MobiSys 2012
Medusa: A Programming Framework
for Crowd-Sensing Applications
Moo-Ryong Ra, Bin Liu
Tom La Porta, and Ramesh Govindan
2
Crowd-Sensing
Crowds
Sensing
2016
2012Camera/Mic.
GPS/WPS
Accelerometer
Compass/Gyro
Growing Proximity
Smartphone Users
Motivation
Problem
Design
Evaluation
Conclusion
Example: Documenting Social Events
User
3
Summaries
Full Videos
Worker
Worker
Worker
Worker
Worker
Worker
Motivation
Problem
Design
Evaluation
Conclusion
4
Other Crowd-Sensing Tasks
Auditioning
Collaborative
Learning
Forensic
Analysis
Motivation
Problem
Television producer
wants potential actors
to submit clips of
their acting.
Mobile app developer
wants samples
of
Security
personnel
activities
train
wants
to findtopictures
classifier.
of attack
taken by
mall visitors.
Design
Evaluation
Conclusion
5
The Problem
Recruiting Workers
Stage 1
Providing Incentives
Coordinating worker actions
Can be
extremely
Stage
2
labor-intensive
Stage 3
Gathering contributions
Stage 4
Automation using a high-level programming system
Motivation
Problem
Design
Evaluation
Conclusion
6
Contributions
Describe crowd-sensing requirements
Devise design principles
Design and implement Medusa,
a programming system for crowd-sensing
Evaluate using qualitatively different tasks
Motivation
Problem
Design
Evaluation
Conclusion
7
Crowd-Sensing Requirements
Expressivity
Runtime
Worker Mediation
Privacy
Concurrent Tasks
Unsynchronized Execution
Anonymity
Robustness
Sandboxing
Secure Communication
Resource Usage Controls
Access Stored Data
In-network Processing
Extensibility
Timeliness Support
Incentives
Motivation
Problem
Design
Evaluation
Conclusion
Runtime Design Principles
8
Partitioned Services
Task Management
Worker Recruiting
Incentive Management
Cloud
Runtime
Cloud (Server)
Opt-in Data
Transfer
Mobile Devices
Sensing
Client
Runtime
In-network Processing
Dumb Smartphone
Motivation
Problem
Design
Evaluation
Conclusion
Medusa Design
9
MedScript
Programming Language
Medusa Runtime
Motivation
Problem
Design
Evaluation
Conclusion
Expressivity of MedScript Language
Video
Documentation
Trigger
SPC stage
Preview outgoing data
Processing
Sensing
Recruit
HIT stage
TakeVideo
ExtractSummary
Communication
UploadSummary
Incentives
10
Curate
UploadVideo
Worker Mediation
XML-based extensible language
Motivation
Problem
Design
Evaluation
Conclusion
11
Other Language Features
Location-based
Fork-join
Reverse
Authoring
Incentives
construct
Tool
Trigger
MedAuthor
Triggering
Stage Execution
GetAccel.
Start
Join
GetGPS
Reverse
Typical
Motivation
Problem
Design
New!
Evaluation
Conclusion
12
Task Execution Example
Resource
Cloud-Initiated
UsageExecution
Controls
Stage
Task
Initiation
Alice
User
Create
‘Recruit’
Medusa
Initiate
‘TakeVideo’ …
Initiate
‘UploadVideo’
Unsynchronized
Runtime
Execution
Sign-up to
Worker
Sign-up to
‘Recruit’
Charlie
Motivation
Sending the Video
to Alice
Taking … Uploading
Video
Video
‘Recruit’
Bob
Got
the Video
Taking
Uploading
…
Video
Video
Worker
Problem
Opt-in
Privacy
Time Flow
Design
Evaluation
Conclusion
13
Medusa Runtime Design
Data
Repository
Interpreted
MedScript Program
Worker
Manager
Stage
Library
Task Tracker
C2DM
Data Transfer
Cloud (Server)
Mobile Devices
Stage Tracker
MedBox
Bob
Motivation
stage
Upload
Manager
Sensor
Manager
Transform
Manager
Storage
Manager
Problem
Amazon
Mechanical
Turk
Design
stage
…
Charlie
Evaluation
Conclusion
14
Medusa Prototype
Python + LAMP
Java Servlet + Tomcat
Implementation
Android
10 crowd-sensing tasks
Expressivity
Evaluation
Concurrent Execution
Performance
Robustness
Motivation
Problem
Design
Evaluation
Conclusion
15
Expressivity
Application
LOC
Sensors
Properties
Video Documentation
86
vs Camera
8,763
In-network Processing, Crowd Curation
Collaborative Learning
62
Accel.,
vs
7,076Mic.
Different Sensors
Auditioning
86
Camera
Reverse Incentives, Crowd Curation
Fewer than 100 LOC for tasks.
Forensic Analysis
86
Spot Reporter
45
Road Monitoring
90
Accel.
Citizen Journalist
45
GPS
Party Thermometer
62
GPS, Mic.
WiFi/Bluetooth Scanner
45
Net. Sensors
Motivation
Problem
GPS, Mic.
Access to Stored Data
Camera,
vs
17,238Mic. Text/Voice Tagging
Fork-join Construct (from PRISM, MobiSys’10)
Multiple Triggers (from PRISM, MobiSys’10)
Location triggering (from PRISM, MobiSys’10)
From AnonySense, MobiSys’08
Design
Evaluation
Conclusion
16
Concurrent Execution
19h08m 19h10m
19h12m
Tasks
19h14m
19h18m
19h20m
2
Worker subscribed to the Task
s
(1)
(2)
(3)
19h16m
19h22m
s
s
s
19h32m
Worker I
s
(4)
(5)
(6)
s
Worker II
s
(7)
(8)
s
Stage Execution
Worker III
s
(9)
(10)
s
Worker IV
s
Command Arrival via SMS
Time Flow
4
Concurrent
Workers
Each
Eachtask
stage
may
execution
has
start
has
a Different
aeach
different
onstage
both
number
execution
at
Medusa
different
of stages.
and
time.
times.
users
Motivation
Problem
Design
Evaluation
Conclusion
17
System Performance
Cloud Runtime
Overhead
Mobile Device
Medusa
runtime
Runtime
Overheads
has very small overheads.
Notification
Delay
~30ms
delay,
5ms
for
tracking
stage
Total initialization
runtime
C2DM gives
delay
much
is only
less
100ms
latency
per stage.
Motivation
Problem
Design
Evaluation
Conclusion
18
Related Work
Crowd-Sourcing
Smartphone Sensing
Amazon Mechanical Turk
PEIR, MobiSys’09
CrowdDB, SIGMOD’11
SoundSense, MobiSys’09
Turkit, UIST’10
AnonySense, MobiSys’08
CrowdSearch, MobiSys’10
PRISM, MobiSys’10
Human-in
-the-loop
Motivation
Incentives MedusaIn-network
Processing
Problem
Design
Evaluation
Sensing
Conclusion
Summary
Medusa: A High-Level Programming Framework
for Crowd-Sensing Applications
a. Identified requirements and principles to design high-level
programming framework for crowd-sensing applications.
b. Design and implementation of the Medscript programming
language and associated runtime system.
c. Prototyped 10 crowd-sensing applications.
Code
Demo Video
http://code.google.com/p/medusa-crowd-sensing/
http://youtu.be/jL1dGA21ciA
19