+ All Categories
Home > Documents > CrowdBind: Fairness Enhanced Late Binding Task Scheduling ... · Slide 13/29 5 Design Features of...

CrowdBind: Fairness Enhanced Late Binding Task Scheduling ... · Slide 13/29 5 Design Features of...

Date post: 27-Jul-2020
Category:
Upload: others
View: 6 times
Download: 0 times
Share this document with a friend
29
Slide 1/29 CrowdBind: Fairness Enhanced Late Binding Task Scheduling in Mobile Crowdsensing Heng Zhang, Michael A Roth, Rajesh K. Panta (AT&T), He Wang, Saurabh Bagchi Dependable Computing Systems Lab (DCSL) @ School of Electrical and Computer Engineering Purdue University
Transcript
Page 1: CrowdBind: Fairness Enhanced Late Binding Task Scheduling ... · Slide 13/29 5 Design Features of CrowdBind 1.Late binding: A task submitted by a MCS application does not need to

Slide 1/29

CrowdBind: Fairness Enhanced Late Binding Task Scheduling in Mobile Crowdsensing

Heng Zhang, Michael A Roth, Rajesh K. Panta (AT&T), He Wang, Saurabh Bagchi

Dependable Computing Systems Lab (DCSL) @ School of Electrical and Computer Engineering

Purdue University

Page 2: CrowdBind: Fairness Enhanced Late Binding Task Scheduling ... · Slide 13/29 5 Design Features of CrowdBind 1.Late binding: A task submitted by a MCS application does not need to

Slide 2/29

Outline• Background• Challenges• Motivation Example• CrowdBind Design• Implementation & Evaluation• Conclusion

Page 3: CrowdBind: Fairness Enhanced Late Binding Task Scheduling ... · Slide 13/29 5 Design Features of CrowdBind 1.Late binding: A task submitted by a MCS application does not need to

Slide 3/29

What is mobile crowdsensing (MCS)Build temperature map using temperature sensors on smartphones at different geographic locations

Image adapted from https://www.wunderground.com/

Page 4: CrowdBind: Fairness Enhanced Late Binding Task Scheduling ... · Slide 13/29 5 Design Features of CrowdBind 1.Late binding: A task submitted by a MCS application does not need to

Slide 4/29

What is mobile crowdsensing (MCS)

A method of collecting environmental information by coordinating smartphone users at different locations

Page 5: CrowdBind: Fairness Enhanced Late Binding Task Scheduling ... · Slide 13/29 5 Design Features of CrowdBind 1.Late binding: A task submitted by a MCS application does not need to

Slide 5/29

MCS PlayersMCS

Application

MCS Scheduler

MCS User

Temperature Map, City Noise Map

CrowdBind

Smartphone users who sign up for MCS campaign

Page 6: CrowdBind: Fairness Enhanced Late Binding Task Scheduling ... · Slide 13/29 5 Design Features of CrowdBind 1.Late binding: A task submitted by a MCS application does not need to

Slide 6/29

Outline• Background• Challenges• Motivation Example• CrowdBind Design• Implementation & Evaluation• Conclusion

Page 7: CrowdBind: Fairness Enhanced Late Binding Task Scheduling ... · Slide 13/29 5 Design Features of CrowdBind 1.Late binding: A task submitted by a MCS application does not need to

Slide 7/29

Challenges in MCS

For MCS applications:• Task coverage [Ubicomp-14]

• Data collection cost [Percom-2015]

For MCS users:• Reward [Mobicom-12]

• Detour [MASS-15]

• Smartphone energy cost [SenSys-13]

• Task Fairness?Metric → Gini Coefficient: inequality measure between 0 and 1. 0 means the fairest Gini(1,1,...,1) = 0 and 1 is the most unfair Gini(1,0,...,0,...,0) ≆ 1

Definition: Whether MCS users receive fairly equal number of MCS tasks.Ideally, we want each user to receive the same number of tasks

Page 8: CrowdBind: Fairness Enhanced Late Binding Task Scheduling ... · Slide 13/29 5 Design Features of CrowdBind 1.Late binding: A task submitted by a MCS application does not need to

Slide 8/29

Survey

MCS users do care about fairness. However, current literature only mention it is a desirable feature but do not optimize for it

Page 9: CrowdBind: Fairness Enhanced Late Binding Task Scheduling ... · Slide 13/29 5 Design Features of CrowdBind 1.Late binding: A task submitted by a MCS application does not need to

Slide 9/29

Contribution

Propose two variants of greedy algorithms, each of which has different pros and cons. We evaluate CrowdBind using a real-world user study of 50 students as well as a simulation of 90K for scalability study.

We share the user study mobility trace for future research and it is publicly available at https://github.com/LLADzhang/CrowdBind

CrowdBind: the first MCS scheduler which optimizes the scheduling fairness without degrading two other primary objectives -- task coverage and the average energy cost of completing a task among users

Page 10: CrowdBind: Fairness Enhanced Late Binding Task Scheduling ... · Slide 13/29 5 Design Features of CrowdBind 1.Late binding: A task submitted by a MCS application does not need to

Slide 10/29

Outline• Background• Challenges• Motivation Example• CrowdBind Design• Implementation & Evaluation• Conclusion

Page 11: CrowdBind: Fairness Enhanced Late Binding Task Scheduling ... · Slide 13/29 5 Design Features of CrowdBind 1.Late binding: A task submitted by a MCS application does not need to

Slide 11/29

Terminology

Task: Issued by MCS application and completed by MCS user; Usually specifies sensors, location, range, number of users needed, what time, periodicity, etcTask Instance: For a periodic task, one instance is created for each period of the taskExample: A MCS application issues a task of collecting temperature values around our conference building. It asks N users to report the temperature every hour (periodicity). The task is from 8 am to 10 am. Then 2 task instances. One requires users to upload before 9 am and another one is before 10 am.

Page 12: CrowdBind: Fairness Enhanced Late Binding Task Scheduling ... · Slide 13/29 5 Design Features of CrowdBind 1.Late binding: A task submitted by a MCS application does not need to

Slide 12/29

TerminologyTask Coverage:

N: the total number of task instances for a given taskSi: the number of actually selected users for instance iMinimum Device Requirement (MDR): the minimum number of users required for each task instanceExample: As from the previous example, suppose each of the two task instances has a MDR of 4. If both eventually select 4 users (S1&S2= 4), the task coverage is (4/4+4/4)/2 = 1. If the task instance at 9 am only finds 2 users (S1=2), the coverage reduces to (2/4+4/4)/2=0.75

Page 13: CrowdBind: Fairness Enhanced Late Binding Task Scheduling ... · Slide 13/29 5 Design Features of CrowdBind 1.Late binding: A task submitted by a MCS application does not need to

Slide 13/29

5 Design Features of CrowdBind

1. Late binding: A task submitted by a MCS application does not need to be scheduled right away but it can wait till shortly before each of the task instances to be scheduled

2. Delay tolerance: e.g. the barometer reading can be taken now or within the next 10 minutes

3. Spatial tolerance: A user within some radius of a region can perform a task (e.g. circle with r = 100m)

Page 14: CrowdBind: Fairness Enhanced Late Binding Task Scheduling ... · Slide 13/29 5 Design Features of CrowdBind 1.Late binding: A task submitted by a MCS application does not need to

Slide 14/29

5 Design Features of CrowdBind4. Regular human mobility: People have regular mobility

based on their daily schedule → mobility prediction

4. Lookahead: Schedule multiple task instances in a certain lookahead window

Accuracy is higher with deeper color

Page 15: CrowdBind: Fairness Enhanced Late Binding Task Scheduling ... · Slide 13/29 5 Design Features of CrowdBind 1.Late binding: A task submitted by a MCS application does not need to

Slide 15/29

Illustrative Example

U1, U2, U3, U4 are 4 users

TI1 and TI2 are two instances of a task each needs at least 2 users to complete (MDR = 2)

highlights the time interval that the user is predicted to reside in the region of the task instance

Without mobility prediction and lookahead:TI1: (U1, U3) or (U1, U2) or (U2, U3) TI2: (U1, U4)U1 may be picked TWICE! → unfair

With CrowdBind: TI1: (U2, U3)TI2: (U1, U4)Each user is fairly selected once and both TI are satisfied.

Page 16: CrowdBind: Fairness Enhanced Late Binding Task Scheduling ... · Slide 13/29 5 Design Features of CrowdBind 1.Late binding: A task submitted by a MCS application does not need to

Slide 16/29

Outline• Background• Challenges• Motivation Example• CrowdBind Design• Implementation & Evaluation• Conclusion

Page 17: CrowdBind: Fairness Enhanced Late Binding Task Scheduling ... · Slide 13/29 5 Design Features of CrowdBind 1.Late binding: A task submitted by a MCS application does not need to

Slide 17/29

Architecture

Core greedy algorithms

Page 18: CrowdBind: Fairness Enhanced Late Binding Task Scheduling ... · Slide 13/29 5 Design Features of CrowdBind 1.Late binding: A task submitted by a MCS application does not need to

Slide 18/29

Problem Formulation● User selection under certain constraints has been proved to be

NP-hard in MCS● We formulate it as a minimization problem subject to three

conditions

U: the usage of usersUti: the set of users selected for task instance tiT: the set of tasksTI: the set of task instancesMBRti: minimum battery requirement (e.g. at least 80% to be eligible for selection)CCt: Think of it as the sum of predicted number of available users for each instance of a taskMCCt: Think of it as the sum of minimum needed number of users for each instance of a taskMore details of CCt and MCCt can be found in the paper

Page 19: CrowdBind: Fairness Enhanced Late Binding Task Scheduling ... · Slide 13/29 5 Design Features of CrowdBind 1.Late binding: A task submitted by a MCS application does not need to

Slide 19/29

Two Greedy Scheduling Algorithms

G-Random-X vs G-Heuristic

Different Initialization Phase and same Optimization Phase

Key to optimize fairness: Look all task instances within the next T minutes, predict the available users and map those users to those instances so that their usages are equally distributed

Page 20: CrowdBind: Fairness Enhanced Late Binding Task Scheduling ... · Slide 13/29 5 Design Features of CrowdBind 1.Late binding: A task submitted by a MCS application does not need to

Slide 20/29

Initialization: G-Random-X

• X: X rounds of execution with different random seeds

• In each round, the initialization will randomly select MDR number of users for each of the task instance

– The optimization stage is followed to satisfy the constraints in the problem formulation by altering the initial user selection

• Choose the fairest scheduling out of the X runs

Page 21: CrowdBind: Fairness Enhanced Late Binding Task Scheduling ... · Slide 13/29 5 Design Features of CrowdBind 1.Late binding: A task submitted by a MCS application does not need to

Slide 21/29

Initialization: G-Random-Heuristic

• More complex initialization: we want it to reach the fairest selection faster by a smart initial selection of users

• Key: for each task instance, select users based on certain properties (mobility predictability, sensor integrity, probability to complete a task instance, usage history). Calculate weight of each user based on those properties

• More technical details in the paper

Page 22: CrowdBind: Fairness Enhanced Late Binding Task Scheduling ... · Slide 13/29 5 Design Features of CrowdBind 1.Late binding: A task submitted by a MCS application does not need to

Slide 22/29

• Key: greedily replace some selected users with other unselected users with higher weights

• End until no better replacement can be found

• More technical details in the paper

Optimization: Both

Page 23: CrowdBind: Fairness Enhanced Late Binding Task Scheduling ... · Slide 13/29 5 Design Features of CrowdBind 1.Late binding: A task submitted by a MCS application does not need to

Slide 23/29

Trade-offs

A total of 9 tasks and 6 instances of each task. MDR = 3. The same 9 tasks are repeated 100 times to draw statistically valid results

● G-Random-X runtime is proportional to X.

● X >=10, G-Random-X is more fair and has higher coverage than heuristic

● Heuristic has relative good performance and very fast

● The larger the X, the more likely to achieve global optima <--> Heuristic may initialize to a local optima

Page 24: CrowdBind: Fairness Enhanced Late Binding Task Scheduling ... · Slide 13/29 5 Design Features of CrowdBind 1.Late binding: A task submitted by a MCS application does not need to

Slide 24/29

Outline• Background• Challenges• Motivation Example• CrowdBind Design• Implementation & Evaluation• Conclusion

Page 25: CrowdBind: Fairness Enhanced Late Binding Task Scheduling ... · Slide 13/29 5 Design Features of CrowdBind 1.Late binding: A task submitted by a MCS application does not need to

Slide 25/29

Implementation & Evaluation

• User study:– 50 students, one month– Location data is received every 5 mins– 213 tasks and 3,916 task instances– Compare with 4 prior works: Periodic [baseline], PCS [SenSys-13],

Sense-Aid [Middleware-17], CrowdRecruiter [Ubicomp-14]– Each protocol is implemented as a standalone Android app

• Simulation:– Gowalla mobility dataset with 90K users– One month data is used– 112,920 synthetical tasks at different locations

Page 26: CrowdBind: Fairness Enhanced Late Binding Task Scheduling ... · Slide 13/29 5 Design Features of CrowdBind 1.Late binding: A task submitted by a MCS application does not need to

Slide 26/29

Comparison with prior works and baseline

Simulation:

User study results are consistent with these in the simulation. We also show the user study results of changing delay tolerance and MDR in the paper

CrowdBind algorithms are the fairest, cost the least energy, and achieve near optimal task coverage

Page 27: CrowdBind: Fairness Enhanced Late Binding Task Scheduling ... · Slide 13/29 5 Design Features of CrowdBind 1.Late binding: A task submitted by a MCS application does not need to

Slide 27/29

Outline• Background• Challenges• Motivation Example• CrowdBind Design• Implementation & Evaluation• Conclusion

Page 28: CrowdBind: Fairness Enhanced Late Binding Task Scheduling ... · Slide 13/29 5 Design Features of CrowdBind 1.Late binding: A task submitted by a MCS application does not need to

Slide 28/29

Conclusion

Key to CrowdBind is to allocate task instances to as many users as possible so that inequality (Gini coefficient) and average energy cost are minimized without damaging the task coverage

Page 29: CrowdBind: Fairness Enhanced Late Binding Task Scheduling ... · Slide 13/29 5 Design Features of CrowdBind 1.Late binding: A task submitted by a MCS application does not need to

Slide 29/29

Q&A

Thank you!

Further discussion: [email protected]


Recommended