Screen backlight levels Power consumption Voltage Discharging current P = V * I models Linear Nonlinear

6 Requirements of Power modeling 5 models are built system changes occur system is under control real-time system Every phone is unique è Personalized models 1 User activities are not expected In training stage 2 Model adaption is needed è Fast training

7 External Metering 6 Measure V and I from external hardware Calculate power from V * I Monsoon power monitor BattOr power monitor Drawbacks n Labor-intensive n Inflexible/inaccurate n Expert knowledge required

8 Self Metering 7 Measure V and I from interior battery interfaces Battery interfaces are registers exposed by battery fuel gauge IC Disadvantage n A large number of existing phones cannot support, e.g., Galaxy Nexus

9 SOD Approach 8 Self metering with SOD (State of Discharge) SOD: percentage of energy left in battery Calculate power from the SOD changes Limitations n Slow n There are only 100 discrete SOD values n Wait a long time period to observe value changes n Inaccurate SOD 70% SOD 100%

10 V-edge Metering 9 Leverage battery characteristics Instantaneous current changes lead to instantaneous output voltage dynamics Internal resistance V_r R + V V_out Open Circuit Voltage (OCV) battery I - phone

11 Battery Dynamics 10 Discharging current changes lead to output voltage dynamics V OCV V_r R I battery + V_out - Observed by battery interface V_out = OCV V_r = OCV I * R ΔI = 1/R * ΔV_out V_edge V_edge è I we can infer current from voltage dynamics

12 Theory vs. Experiment 11 V-edge: voltage difference before and after an operation V_edge = V_1 V_2 current change V_1 Voltage drop V_2 Voltage rise V_out voltage 1Hz I CPU busy Long time in between time

13 V-edge is Accurate 12 Stable linear relationship between V-edge and current change Test on 8 batteries of two phones Various current change cases Coefficient of Determination R^2 > 0.995

14 V-edge is Fast and Sensitive 13 Fast As fast as battery interface update rate n E.g., on Nexus S, ¾ sec (V-edge) V.S. 15 min (SOD) Sensitive Detect 4% CPU usage change with 100% success ratio Fine-grained V-edge resolution

15 V-edge Implementation 14 training stage estimation stage only needed in training stage

16 Model Considered 15 CPU Frequency f and utilization U n P = a_f * U + b_f Screen Backlight level L and average pixel color RGB n P = F(L) * ( c_r * R + c_g * G + c_b * B ) Wi-Fi Throughput D n P = d * D + e GPS Service status S n P = f * S

17 Evaluation 16 Training overhead 400+ training programs è 1.2 h building time 100X faster than SOD Accuracy Real energy consumption error n Stricter than model parameter comparison Component model n Random benchmarks on CPU, screen, Real applications n Include video playback, VoIP call, web browsing

18 Accuracy 17 Comparison: G: ground truth measurement V,P: estimations using models V: V-edge P: external-metering-based directly measured power G system activities models predicted power V,P

19 Accuracy CPU and Screen 18 CPU screen

20 Accuracy CPU and Screen 19 error 5.8% diff. 3.7% CPU error: V-edge error compared to ground-truth diff.: V-edge value difference compared to externalmetering-based power modeling error 5.8% diff. 3.5% screen

21 Accuracy WiFi, GPS, and Apps V edge model meter model ground truth WiFi GPS normalized energy cost real Apps 0 GALLERY BROWSER ANGRY BIRDS VIDEO SKYPE GPS STATUS

22 Accuracy WiFi, GPS, and Apps V edge model meter model ground truth error 14% diff. 3.8% WiFi error 10% diff. 6.5% GPS normalized energy cost error 10% diff. 3.8% real Apps 0 GALLERY BROWSER ANGRY BIRDS VIDEO SKYPE GPS STATUS

23 Conclusions 22 Key finding on battery powered devices current change can be determined from instantaneous voltage change A new self-constructive power model building with only V readings Works for most phones 100X faster than SOD method Within 4% difference to models using external metering Evaluations demonstrated the effectiveness in power modeling

24 Thank you 23 Any questions?