1. Smartphone Energy Drain in the Wild: Analysis and Implications Authors: Xiaomeng Chen, Ning Ding, Abhilash Jindal†, Y. Charlie Hu†, Maruti Gupta, Rath Vannithamby

2. Contents  Objective and contributions  A hybrid power model  Measurement methods  Result and conclusion  Objective and contributions  A hybrid power model  Measurement methods  Result and conclusion

3. Objectives and contributions  Objective  In order to extending smartphone battery life, we need to have a better understanding on where and how is energy drain happening on users’ phones under normal usage, for example, in a one-day cycle.  Objective  In order to extending smartphone battery life, we need to have a better understanding on where and how is energy drain happening on users’ phones under normal usage, for example, in a one-day cycle.  Contributions  A hybrid power model that integrates utilization-based models and FSM- based models for different phone components.  Through analyzing traces collected on1520 devices in the wild, we present detailed analysis of where the CPU time and energy are spent.  Contributions  A hybrid power model that integrates utilization-based models and FSM- based models for different phone components.  Through analyzing traces collected on1520 devices in the wild, we present detailed analysis of where the CPU time and energy are spent.

4. A hybrid power model  Utilization- based: CPU, GPU  FSM-based: WiFi, 3G, LTE  Constant: WiFi beacon, cellular paging and SOC suspension  Utilization- based: CPU, GPU  FSM-based: WiFi, 3G, LTE  Constant: WiFi beacon, cellular paging and SOC suspension

5. Utilization-based power model  CPU: Utilization + frequency  Trace data: app-wise CPU usage, per- core CPU usage and the duration staying on different frequencies.  CPU: Utilization + frequency  Trace data: app-wise CPU usage, per- core CPU usage and the duration staying on different frequencies.

6. Utilization-based power model  GPU  Three power states: Active, Nap, Idle  Power states + frequency  GPU  Three power states: Active, Nap, Idle  Power states + frequency

7. FSM-based power model  Network events + signal strength  The state-machine models for wireless interfaces are driven by network events collected in a packet trace or a network system call trace. On an unmodified user phone, we can Estimating network events from network usage  Network events + signal strength  The state-machine models for wireless interfaces are driven by network events collected in a packet trace or a network system call trace. On an unmodified user phone, we can Estimating network events from network usage

8. constant power model  WiFi Beacon: periodically wakes up to receive beacons from the AP.  We can average the energy of 50 spikes over the duration of the spikes and then model the WiFi beacon power as constant current.  Cellular paging: similar with WiFi Beacon.  SOC suspension: When the CPU and other hardware components are offline, the entire SOC is suspended and draws a constant current.  WiFi Beacon: periodically wakes up to receive beacons from the AP.  We can average the energy of 50 spikes over the duration of the spikes and then model the WiFi beacon power as constant current.  Cellular paging: similar with WiFi Beacon.  SOC suspension: When the CPU and other hardware components are offline, the entire SOC is suspended and draws a constant current.

9. Measurement methods  A free android app called eStar.

10. Measurement Model Validation CPU CPU + Network CPU+ Screen+ GPU

11. Results and implications-- CPU Time Analysis  Divide the users into 5 groups by total screen-on time  A huge portion of screen-off CPU time is spent idle (50.4%) which should ideally be close to zero.  Divide the users into 5 groups by total screen-on time  A huge portion of screen-off CPU time is spent idle (50.4%) which should ideally be close to zero.

12. Results and implications-- Energy Analysis  Energy breakdown by activities  45.0% of the total energy drain in a day occurs during screen-off periods.  21.2% of the total energy drain due to SOC base power, WiFi beacon, WiFi scanning and cellular paging activities during screen-off periods does not contribute to any useful work.  Out of the 55.0% energy incurred during screen-on periods, 24.0% is spent in screen.  Energy breakdown by activities  45.0% of the total energy drain in a day occurs during screen-off periods.  21.2% of the total energy drain due to SOC base power, WiFi beacon, WiFi scanning and cellular paging activities during screen-off periods does not contribute to any useful work.  Out of the 55.0% energy incurred during screen-on periods, 24.0% is spent in screen.

13. Results and implications-- Energy Analysis  Energy breakdown by components  SOC: On average the SOC accounts for 5.2% of the daily total energy drain.  Cellular paging vs. WiFi beacon: The energy drain from cellular paging and WiFi beacon are 12.5% and 2.3%, respectively.  Cellular vs. WiFi: The energy drain over cellular (LTE and 3G) and over WiFi are 11.7% and 1.3%, respectively.  a significant portion of wireless interface energy drain are tail energy, 89.3% out of total cellular energy, and 90.4% out of the total WiFi energy  Energy breakdown by components  SOC: On average the SOC accounts for 5.2% of the daily total energy drain.  Cellular paging vs. WiFi beacon: The energy drain from cellular paging and WiFi beacon are 12.5% and 2.3%, respectively.  Cellular vs. WiFi: The energy drain over cellular (LTE and 3G) and over WiFi are 11.7% and 1.3%, respectively.  a significant portion of wireless interface energy drain are tail energy, 89.3% out of total cellular energy, and 90.4% out of the total WiFi energy

14. Results and implications-- Device Evolution: S3 vs. S4  S4/JB CPU idle time increases by 72.3% than A3/JB and CPU busy time by foreground apps, background apps and background services shrink by 62.7%, 48.8% and 71.3% respectively, due to a faster CPU.

15. Result and conclusion-- Android Evolution: Jellybean vs. Kitkat  Background CPU time increased due to more apps on KitKat devices.  on average KitKat devices have 68.2% fewer wakeups than Jellybean devices due to a new API that can aggregate wakeup activities(background activities and network activities).  Background CPU time increased due to more apps on KitKat devices.  on average KitKat devices have 68.2% fewer wakeups than Jellybean devices due to a new API that can aggregate wakeup activities(background activities and network activities).

16. Conclusion  we make effort towards understanding where and how energy drain happens in smartphones running in the wild.  We developed a hybrid utilization-based and FSM-based model that accurately estimates energy breakdown among activities and phone components without changing the Android framework or rooting the phone.  We draw implications to the phone vendors, SOC vendors, cellular carriers, and app developers on better system, network, and app design to extend battery life.  we make effort towards understanding where and how energy drain happens in smartphones running in the wild.  We developed a hybrid utilization-based and FSM-based model that accurately estimates energy breakdown among activities and phone components without changing the Android framework or rooting the phone.  We draw implications to the phone vendors, SOC vendors, cellular carriers, and app developers on better system, network, and app design to extend battery life.

17.  THANK YOU