33
Into the Wild: Studying Real User Activity Patterns to Guide Power Optimizations for Mobile Architectures Alex Shye, Benjamin Scholbrock, and Gokhan Memik Northwestern University Electrical Engineering and Computer Science Department Energy Efficiency – ELEC 518 Spring 2011 Jash Guo, Myuran Kanga Rice University Houston, TX Mar 17, 2011

Into the Wild: Studying Real User Activity Patterns to Guide Power Optimizations for Mobile Architectures Alex Shye, Benjamin Scholbrock, and Gokhan Memik

Embed Size (px)

Citation preview

Page 1: Into the Wild: Studying Real User Activity Patterns to Guide Power Optimizations for Mobile Architectures Alex Shye, Benjamin Scholbrock, and Gokhan Memik

Into the Wild: Studying Real User Activity Patterns to Guide Power

Optimizations for Mobile Architectures

Alex Shye, Benjamin Scholbrock, and Gokhan Memik Northwestern University Electrical Engineering and Computer

Science Department

Energy Efficiency – ELEC 518 Spring 2011

Jash Guo, Myuran KangaRice UniversityHouston, TXMar 17, 2011

Page 2: Into the Wild: Studying Real User Activity Patterns to Guide Power Optimizations for Mobile Architectures Alex Shye, Benjamin Scholbrock, and Gokhan Memik

2

Agenda

• Background

• The Paper– Introduction– Experiment– Findings– Evaluation– Conclusions

• Related Works/Topics

Page 2

Page 3: Into the Wild: Studying Real User Activity Patterns to Guide Power Optimizations for Mobile Architectures Alex Shye, Benjamin Scholbrock, and Gokhan Memik

3

Page 3

Background

• Venue– Proceedings of the 42nd Annual IEEE/ACM

International Symposium on Microarchitecture– MICRO 2009: December 12-16, 2009 – 52 out of 209 submissions

Page 4: Into the Wild: Studying Real User Activity Patterns to Guide Power Optimizations for Mobile Architectures Alex Shye, Benjamin Scholbrock, and Gokhan Memik

4

Page 4Gokhan Memik

Associate Professor

EECS, Northwestern

Alex Shye

PhD Student 2010’

EECS, Northwestern

Ben Scholbrock

PhD Student

EECS, Northwestern

Page 5: Into the Wild: Studying Real User Activity Patterns to Guide Power Optimizations for Mobile Architectures Alex Shye, Benjamin Scholbrock, and Gokhan Memik

Introduction●Increased need for mobile computing●Batch jobs/Long running services disabled – iPhone●End-user activity (Workload)●Android G1 logger – User power consumption●CPU frequency scaling/Screen Brightness

Page 5

http://www.mobilecrunch.com/wp-content/uploads/2010/06/iphone4_2up_angle.jpg

http://rdn-consulting.com/blog/2007/12/21/bci-brain-computer-interface/

Page 6: Into the Wild: Studying Real User Activity Patterns to Guide Power Optimizations for Mobile Architectures Alex Shye, Benjamin Scholbrock, and Gokhan Memik

Experiment• Architecture – HTC Dream• Power Estimation Model –

Using real measurements• Logger application• Deployment• Useful data

Page 6

High-level overview of the target mobile

architecture

Page 7: Into the Wild: Studying Real User Activity Patterns to Guide Power Optimizations for Mobile Architectures Alex Shye, Benjamin Scholbrock, and Gokhan Memik

Power Model/Building Estimation• Power states: Active/Idle• Choosing parameters• Estimation model build

• Real-time Measurements• R-tool – Linear Regression Model

Page 7

Parameters used for linear regression in

power estimation model

Page 8: Into the Wild: Studying Real User Activity Patterns to Guide Power Optimizations for Mobile Architectures Alex Shye, Benjamin Scholbrock, and Gokhan Memik

Model Validation• Additional logs recorded

• Strict hardware• Scenario

• Accurate power estimation – Median 6.6%

Page 8

Cumulative total energy error

Cumulative distribution of power estimation error

Page 9: Into the Wild: Studying Real User Activity Patterns to Guide Power Optimizations for Mobile Architectures Alex Shye, Benjamin Scholbrock, and Gokhan Memik

Per Component Power• Measured and predicted power consumption• Surfing the internet and streaming media for 160sec• Actual usage varies by workload• Similar breakdown for all components (next slide image)

Page 9

Power Consumption Timeline

Page 10: Into the Wild: Studying Real User Activity Patterns to Guide Power Optimizations for Mobile Architectures Alex Shye, Benjamin Scholbrock, and Gokhan Memik

Power Breakdown• Idle time a significant issue• Varying solutions based on workload• Summary

• Accurate total system power estimation• Power breakdown – Highly dependent on workload

Page 10

User power breakdown User power breakdown excluding idle time

Page 11: Into the Wild: Studying Real User Activity Patterns to Guide Power Optimizations for Mobile Architectures Alex Shye, Benjamin Scholbrock, and Gokhan Memik

Idle Time“Energy Efficiency of Handheld Computer Interfaces: Limits, Characterization and Practice,” Lin Zhong and Niraj K. Jha Department of Electrical Engineering - Princeton University

•Human sensory limits•Speech recognition rates vs. typing•Interface cache•User acceptance

Page 11

Interface cache wrist-watch device

Page 12: Into the Wild: Studying Real User Activity Patterns to Guide Power Optimizations for Mobile Architectures Alex Shye, Benjamin Scholbrock, and Gokhan Memik

12

Page 12

Findings

• The end user is the workload

• Variation in the power break-down between users

• The CPU and the screen are the two most power-consuming components

Page 13: Into the Wild: Studying Real User Activity Patterns to Guide Power Optimizations for Mobile Architectures Alex Shye, Benjamin Scholbrock, and Gokhan Memik

13

Page 13

Characterizing Real User Workloads

• The workload of a mobile architecture has a large effect on its power consumption

• The hardware components that dominate power consumption vary drsticaly depending upn the workload

• The user determines the workload for a mobile architecture

Page 14: Into the Wild: Studying Real User Activity Patterns to Guide Power Optimizations for Mobile Architectures Alex Shye, Benjamin Scholbrock, and Gokhan Memik

14

Page 14

Power Breakdown Including Idle Time

Page 15: Into the Wild: Studying Real User Activity Patterns to Guide Power Optimizations for Mobile Architectures Alex Shye, Benjamin Scholbrock, and Gokhan Memik

15

Page 15

Power Breakdown Excluding Idle Time

Page 16: Into the Wild: Studying Real User Activity Patterns to Guide Power Optimizations for Mobile Architectures Alex Shye, Benjamin Scholbrock, and Gokhan Memik

16

Page 16

The Paper Focus on Active State

• Idle State (about 68 mW)

• Active State (up to 2000 mW)

• Active state contributes highly to the user experience

• Active state accounts for 50.7% of the total system power

Page 17: Into the Wild: Studying Real User Activity Patterns to Guide Power Optimizations for Mobile Architectures Alex Shye, Benjamin Scholbrock, and Gokhan Memik

17

Page 17

Screen Usage of Real Users

• Screen Interval: a continuous block of time where the screen is on

• Duration: the length of time corresponding to the interval

• 70% of total screen duration > 100s

• The total duration time is dominated by a relatively small percentage of long screen intervals

Page 18: Into the Wild: Studying Real User Activity Patterns to Guide Power Optimizations for Mobile Architectures Alex Shye, Benjamin Scholbrock, and Gokhan Memik

18

Page 18

User-Aware Optimizations

• A few long screeen intervals dominate the overall screen duration time

• The power consumption during Active time is dominated by the screen and the CPU

• Change Blindness: the inability for humans to detect gradual/large changes in their surrounding environment

Page 19: Into the Wild: Studying Real User Activity Patterns to Guide Power Optimizations for Mobile Architectures Alex Shye, Benjamin Scholbrock, and Gokhan Memik

19

Page 19

Change Blindness

Page 20: Into the Wild: Studying Real User Activity Patterns to Guide Power Optimizations for Mobile Architectures Alex Shye, Benjamin Scholbrock, and Gokhan Memik

20

Page 20

Solutions

• Develop an accurate estimation model

• Slowly decrease CPU frequency

• Slowly decrease screen brightness

Page 21: Into the Wild: Studying Real User Activity Patterns to Guide Power Optimizations for Mobile Architectures Alex Shye, Benjamin Scholbrock, and Gokhan Memik

21

Page 21

CPU Optimization

• Dynamic frequency scaling (DFS) algorithm

• ondemand DFS governor

Page 22: Into the Wild: Studying Real User Activity Patterns to Guide Power Optimizations for Mobile Architectures Alex Shye, Benjamin Scholbrock, and Gokhan Memik

22

Page 22

Screen Optimization

• Decrease the brightness by 7 units every 3 seconds until 60% threthold

• Affect only long screen inervals

• Maintain user perception

Page 23: Into the Wild: Studying Real User Activity Patterns to Guide Power Optimizations for Mobile Architectures Alex Shye, Benjamin Scholbrock, and Gokhan Memik

23

Page 23

Experimental Results

• Screen Ramp• CPU Ramp

• Screen Drop• CPU Drop

• Emulate the optimizations on the user logs• Conduct a user study

Page 24: Into the Wild: Studying Real User Activity Patterns to Guide Power Optimizations for Mobile Architectures Alex Shye, Benjamin Scholbrock, and Gokhan Memik

•Power savings•User satisfaction•Evaluation with blind use of optimizations•Single run evaluations

Page 24

Results/Evaluation

Total system power savings for each of the optimizations as estimated by our power model

Page 25: Into the Wild: Studying Real User Activity Patterns to Guide Power Optimizations for Mobile Architectures Alex Shye, Benjamin Scholbrock, and Gokhan Memik

Page 25

User Satisfaction

Reported user satisfaction

Page 26: Into the Wild: Studying Real User Activity Patterns to Guide Power Optimizations for Mobile Architectures Alex Shye, Benjamin Scholbrock, and Gokhan Memik

Feedback and Solution Acceptance• User disclosure – Screen/CPU significance• Feedback based on input response• CPU frequency change – Jitter• Change blindness beneficial• Optimization On/Off tool?• User pattern essential to proper power consumption

reduction

Page 26

Glitchy Screenhttp://www.flickr.com/photos/aparrish/5515150358/sizes/l/in/set-72157626237465468/

Page 27: Into the Wild: Studying Real User Activity Patterns to Guide Power Optimizations for Mobile Architectures Alex Shye, Benjamin Scholbrock, and Gokhan Memik

27

Page 27

Conclusion

• Mobile architectures – natural environment• Logger application to collect logs• Develop power estimation model• Findings show CPU and screen dominate usage• Optimizations based on user behavior• Change blindness utilized for 10% total savings

Page 28: Into the Wild: Studying Real User Activity Patterns to Guide Power Optimizations for Mobile Architectures Alex Shye, Benjamin Scholbrock, and Gokhan Memik

28

Page 28

Benefits/Criticisms

• Pros– Estimation Model– The Logger– Real Users– Real Patterns– Usage interval

awareness– Change Blindness

• Cons– Linear?– Logger Overhead– Sample Size

• Single model• 20 users• 145/250 days

– Device/User Gap– Major focus on CPU– Future Trends

• More WiFi, EDGE

Page 29: Into the Wild: Studying Real User Activity Patterns to Guide Power Optimizations for Mobile Architectures Alex Shye, Benjamin Scholbrock, and Gokhan Memik

29

Page 29

Related Paper

• Power to the People: Leveraging Human Physiological Traits to Control Microprocessor Frequency (2008)

Power saving by better understand the individual user satisfaction

Page 30: Into the Wild: Studying Real User Activity Patterns to Guide Power Optimizations for Mobile Architectures Alex Shye, Benjamin Scholbrock, and Gokhan Memik

30

Page 30

Related Paper

• Energy Efficiency of Handheld Computer Interfaces: Limits, Characterization and Practice(2005)

Utilize interface cache for small tasks

Typical text entry speeds

Page 31: Into the Wild: Studying Real User Activity Patterns to Guide Power Optimizations for Mobile Architectures Alex Shye, Benjamin Scholbrock, and Gokhan Memik

31

Page 31

Related Paper

• Energy-aware adaptation for mobile applications (1999)

Tradeoff between energy conservation and application quality

Page 32: Into the Wild: Studying Real User Activity Patterns to Guide Power Optimizations for Mobile Architectures Alex Shye, Benjamin Scholbrock, and Gokhan Memik

32

Page 32

Related Paper

• Human Generated Power for Mobile Electronics (2004)

Alternatives to batteries: additional power sources

Human power generation

Page 33: Into the Wild: Studying Real User Activity Patterns to Guide Power Optimizations for Mobile Architectures Alex Shye, Benjamin Scholbrock, and Gokhan Memik

33

Page 33

Human Factors in Power Savings• Source

– Better batteries– Additional power sources

• Hardware

• Software

• Monitoring

• Alarming

• Perception– User satisfaction– Quality vs. performance sacrifice

Human power generation – Proof of concept