Mobilesoft 2017 Keynote

Future of Mobile Software
Energy and Performance
Abhik Roychoudhury
abhik@comp.nus.edu.sg
http://www.comp.nus.edu.sg/~abhik
1Mobilesoft 2017 Keynote
Keynote Talk
Joint work with Abhijeet Banerjee

Devices
• Many, but concretely …
Mobilesoft 2017 Keynote 2

Processor & Application
• Developments in mobile processor level for energy and performance
• What about apps?
o How much can apps help?
o Are energy and performance synonymous?
ARM big.LITTLE

Energy vs. performance
• Impact of CPU frequency scaling on energy and
performance.
do{
scrape_puzzle(); // I-O intensive
process_puzzle(); // CPU intensive
} while (more puzzles to scrape and process);
Observe time and energy for different frequencies.

Energy vs. Performance
Energy
consumption
and
performance do
not rise and fall
together.

Energy vs. Performance
Source: “Energy consumption in mobile phones: A measurement study and implications for
network applications”, N. Balasubramanian et al, SIGCOMM 2009.
Speed
GSM < 3G < Wifi
3G consumes more
energy than GSM.
Wifi consumes less
energy than 3G,
GSM.

So far
• Energy and Performance are not synonymous.

Detecting Energy-Inefficiency -
8
Energy Consumption (High
Utilization*)
Energy Consumption (Low
Utilization*)
>
* Of device’s components
=> E/U ratio is a suitable measure of energy-
inefficiency of an application
Mobilesoft 2017 Keynote

Measuring
• PowerTutor: Tracks
components and their power
states, and uses inbuilt power
model
• Dumpsys: Uses power profile
(XML file) shipped with devices

Measurements
• Testing, Debugging, Re-factoring …
• Power measurements to guide Mobile SE activities

So far
• The need for energy measurements.

Energy-Inefficiencies: Energy Bugs
Prevents device from becoming idle even after app has
completed execution and there is no user activity.
12
PRE EXC REC POST
E/URatioE/URatio
No Bug/Hotspot
With Bug
Find dissimilarity between PRE and POST stages

Energy-Inefficiencies: Energy Hotspot
13
Causes device to consume abnormally high amount of battery power
even though the utilization of hardware resources is low
PRE EXC REC POST
E/URatioE/URatio
No Bug/Hotspot
With Hotspot
Anomaly Detection Techniques: Jmotif
"HOT SAX: Efficiently Finding the Most Unusual Time Series Subsequence"
E. Keogh, J. Lin and A. Fu (2005).
D1
D2

Energy Bugs & Hotspots
14
Category Energy Bug Energy Hotspot
Hardware
Resources
Resource Leak Suboptimal
Resource Binding
Sleep-state
transition heuristics
Wakelock Bug Tail-Energy
Hotspot
Background
Services
Vacuous
Background Services
Expensive
Background Services
Defective
Functionality
Immortality Bug Loop-Energy
Hotspot
App Starts
Resource
Acquired
Resource Used
Energy Wasted
App Stops
Service Started Service not stopped
Energy Wasted
App Starts
Buggy App
Starts
Reset Using Task Manager /
Restart
Buggy App
Re-starts

Hotspot / Bug
Detection
Component
Guidance
Component
Power
Meter
Hotspot/Bug
Detection
EFG
Extraction
Guidance
Module
Database
Event Trace
Generation
Smartphone
Energy
Consumption Trace
Utilization Trace
Hotspot
/ Bugs
Event Flow
Graph
Event Trace
Application
Detection
15Test SuiteMobilesoft 2017 Keynote

Example: Energy Bug
16
E/U Trace before fix has been applied
E/U Trace after fix has been applied
Aripuca
Records tracks and
Waypoints
Energy Bug
Vacuous Background
Services
public void onPause() {
unregisterReceiver(compassBroadcastReceiver);
unregisterReceiver(locationBroadcastReceiver);
//FIX
serviceConnection.getService().stopLocationUpdates();
serviceConnection.getService().stopSensorUpdates();
serviceConnection.unbindAppService();
}
public void onPause() {
unregisterReceiver(compassBroadcastReceiver);
unregisterReceiver(locationBroadcastReceiver);
//
//
//
serviceConnection.unbindAppService();
}
Bug-revealing event-traces:
Event_1_1 --> Event_1_2 --> Event_2_1 --> Event_2_2
Event_1_1 --> Event_1_2 --> Event_2_1 --> Event_3_1
Fix: Remove location updates before exiting

Example: Energy Hotspot
E/U Trace before fix has been applied
E/U Trace after fix has been applied
Montreal Transit
Fetches Transit Related Info
High E/U ratio for a short period (~ 5 seconds) after EXC stage
High E/U does not persist, therefore hotspot and not a bug
Energy Hotspot
Expensive Background
Services,
Suboptimal Resource Binding
Main Functionality
+ Get Location
Load Ads
+ Get Location
User
Navigates Away
loading ads
Stop
GPS released
User
Starts App
Energy
Wasted
Main
Thread
Main Functionality
+ Get Location
Load Ads
Share Location
User Navigates Away
+ GPS Released
loading ads
StopUser
Starts App
Main
Thread
Async
Thread
17

So far
• What are energy bugs and hotspots?

Re-factoring
19
Resources such as I/O Components and power management
utilities have the biggest impact on energy-consumption,
hence their usage must be reduced as much as possible
without affecting the functionality of the app.
PowerConsumption
(mW)

Re-factoring
20
Design
Extraction
Guideline-
based
Refactoring
Code
Generation
App
Source
Refactored
App
Source
Design
Expression
Refactored
Design
Expression
Design
Extraction

Extracting Design
Expression
21
Ap
p
Explore
UI
Analyze
Bytecode
Event
Flow
Graph
Convert
to DFA
Deterministic
Finite
Automata
DFA to
RE
Design
Express
ion
xr is acquire
yr is release
Change to ab*cxrurdyr

So far
• Re-factoring

Field Failures
23
Omnidroi
d
Issue 98
Battery Drain Concerns …
I know it was omnidroid
because I looked at the battery
use graph and omnidroid was
using 8 times more battery than
the next highest app
Field Failure
A defect that is observed by the user, after the software
has been shipped
Defect Report
A report describing the observed
defect.
Often provided by the user. May
or may not be sufficient to recreate
and/or fix the defect.
Energy
Inefficiency
A key concern
in battery
constrained
mobile
devices

Field Failures
24
App
App User
Reviews/
Bug Reports
App
Development
Challenges on
the
User Side
Challenges on
the
Developer Side
Replicate
Defect
Localize
Defect
Visualize
Defect
Patch Defect
Record
Sufficient
Information
to Recreate
Defect
Convey
Observed
Defect
Information
Succinctly

Going beyond
How can we utilize user comments related to energy consumption to
drive present and future mobile SE?

Reviews
Why do users downvote apps?
Crash
Advertisement
Performance
Energy
Other functionality
Gathered reviews of 170,000 Android apps in Google Play Store.

Uninstall ratio - plot
27Uninstallcategory
Reviewscategory
Highest uninstall ratio observed for energy related issues.

Possibility
28
User
Review
Temporal
Logic
formula
App
EventFlo
w Graph
Model
Test
Generator
Tests
Event Keywords: Open, Click, Tap, Drag, …
Temporal Keywords: Then, Until, Finally, …
HW State Keywords: GPS is on, …
UI State Keywords: <different UI states of the app>

Example
I 'm having a similar problem I have a Dell Streak running dj_steve 's custom
Android rom When I open Bitcoin-Android the circular thinking/downloading
thing in the top right corner keeps on spinning forever When I close that window
the app is still draining my cpu at over 90 % I tried sending 0.01 BTC to it and
it never showed up in the app but is confirmed on my pc client I was wondering if
this has something to do with me running a custom rom even though every other
app seems to work fine
G(user_open  (downloading U (user_close  F battery_drain))
G(user_open  (downloading U (user_close 
F (user_BCsnd  GBC_rcv ))
Test generation, not Model checking

Example
31
Had this issue as well turns out the block chain just takes a LONG time
to download Once the chain is downloaded took me about 5 hours
over 3G my transfers showed up new transfers showed up instantly and
the app does n't seem to be draining the battery at all Only dues it when I
open it and it needs a few seconds to a minute to download the new
blockchain piece
G(user_open  (downloading U (downloaded
G(battery_ok  (user_BCsnd  F BC_rcv))
Compare with
G(user_open  (downloading U (user_close  F battery_drain))
G(user_open  (downloading U (user_close 
F (user_BCsnd  GBC_rcv ))

32
Tests
App Execution
Logs
Fault
Localization
Fault
Visualization
Patch
Suggestion
Possible Workflow

So far
• Re-factoring
• Mobile SE based on field failures and reviews

Looking forward
Rise of drones - operations in
risky tasks
- Disaster Management
- Stealth Monitoring
Distributed Energy
Management
- Distributed among tasks
- Caters for many
operational scenarios

Futuristic Disaster Management
Tasks Camera Sensors GPS Propulsion Processor Radio Priority
Obstacle
Avoidance
X X X 3
Navigation X X 2
Data
gathering +
processing
X X 1
Rcv cmd,
send
telemetry
X X 2
Price per
unit
… … … … … …

Price Theory
Virtual Market-
place
Tasks Resources
1. Obstacle Avoidance
2. Navigation
3. Telemetry
4. Data
gathering/processing
1. Camera
2. Sensors
3. Propulsion
4. Processor
5. Radio
Demand Supply

Virtual Market-place
Battery Agent
Task
Agents
Resource
Agents
Acquisition
Resource Usage

Bidding, Price
Bn,t,r = min(( Alloct,n + Savet,n), (Bn-1,t,r + (Dt,r - St,r) * Pr)
• n is the round number
• t is the task
• r is the resource
Pr = (t Bt,r / Sr)
Supply of resource r is increased by transitioning to high energy mode.
This is done when Pr becomes too high.
Alloct,n captures the decision of the battery agent to allocate battery power
so that a maximum battery discharge rate is not crossed.

Resource Allocation
• How Res. are purchased using Virtual Credits?
o Task agents bid for Resources
39
D = 100
S = 50
T1
T2
D = 150
S = 50
S = 100
C1
• Price
– Price per Res
• $0.02 per
Res
• Resources
allocated
based on bids
– T1 and T2
get 50 Res
50
50
$1
$1
Price = ($1+$1)/100 = $0.02

Resource Allocation
• How Res are purchased using Virtual Credits?
o Task agents bids more due to supply-demand mismatch
40
D = 100
S = 50
T1
T2
D = 150
S = 50
S = 100
C1
• Price
– $0.035 per
Res
– Increased
• T1 gets
$1.5/0.035 = 43
units, T2 gets 57
units.
• Inflation
Scenario
$1.5
$2.0
Inflation: Supply < Demand
Resources Transition to High Energy
Mode

Resource Allocation
• How Res are purchased using Virtual Credits?
41
D = 100
S = 50
T1
T2
D = 150
S = 50
S = 300
C1
• Price per Res
– $0.012 per
Res
– Decreased
• T1 gets.
$1.5/$0.012 =
125 units, T2
gets. $2/0.012
= 166 units !
$1.5
$2.0
Deflation: Supply > Demand
Resources Transition to Low Energy
Mode
Can even shut down some of the units?

Resource Allocation
• How PUs are purchased using Virtual Money?
42
D = 100
S = 100
T1
T2
D = 150
S = 150
S = 250
C1
• Supply meets
Demand
– No change
– Price is constant
100 PUs
150 PUs
Stable: Supply = Demand
Energy Mode not changed

Resource Allocation
• Summary
o Task bids
• More – Under-Supplied
• Less – Over-Supplied
• Constant – Supply meets Demand
o Prevent Inflation/Deflation in prices
• By transitioning the resources between energy modes
• Maintain a maximum cutoff battery discharge rate
• Maintain expected battery life, while maximizing functionality
• Energy management of mobile devices in many operational scenarios.
43
Detailed example of distributed energy management in
drones can be seen from the accompanying paper.

In this talk
• Re-factoring
• Mobile SE based on field failures and reviews
• Energy management in other mobile devices.

Outlook
Exploiting user reviews in apps Energy management in drones

Mobilesoft 2017 Keynote

Recommended

Recommended

More Related Content

What's hot

What's hot (20)

Similar to Mobilesoft 2017 Keynote

Similar to Mobilesoft 2017 Keynote (20)

More from Abhik Roychoudhury

More from Abhik Roychoudhury (9)

Recently uploaded

Recently uploaded (20)

Mobilesoft 2017 Keynote