“ continuous all k nearest neighbor queries in smartphone networks ”

Dagstuhl Seminar 10042, Demetris Zeinalipour, University of Cyprus, 26/1/2010

MDM 2012 © Chatzimilioudis, Zeinalipour-Yazti, Lee, Dikaiakos 1

“Continuous All k Nearest Neighbor Queries in Smartphone Networks”

Tuesday, July 24, 201213th IEEE Int. Conference on Mobile Data Management

(MDM’12), Bangalore, India

Georgios Chatzimilioudis

Demetrios Zeinalipour-Yazti

Marios D. Dikaiakos

Wang-Chien Lee


MDM 2012 © Chatzimilioudis, Zeinalipour-Yazti, Lee, Dikaiakos

• Smartphone: A powerful sensing device!– Processing: 1.4 GHz quad core (Samsung Exynos)

– RAM & Flash Storage: 2GB & 48GB, respectively

– Networking: WiFi, 3G (Mbps) / 4G (100Mbps–1Gbps)

– Sensing: Proximity, Ambient Light, Accelerometer, Microphone, Geographic Coordinates based on AGPS (fine), WiFi or Cellular Towers (coarse).

• In-House Applications!

Smartphones

SmartTrace (ICDE’09,MDM’09,TKDE’12)

Airplace(MobiSys’12, MDM’12)

SmartP2P(MDM’11, MDM’12)

2



• Crowdsourcing with Smartphones– A smartphone crowd is constantly moving and sensing providing

large amounts of opportunistic data that enables new services and applications

Motivation

"Crowdsourcing with Smartphones", Georgios Chatzimiloudis, Andreas Konstantinides, Christos Laoudias, Demetrios Zeinalipour-Yazti, IEEE Internet Computing, Special Issue: Crowdsourcing (Sep/Oct 2012), accepted May 2012. IEEE Press, 2012

3



Continuous k Nearest Neighbor Queries

Find 2 Closest Neighbors for 1 User



Motivation

5

“Create a Framework for Efficient Proximity Interactions”

Screenshots from a prototype system we've implemented for Windows Phonehttp://www.zegathem.com/



Continuous All k Nearest Neighbor Queries

Find 2 Closest Neighbors for ALL User



• Motivation• System Model and Problem

Formulation• Proximity Algorithm• Experimental Evaluation• Future Work

7

Presentation Outline



• A set of users moving in the plane of a region (u1, u2, … un)

• Area covered by a set of Network Connectivity Points (NCP)– Each NCP creates the notion of a cell

– W.l.o.g., let the cell be represented by a circular area with an arbitrary radius

• A mobile user u is serviced at any given time point by one NCP.

• There is some centralized service, denoted as QP (Query Processor), which is aware of the coverage of each NCP.

• Each user u reports its positional information to QP regularly

8

System Model

C

.u7

.u4

u5.

u2.

u6.

u0.

u3.

.u1

Query Processor

QP



QueryProcessor

• Definition of the CAkNN problem:Given a set U of n users and their location reports ri,t R∈ at timestep t T ∈ , then the CAkNN problem is to find in each timestep t T ∈ and for each user ui U ∈ the k objects Usol U − u⊆ i such that for all other objects uo ∈ U − Usol − ui, dist(uk, ui) ≤ dist(uo, ui) holds

9

Problem Definition

C

.u7

.u4

u5.

u2.

u6.u0.

u3.

.u1



Query Processor

10

Naïve Solution / Related Work

Find 2-NN for u0 at timestep t.

C

.u7

.u4

u5.

u2.

u6.u0.

u3.

.u1

For u5?

Look inside your cell!

Perform iterative deepening!

TOO EXPENSIVE!

WRONG!(u1 closer)



Related Work• Existing algorithms for CkNN (not CAkNN)

Yu et al. (YPK) 11 and Mouratidis et al. (CPM) 12

• Stateless Version: Iteratively expand the search space for each user into neighboring cells to find the kNN (like previous slide)

• Stateful Version: Improve the stateless version by utilizing previous state, under the following assumptions:

– i) Static Querying User (i.e., designated for point queries)– ii) Target users move slowly (i.e., state does not decay)– iii) Few Target users

11. Yu, Pu, Koudas. “Monitoring k-nearest neighbor queries over moving objects,” ICDE ’0512. Mouratidis, Papadias, Hadjieleftheriou, “Conceptual partitioning: an efficient method for continuous nearest neighbor monitoring,” SIGMOD ’05.

Not Appropriate

for CAkNN





12




Proximity Overview

• The first specialized algorithm for Continuous All k-NN (CAkNN) queries

• Important characteristics:– Stateless (i.e., optimized for high mobility)– Batch processing (i.e., with search space

searching)– Parameter-free (i.e., no tuning parameters)

• Generic operator for proximity-based queries

– See Crowdcast application presented later.

13



QueryProcessor

C

.u7

.u4

u5.

u2.

u6.u0.

u3.

.u1

Proximity Outline• For every timestep:

1. Initialize a k+-heap for every cell

2. Insert every user’s location report to every k+-heap• Notice that k+-heap is a heap-based structure and most location

reports will be dropped as a result of an insert operation

3. For every user scan the k+-heap of his cell to find his k-NN

14



QueryProcessor

C

.u7

.u4

u5.

u2.

u6.

u0.

u3.

.u1

Intuition behind Proximity• Users in a cell will share the same search

space (search space sharing)• Compute 1 search space per cell only!

15

Search Space(contains the right answers for all users in

C)

Cell



• The k+-heap structure for a cell

Proximity k+-heap

16

All users inside the

cell

K nearest users outside the cell

Beyond k nearest outside users (correctness)

k+-heap structure

Ktop-k heap

Bordered list

Ulist



k+-heap structure

Ktop-k heap

Bordered list

Ulist

k+-heap Construction (for Cell C)

17

ArrivingReports

StructureU

StructureK

StructureB

u6 u6

u4 u6 u4

u7 u6 u7, u4

u2 u6 u4, u2 u7

u3 u6 u3, u2 u4, u7

u1 u6 u2, u1 u3, u4

u5 u6 u2, u1 u3, u4, u5

u0 u6, u0 u2, u1 u3, u4, u5

Assume k=2.

C

.u7

.u4

u5.

u2.

u6.

u0.

u3.

.u1



k+-heap structure

Ktop-k heap

Bordered list

Ulist

k+-heap Construction (for Cell C)

18

ArrivingReports

StructureU

StructureK

StructureB

u6 u6

u4 u6 u4

u7 u6 u7, u4

u2 u6 u4, u2 u7

u3 u6 u3, u2 u4, u7

u1 u6 u2, u1 u3, u4

u5 u6 u2, u1 u3, u4, u5

u0 u6, u0 u2, u1 u3, u4, u5

Assume k=2.

C

.u7

.u4

u5.

u2.

u6.

u0.

u3.

.u1

2 nearest users to cell

x.

Diameter +

d k





19




Experimental Methodology

• Dataset:– Oldenburg Dataset:

• Spatiotemporal generator with roadmap of Oldenburg as input (25km x 25km)• 5000 maximum users (Oldenburg population: 160.000)

– Manhattan Dataset:• Vehicular mobility generator with roadmap of Manhattan as input (3km x 3km)• 500 users

• Network connectivity points (NCPs) uniformly in space– Communication ranges 1km, 4km and 16km for the Oldenburg

dataset and ranges 1km and 4km for the Manhattan dataset

• Query: CAkNN• Comparison: Proximity vs YPK vs CPM (using the

optimal parameter value for YPK and CPM)

20



Proximity - Experiments

• Bottleneck for Proximity: build time• Bottleneck for the adapted YPK and CPM: search time

21




• Benefit of Proximity in the search time• Proximity is scalable with respect to k

– Search space for Proximity is not proportional to k like YPK, CPM

22




• Proximity: build time drops as k increases!• Machine’s memory scheduler makes more efficient use of

buffers when the search spaces are larger

23




• Proximity scales with the number of users• Proximity outperforms YPK, CPM by an order of magnitude• Proximity does not converge to YPK, CPM for higher values of Nmax.

24



Crowdcast Suite

The Crowdcast Application Suite

PROXIMITY

WIN7 API

…

Ranked 3rd in Microsoft’s ImagineCup contest local competition (to appear in the next demo session!)



Crowdcast: MsgCast

"Get your location-based questions answered!"

MsgCast- Location-based Chat Channel (Post / Follow)- Provide Guidelines



Crowdcast: HelpCast

HelpCast"The Ubiquitous Help Platform for Anyone in Need”



Crowdcast: EyeCast

"Extend your Vision beyond your 2 eyes!"

EyeCast- Disaster Recovery Operations- Indoor Video Conferencing Network





29




Proximity – Future Work

• Privacy extensions (e.g., spatial cloaking, strong privacy)

• Parallelizing server computation

• Proximity for cloud server

• User-defined k values

30



Future Work: SmartLab

31

SmartLab.cs.ucy.ac.cy

Install APK, Upload File,Reboot, Screenshots, Monkey Runners, etc.…

Programming cloud for the development of smartphone network applications & protocols as well as experimentation with real smartphone devices.

"Demo: A Programming Cloud of Smartphones", A. Konstantinidis, C. Costa, G. Larkou and D. Zeinalipour-Yazti, "Demo at the 10th International Conference on Mobile Systems Applications and

Services" (Mobisys '12), Low Wood Bay Lake District UK, 2012



Future Work: SmartLab

32



“Continuous All k Nearest Neighbor Queries in Smartphone Networks”

Tuesday, July 24, 201213th IEEE Int. Conference on Mobile Data Management

(MDM’12), Bangalore, India

Georgios Chatzimilioudis

Demetrios Zeinalipour-Yazti

Marios D. Dikaiakos

Wang-Chien Lee

Thanks!Questions?

“ continuous all k nearest neighbor queries in smartphone networks ”

Documents