1. Smartphone Background Activities in the Wild: Origin, Energy Drain, and Optimization

2. Outline Introduction Related Work
Characterizing Background Activities in the Wild
Trace Collection
Screen-off Intervals
Background Activities in Screen-off
Energy Drain
Methodology
Background Energy Analysis
Key Idea
BFC Analysis
Conclusion

3. Introduction
Many apps on smartphones wake up periodically to run when users are not actively interacting with them.
to perform a refresh of the app state
to sync with cloud services and get status updates or notifications
to support non-touch based user interactions
On average 28.9% of the daily energy drain is due to apps and services running during screen-off intervals.
(1) e.g., in apps that provide
updates for news, financial stocks, weather or twitter feed updates,
(2) e.g., in social networking apps such as Facebook, WeChat,
and Google+,
(3) such as in audio apps Pandora and Spotify which periodically download
files to play songs, or apps used for location tracking or navigation
which allow a user to listen to directions without screen interactions.

4. Introduction Turn off all the background activities.
iOS: Disable Background Refresh
Android: Restrict Background Data
Disable useful background activities, affecting user experience
Disable background activities respectively.

5. Introduction
Background activities, foreground activities, and user experience.
The usefulness of background activities of an app is likely to be user-dependent.
In this paper, we explore effective ways of optimizing background
activities of apps and services. Our exploration is motivated
by two hypotheses:

6. Introduction Background to Foreground Correlation(BFC).
The level of background activities of an app should be personalized to individual users.
A screen-off energy optimizer on Android: HUSH.
Saves screen-off energy of smartphones by 15.7% on average.
1. A metric to measure usefulness
2. Experiment
3. HUSH monitors the BFC of all apps on a phone online
and automatically identifies and suppresses app background activities
during screen-off intervals that are not useful to the user experience.
4. HUSH saves screen-off energy of smartphones
by 15.7% on average with minimal impact on the user experience
with the apps.

7. Related Work Most focus on the smartphone app traffic.
Falaki et al. characterized smartphone traffic based on traffic traces and showed how power consumption could be reduced
Huang et al. collected traffic and studied screen-off radio energy consumption
None of the work provided a thorough analysis of how energy is consumed on the smartphones in the wild.
Only a few have looked at the impact of background traffic activities on the power consumption.

8. Trace Collection Coarse-grained logging Fine-grained logging
On-demand event logging
Trace collection
Logging overhead
CPU usage
network usage
dynamic events: screen being switched on and off,WiFi being associated and scanning,WiFi and cellular signal strength change, battery level change (1% granularity) and the time each app starts and stops.
2000 Galaxy S3 and S4 devices;
2.4% of the total CPU time, 0.3% of the total network bytes, and 0.6% of the total energy drain.

9. Screen-off Intervals
The duration of screen-on/off intervals differ significantly.
The screen-off intervals tend to last much longer than screen-on intervals.

10. Background Activities in Screen-off
A longer screen-off interval does not necessarily imply more background activities and more energy drain.
9.8 vs 16.7
50.8 vs 22.7

11. Energy Drain - Methodology
A hybrid power model
power draw has linear correlation with utilization
FSM-based modeling for wireless interfaces such as WiFi/3G/LTE
The component power draw are largely independent, and hence the total power draw can be approximated by summing the power draw of individual components.

12. Energy Drain – Background Energy Analysis
Overall screen-on vs. screen-off
Back-ground app/service energy in screen-off
CPU idle energy in screen-off
Overall screen-on vs. screen-off: =45.9% of the total energy drain in a day occurs during screen-off
Back-ground app/service energy in screen-off: =28.9% screen off % screen on
CPU idle energy in screen-off: 50.8% of the total CPU time in idle in screen-off, but it only drains on average 6.0% of the total energy

13. Key Idea Not all of the background activities are necessary
Usefulness of app screen-off activities is
app-dependent
user-dependent
Premise
Background activities of apps are meant to improve user app experience but they are only useful if the user interacts with those apps in foreground some time during the next screen-on interval.
The usefulness of background activities of an app is likely to be user-dependent and thus their occurrences should be personalized.
Whether it would run in the foreground during the next screen-on interval.

14. Outline How to quantify usefulness?
Test the hypothesis
How to develop an online algorithm to optimize screen-off energy?

15. Quantify Usefulness: Background-Foreground Correlation (BFC)
1. Define per-interval
Screen-off interval
Screen-on interval
Background activity
Foreground activity b1 b2
time
2. BFC is the average of
0  low correlation  useless
1  high correlation  useful

16. BFC of 2000-User Traces BFC is app-dependent BFC is user-dependent
60% of apps have zero BFC
BFC is user-dependent

17. Stability of BFC
The BFC for the same app on the same device is fairly steady or changes slowly
BFC for the same app may vary sharply across users, it is fairly stable for individual users.

18. Prediction-based Online Algorithm
1. Keep track of per-app BFC for each user using exponential moving average ,
2. Suppress background activities in interval if

19. Choosing Parameters
1. 22% apps have more than 10 background activities per day
2. 10% apps have more than 36 background activities per day
3. 77% of apps have more daily background activities than daily foreground activities

20. Choosing Parameters
Case 1: there are few foreground activities and hence many background activities per foreground activity (e.g., Google Now, and Gmail).

21. Choosing Parameters
Case 2: there are many foreground activities and hence few background activities per foreground activity (e.g., Facebook-User 2, Whatsapp-User 2).

22. Choosing Parameters
Case 3: the app contains alternating phases with few and many foreground activities (e.g., Facebook-User 1, Whatsapp-User 1).

23. Evaluation Metrics 1. Energy saving: 2. Staleness: time
Background activity
Foreground activity

24. Evaluation of Prediction-based Online Algorithm
2.5x staleness increase
16.4% avg. energy saving (upper bound = 29%)
Can we improve staleness and maintain energy saving?

25. Analysis of High Staleness

26. Exponential Backoff Algorithm
Original algorithm：
Foreground activity
Background activity
time
staleness
Relax the strictness of suppressing
Exponential backoff：
time
threshold time:

27. Exponential Backoff Algorithm
Original algorithm：
Foreground activity
Background activity
time
staleness
Relax the strictness of suppressing
Exponential backoff：
time
staleness

28. Evaluation of Exponential Backoff Algorithm
staleness increase 2.5x 1.3x
avg. energy saving 16.4%  15.7%
staleness of individual apps reduces

29. LocationManagerService
Architecture of HUSH
ActivityManagerService
BatteryStatsImpl.Uid.Pkg.Serv
updateBg
updateFg
Intercept framework modules to suppress background activities on behalf of apps
allowHush
LocationManagerService
TelephoneRegistry
PendingIntentRecord
BroadcastQueue …
BatteryStatsImpl.Uid.Pkg{
long mBgTime;
long mThrTime;
void updateFg(){…}
void updateBg() {…}
boolean allowHush() {…} }

30. Early Evaluation of HUSH
2 Users: 3 days with original Android, 3 days with HUSH
User - 1
User - 2
Number of installed apps 73 52
Daily screen-on intervals 85 29
Daily screen-on time (min)
82.35
49.95
Daily suppressions by HUSH
4400
5543
Android
HUSH
Daily CPU busy time (min)
164.2
97.40
60.81
27.24
Maintenance power (mA)
12.76
12.12
Avg. screen-off power (mA)
15.57
5.27 
3.19
2.18 
Avg. screen-on power (mA)
316.8
323.5 
271.4
273.0 
Overall avg. power (mA)
45.50
36.34 
27.32
18.99  3x
1.5x
1.3x
1.4x

31. Conclusion Energy measurement study in the wild
29% of daily energy due to background activities during screen-off
Quantify usefulness of background activities
Background-Foreground Correlation
Usefulness is app-dependent and user-dependent
Screen-off energy optimizer: HUSH
Save 15.7% daily energy on average
Available at