Page 1
•Sharanya Eswaran, Penn State University•Matthew Johnson, CUNY•Archan Misra, Telcordia Technolgoies•Thomas La Porta, Penn State University
Utility-driven Energy-aware In-network Processing for Mission-oriented Wireless
Sensor Networks
Annual Conference of ITA (September 24, 2009)
Page 2
The Problem
. . .
Sensor Resources Network ResourcesMissions/Applications
“How to share the network resources (bandwidth, energy) to maximize the effectiveness of sensor-enabled applications (missions)?”
• Limited bandwidth• Limited energy • Heterogeneous missions utilizing multiple types of sensors• Variable degrees of in-network processing
- Forwarding nodes may compress or fuse data
Perimeter monitoring
Gunfire localization
Mobile insurgent tracking
Surveillance
...Image fusion
Correlation
Page 3
In-network Processing• In-network processing is an attractive option conserving bandwidth
and energyo Compression o Fusion
• Non-negligible energy footprint for streaming applications• Stream-oriented data comprise sophisticated DSP-based operations
(e.g., MPEG compression, wavelet coefficient computation)
• Forwarding nodes can compress on the flyo With variable compression ratios
• Forwarding nodes can fuse multiple streamso the location of these fusion points can be determined on the fly
• Dual trade-off o Bandwidth vs. loss of informationo Communication cost vs. computation cost
Page 4
Adaptive In-network Processing
• Variable quality compression– Each forwarding node compresses data to different ratios, depending
on• Residual energy at that and downstream nodes• Congestion in the region• Effect of compression on application
• Dynamic fusion operator placement– Select best node in the path each time for fusion, depending on
• Residual energy at that and downstream nodes• Congestion in the region
• Variable source rate
1 2
A
C
M
B
Page 5
Our Approach
Each mission has a “utility”:• A measure of how “happy” the mission is• A function of rates received from all its sensors
Allocate WSN resources (bandwidth and energy of
nodes) to maximize cumulative utility.
Network Utility Maximization (NUM) A Distributed, Utility-Based Formulation of Resource Sharing
Objective: “Joint Congestion and Energy Control for Network Utility Maximization”
Page 6
Optimization Problem
Mm knodes
ktotmsets
recsm PxU
,)( )}({ maximize
Qqc
kixqik ki
out
,1),(:constraintCapacity i)
subject to
),(
kcomp
ktrans
krec
ktot
kktot
PPPP
knodesPP
where
,
:constraintEnergy ii)
max,
Page 7
Background: WSN-NUM Model
Airtime constraint over “transmission-specific” cliques
Cliques => “contention region” No two transmissions in a clique
can occur simultaneously Llcx
)(XU
LUSENSOR
l(k,s) k,s
s
Mmmm
clique maximaleach for
1
:subject to
maximize:),(
Connectivity graph Multicast trees (with broadcast transmissions)
Transmission-basedConflict graph
2 1 3
4 5
m1 m2 m3
Page 8
WSN-NUM Protocol Price-based, iterative, receiver-
centric scheme
Solve two independent sub-problems
• Network nodes: • Aim to maximize “revenue”• Compute Clique cost: degree of
congestion in the clique• Flow cost = sum of costs of all cliques
along the flow
• Mission (sink): • Aims to maximize its utility minus the
cost• Sends path cost to each source• Sends ‘willingness to pay’ for each
source
• Sensor (source):• Adjusts rate to drive gradient to zero
.0over
, cliqueeach for ,1 subject to
);log( maximize:);(
),( ,
s
sk sk
s
Ss Mmsms
x
Llcx
xwwLNETWORK
2
),( ,
1)()(
lsk sk
sl c
txt
s
msms x
Utxw
)(
)( ),( ,)(
)1*)((*)()()(sflowl lsk sk
lssMissm
mss cttxtwtx
dtd
ss to pay""willingnexwx
wwU
USINK
smsms
s
msetsms
m
mm
mmm
:*.0over
)( maximize
:);(
)(
(1)
(2)
(3)
(4)
Page 9
Distributed Solution for INP-NUM
)( ),(
),()(sMissm k q qsk sks
outq
s
ktot
ks
ms
s
xCksx
xP
xUx
dtdx
)( ),(
),()(iMissm v q qiv kivi
outq
ki
vtot
vki
mki
ki
lCvix
lP
lUl
dtdl
Impact on utility
1 2
A
C
M
B
At each source:
Energy cost Congestion cost
At each forwarding node:
Impact on utility Energy cost Congestion cost
• Two penalty values:- Congestion cost, µ- Energy cost, η
Page 10
Adaptive Operator Placement
• We assume that fusion can be shared across multiple nodes– Can be thought of as time-sharing
• Each candidate node fuses a fraction (θ) of the flow– Sink receives multiple sub-flows, each fused at a different node
• Optimize θ such that fusion is most efficient
1 2
A
C
M
B
)( ),( ,,,
,, ),()(
iMissm v q qivk
sopvi
outqk
sop
vtot
vksop
mksop
ksop
CvixPU
dtd
Page 11
1
A
m
2
B
C
Flow 1: x1Flow 2: x2
),( 2211 xxf AA Afl
11 )1( xA1Al
22 )1( xA2Al
`
Illustration of INP-NUM
Fused flow f
Page 12
Challenges in INP-NUM Protocol
• Missions do not know about original flow and the transformations (compression and fusion)
• Fusion placement and compression ratio adaptation require different sets of data.
• Feedback received and processed by each forwarding node in the path– It is modified before forwarding upstream
• If it is a fusion point, it updates the feedback to include the effect of fusion– Based on chain rule of differentiation
dxdx
dxdx
dxdx
dxdU
dxdU nrec
rec
2
1
1
Page 13
Illustration of INP-NUM Feedback
1
A
m
B
ff x
Ux *pay toswillingnes m
fCf
Bf
Aff xxxxx
C
2
21
2fA
fB
Cumulative Info
21
2fA
Cumulative Info
Rate Info Energy Info Congestion Info
Rate Info Energy Info Congestion Info
1
Rate Info Energy Info Congestion Info
Cumulative Info
2 Rate Info Energy Info Congestion Info
Cumulative Info
Page 14
Addressing Practical Constraints
• Often in reality, fully elastic compression may not be possible
– Only discrete levels of compression
• E.g., JPEG allows 100 discrete values for compression ratio, video may be
encoded in a finite set of bitrates depending on the encoding technique
• Similarly, partial fusion may not be feasible
– Fusion operation may need to take place at a solitary node.
• NP-hard to solve both problems without these assumptions
• We can use approximation heuristics
• Determine nearest valid compression ratio
• Pick node with most responsibility for solitary fusion
Page 15
Evaluation
High Utility Medium Utility Low Utility
Page 17
Effect of Discretization
Page 18
Conclusion
• Protocol for adaptive compression and fusion placement– Fully distributed– Low overhead– Provably optimal utilization of bandwidth and energy
• Heuristics for realistic constraints provide near-optimal solution
• In future, we will develop a model taking lifetime requirements of missions into account