PPT

18
INTO THE WILD Studying Real User Activity Patterns to Guide Power Optimizations for Mobile Architectures Alex Shye Ben Scholbrock Gokhan Memik Northwestern University, EECS Empathic Systems Project 12/14/200 9 International Symposium on Microarchitecture. New York, NY. 1

Transcript of PPT

INTO THE WILDStudying Real User Activity Patterns to Guide Power

Optimizations for Mobile Architectures

Alex ShyeBen ScholbrockGokhan Memik

Northwestern University, EECSEmpathic Systems Project

12/14/2009

International Symposium on Microarchitecture. New York, NY. 1

The rise of mobile architectures

12/14/2009

International Symposium on Microarchitecture. New York, NY. 2

ENIAC1946

IBM System/3601964

Apple II

1977

IBM Thinkpad 7001992

1950 1960 1970 19901940 2000 20101980

Media players, PDAs, smartphones, netbooksToday

Trends Implications

“Personal” nature The user drives execution

Form factor Total energy store

Power optimization is important

Summary

12/14/2009

International Symposium on Microarchitecture. New York, NY. 3

Observation: For mobile architectures, the user is the workload

Claim: Computer architects should study user activity to:

(1) Characterize power consumption

(2) Guide the development of power optimizations

Summary

12/14/2009

International Symposium on Microarchitecture. New York, NY. 4

Observation: For mobile architectures, the user is the workload

Claim: Computer architects should study user activity to:

(1) Characterize power consumption

(2) Guide the development of power optimizations

• Log the activity of real users using the Android G1

• Develop an accurate power model for the Android G1

• Power consumption varies across users

• The screen and CPU consume the most power

• Screen time dominated by long screen intervals

• Develop optimizations leveraging change blindness

• Save ~10% total system power with minimal change to user satisfaction

Target Mobile Architecture• We study the HTC Dream mobile platform (Android G1)

– First platform based upon Android OS– 2 phones, identical hardware:

• Google Android Developer Phone 1 (ADP1)• T-Mobile G1 – commercialized version of the G1

12/14/2009

International Symposium on Microarchitecture. New York, NY. 5

[Google ADP1]

How do we proceed with power optimizations?Which components consume the most power?

Logging Real User ActivityNU JamLogger• Runs on any Android G1• Logs system activity

– CPU utilization, Wifi traffic, SD card traffic, screen usage, etc.

• Periodically sends logs to our server• Lightweight

– < 5% CPU overhead during active phone use

– Logs compressed with gzip before upload

12/14/2009

International Symposium on Microarchitecture. New York, NY. 6

Getting Users• Released NU JamLogger on Android Market• Advertised at Northwestern University, University of Michigan, and

online (Slashdot, online Android forums, etc)• 20 users, ~250 days of cumulative user activity

• Each user has over 1 weeks worth of logs

Power Estimation Model

12/14/2009

International Symposium on Microarchitecture. New York, NY. 7

Power Measurements

InstrumentedBattery via

Current Clamp

HW Unit Parameter

CPU hi_cpu_utilmed_cpu_util

Screen screen_onscreen_brightness

Call call_ringingcall_off_hook

EDGE edge_has_trafficedge_traffic

Wifi wifi_onwifi_has_trafficwifi_traffic

SD Card sdcard_traffic

DSP music_on

System system_on

Idle idleLinear

RegressionPower Model

Time-based Traces

Jam Logs

…49446 : CPU_Utilization 23.36 21.50 1.8749491 : Load_Avg 4.14 3.49 3.30 1 185 36350343 : Screen_Off50557 : Wifi_Traffic 10 0…

1 sec. samplesw/ parameters

Accuracy of Power Model

12/14/2009

International Symposium on Microarchitecture. New York, NY. 8

• Built power model with one ADP1, validated on a separate ADP1• 6.6% median error per sample• < .1% error summed across all samples (total energy)

Hardware Specific Runs Usage Scenarios

Using the Power Model

• Estimated power closely tracks real-time measurements• Can sum parameters to derive a power breakdown

12/14/2009

International Symposium on Microarchitecture. New York, NY. 9

CPU

Wifi

DSP

Screen

Power Breakdown (20 Users)

12/14/2009

International Symposium on Microarchitecture. New York, NY. 10

• There exists a significant variation in users• Idle is very important, consuming 49.3% of total energy

Power Breakdown (20 Users w/o Idle)

12/14/2009

International Symposium on Microarchitecture. New York, NY. 11

• Ignoring Idle time, the screen and CPU are most power hungry– Screen: 35.5% (19.2% from brightness)– CPU: 12.7%

Studying User Activity for Optimization

• Screen activity is a good indicator of usage activity

• We study screen intervals, periods of time the screen is on

• Observation: The total screen time is dominated by a small number of long screen intervals

– Screen intervals of 100+ seconds account for ~70% of screen time

• We develop a power optimization strategy that leverages change blindness to target the screen and CPU

12/14/2009

International Symposium on Microarchitecture. New York, NY. 12

Change BlindnessThe inability to distinguish changes in a stimulus

12/14/2009

International Symposium on Microarchitecture. New York, NY. 13

[Simons 99]

Ramp Optimizations• Can change blindness be applied for power optimization?

• We target the screen and CPU for optimization:– Screen Ramp: slowly dim brightness

• Decreases brightness by 7 units (max 255) until 60% of starting brightness

– CPU Ramp: slowly decrease effective frequency• Reaches 70% of maximum frequency in 40 seconds

• Implemented by tuning ondemand DFS governor

• We compare to Drop optimizations– Screen Drop: drop to 60% brightness after 40 seconds– CPU Drop: drop to 70% frequency after 40 seconds

12/14/2009

International Symposium on Microarchitecture. New York, NY. 14

Power Savings

• Simulate optimizations on real user traces

• CPU Ramp saves 4.9% system power (22.8% of CPU power)• Screen Ramp saves 5.7% system power (19.1% of screen power)

• Ramp optimizations outperform Drop optimizations

12/14/2009

International Symposium on Microarchitecture. New York, NY. 15

Impact on User ExperienceUser study design• Run four optimizations and a control run (in random order), on:

– Web browsing: Browse Wikipedia– Game: BreakTheBricks game– Video: Watching video with PlayVideo

• Ask for verbal satisfaction rating from 1 (low) – 5 (high)• At end of study, debrief user, and discuss acceptance of optimizations

12/14/2009

International Symposium on Microarchitecture. New York, NY. 16

Results• No significant difference in user satisfaction when compared to control

(using paired t-test) except for small changes in:– All optimizations with the Game– Screen Drop with Video

• Users mostly rated based upon feeling of jumpiness/jitter• Works well for screen; most users couldn’t tell screen was changing• 15 users would use some of optimizations, 4 would not, 1 apathetic

Conclusion• We study real users on real devices in real environments

• Develop and user power model to characterize power consumption– Show that power consumption does vary across users– When active, the screen and CPU consume the most power

• Present an example of a user-activity-driven power optimization– Develop ramp optimizations targeting the screen and CPU– Save 5.7% and 4.9% of total system power for the screen and CPU,

respectively– Ramp optimizations well-accepted by users, especially for the screen

12/14/2009

International Symposium on Microarchitecture. New York, NY. 17

Computer architects should study real user activity to:(1) Characterize power consumption

(2) Guide the development of power optimizations

Thank You!• Questions?

12/14/2009

International Symposium on Microarchitecture. New York, NY. 18

Alex Shyehttp://www.ece.northwestern.edu/[email protected]

ESP: Empathic Systems Projecthttp://www.empathicsystems.org

NU JamLogger: Available on Android Markethttp://www.ece.northwestern.edu/microarchitecture/jamlogger/