1. Fine-Grained Power Modeling for Smartphones Using System Call Tracing Abhinav Pathak, Y. Charlie Hu Purdue University Ming Zhang, Paramvir Bahl, Yi-Min Wang Microsoft Research 2011 EuroSys

2. Motivation Accurate, fine-grained online energy estimation of mobile devices. ◦ Understanding and debugging the energy consumption of mobile applications. Current utilization-based power modeling is insufficient for modeling the power consumption of mobile devices.

3. Utilization-Based Power Modeling The power state of a device is correlated to the utilization of its hardware components. ◦ Ex: CPU utilization 0%=> 150W CPU utilization 50%=> 250W CPU utilization 90%=> 340W This model can be used to perform online estimation of power consumption. ◦ Collects the utilization of each components and feeds them into the model.

4. New Challenges Several components have tail power states. ◦ Tail power: a component stays at high power state for a period of time after active I/O activities. System calls that do not imply utilization can change power states ◦ Could be due to power optimizations programmed in device drivers.

5. New Challenges(cont.) Several components do not have quantitative utilization. ◦ “Exotic” components such as GPS and camera.

6. System-Call-Based Power Modeling Based on the following observations: ◦ System calls is the only way an app can access hardware (I/O) components. ◦ System calls can trigger power state transition. ◦ System calls that turn on/off “exotic” components trigger power state change immediately. ◦ Using system calls as triggers naturally suggests using a FSM to model state transitions. ◦ System calls can be easily related back to the calling subroutine.

7. System-Call-Based Power Modeling(Cont.) Model the power state of a device using FSM. ◦ Each state of the FSM is a power state of the device.  Can be easily annotated with timing and workload of recent events. ◦ The state transition is triggered by system call.

8. Construction Step I Modeling single system call power consumption

9. Construction Step II Modeling Multiple System Calls(to the same component) ◦ Concurrent system calls  The second system call is invoked before the component is out of the productive or tail state due to the first system call. ◦ Power states of a component  Take the union of all power states discovered by modeling individual system calls.

10. Construction Step II(Cont.) ◦ Modeling state transition  A system call arrives after the previous one is out of its productive power state. => superimposing the FSM of the second on top of the first.  A system call arrives while the previous one is still in its productive power state. =>  Workload-based system call(ex: read/write)  workload_first => workload_first + workload_second  Initialization-based system call(ex: open)  Productive state => Tail state => Superimposing

11. Construction Step III Modeling Multiple Components ◦ Android  The total power consumption is the summation of those of individual components when active in isolation. ◦ Windows Phone  Tail states of different components interfere with each other.  Try out all possible combinations of the sets of conditions, each set for driving one component into all possible states, and measure the corresponding total power consumption.

12. CTester Applications A testing benchmark suite. ◦ An application for each component carefully designed to exercise all the relevant system calls. ◦ A wrapper application that invokes individual applications at predetermined timing. Run CTester and measure the power dissipated through an external power meter.

13. Implementation Tracing System Calls ◦ Windows Mobile 6.5 with Windows CE 5.2 kernel ◦ Android 2.2 running Linux Kernel 2.6.34 Flash Customized kernel Images

14. Hardware Platform Phone is connected to a power supply with power monitor. Power monitor samples every 200 ms. NameHandset HTC-CPU (MHz)OS (kernel)BasePower Magic 528Android 2.0 (Linux 2.6.34) 160mA Touch 528WM6.1 (CE5.2)250mA Tytn2Tytn II400WM6.1 (CE5.2)130mA

17. Energy Consumption Estimation Compare the accuracy of fine-grained and whole-application energy consumption estimation. ◦ System-call-based ◦ Utilization-based  LR model ◦ Only consider LCD, CPU, sacard, and wifi NIC.

19. Fine-grained Interval: 50ms

20. Whole-application

21. Conclusion This paper presents the design and implementation of a system-call-based power modeling approach. The experimental results on Android and Windows Mobile using a diverse set of applications show that the new model drastically improves the accuracy of fine- grained energy estimation as well as whole application energy estimation.