algorithmic aspects of throughput-delay performance for...
TRANSCRIPT
Algorithmic Aspects of Throughput-Delay Performance for FastData Collection in Sensor Networks
Amitabha GhoshCommittee:
Bhaskar KrishnamachariC.S. Raghavendra
Gaurav S. Sukhatme
Autonomous Networks Research GroupMing Hsieh Department of Electrical Engineering
USC Viterbi School of [email protected]
http://anrg.usc.edu/˜amitabhg
July 27, 2010
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Thesis Statement:
Multi-channel scheduling, optimal routing topologies, and transmission powercontrol are necessary in order to improve the throughput-delay trade-off foraggregated convergecast in large-scale, multi-hop, wireless sensor networks.
2 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Outline
1 Maximizing Convergecast ThroughputMotivation and PreliminariesScheduling with Unlimited FrequenciesApproximations on Unit Disk GraphsApproximations on General Disk Graphs
2 Multi-Channel Scheduling in SINRMotivation and Interference ModelsProblem Formulation and ApproachApproximations in SINR
3 Throughput-Delay Trade-offMotivationBi-criteria FormulationOptimal Spanning Trees (parts)
4 Distributed Topology Control in 3-DMotivationTwo Approaches: Orthographic Projections, SDT
3 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Maximizing Convergecast Throughput
Outline
1 Maximizing Convergecast ThroughputMotivation and PreliminariesScheduling with Unlimited FrequenciesApproximations on Unit Disk GraphsApproximations on General Disk Graphs
2 Multi-Channel Scheduling in SINRMotivation and Interference ModelsProblem Formulation and ApproachApproximations in SINR
3 Throughput-Delay Trade-offMotivationBicriteria FormulationOptimal Spanning Trees
4 Conclusions
4 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Maximizing Convergecast Throughput
Motivation and Preliminaries
Model and Assumptions
Secondary conflict
e.g., Applications with fast &efficient data collectionrequirements.
5 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Maximizing Convergecast Throughput
Motivation and Preliminaries
Model and Assumptions
Secondary conflict
e.g., Applications with fast &efficient data collectionrequirements.
TDMA periodic scheduling(- every node generates a single packet at thebeginning of each frame- every node transmits once per slot)
Graph-based network & interference model
Half-duplex, single transceiver
Receiver-based Frequency Assignment
Non-interfering orthogonal channels
In-network aggregation(distributive, algebraic- aggregation happens at each hop)
1 2 3 k. . .
Schedule length
5 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Maximizing Convergecast Throughput
Motivation and Preliminaries
Convergecast and Benefits of Multiple Frequencies
Convergecast
A many-to-one communication pattern (i.e., flow of data): from all the nodesto a sink (opposite of broadcast).
a b c
d e f g
s
f1 f1f1
f1
1
12
23 4
5
Figure: Schedule length is 5.
6 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Maximizing Convergecast Throughput
Motivation and Preliminaries
Convergecast and Benefits of Multiple Frequencies
Convergecast
A many-to-one communication pattern (i.e., flow of data): from all the nodesto a sink (opposite of broadcast).
a b c
d e f g
s
f1 f1f1
f1
1
12
23 4
5
Figure: Schedule length is 5.
a b c
d e f g
s
f1 f1f2
f1
1
12
21 2
3
Figure: Schedule length is 3.
6 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Maximizing Convergecast Throughput
Motivation and Preliminaries
Convergecast and Benefits of Multiple Frequencies
Convergecast
A many-to-one communication pattern (i.e., flow of data): from all the nodesto a sink (opposite of broadcast).
Frame 1 Frame 2
Receiver Slot 1 Slot 2 Slot 3 Slot 1 Slot 2 Slot 3
sab
c
c
de
a d
fg
bef c g
de f
a d bef
g
Pipelining
a b c
d e f g
s
f1 f1f2
f1
1
12
21 2
3
Figure: Schedule length is 3.
7 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Maximizing Convergecast Throughput
Motivation and Preliminaries
Problem Formulation
Joint Frequency and Time Slot Assignment
Given a connected network G = (V, E) of arbitrarilydeployed nodes, K orthogonal frequencies, and a routingtree T ⊆ G rooted at the sink s ∈ V
Assign a frequency to each receiver in T
Assign a time slot to each edge in T
such that the aggregated throughput at the sink is max-imized, where
Aggregated Throughput = 1/Schedule Length(equivalent to minimizing the Schedule Length) Secondary conflict
8 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Maximizing Convergecast Throughput
Motivation and Preliminaries
Scheduling Complexity
Theorem
Joint Frequency and Time Slot Assignment is NP-hard on general graphs.Proof: Reduction from Vertex Color.
Questions
Does the problem still remain NP-hard with unlimited frequencies?
If not, how many frequencies (at a min, or at most) are needed?
Can we design algorithms that have provable, worst-case guarantee?
9 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Maximizing Convergecast Throughput
Motivation and Preliminaries
Scheduling Complexity
Theorem
Joint Frequency and Time Slot Assignment is NP-hard on general graphs.Proof: Reduction from Vertex Color.
Questions
Does the problem still remain NP-hard with unlimited frequencies?
If not, how many frequencies (at a min, or at most) are needed?
Can we design algorithms that have provable, worst-case guarantee?
Theorem
Given a spanning tree T on a graph G = (V, E), findingthe minimum number of frequencies required to removeall secondary conflicts is NP-hard.
f1
f2
1 1
Secondary conflict
9 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Maximizing Convergecast Throughput
Scheduling with Unlimited Frequencies
A Frequency Upper Bound
Constraint Graph
For each receiver in G create a node in GC .
Create an edge between two nodes in GC if the corresponding receivers inG form a secondary conflict.
1 2
3 4 5
6 7 8
9
10 11
12 13 15 1614
s
f11 2
3 4 5
6 7 8f4
f4
f1
f2f2
f3f3
f3sLargest DegreeFirst
Lemma
Kmax ≤ ∆(GC) + 1, where ∆(GC): max degree in GC . (Vertex Color)
10 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Maximizing Convergecast Throughput
Scheduling with Unlimited Frequencies
Time Slot Assignment
Algorithm: BFS-TimeSlotAssignment
1. while ET 6= φ do2. e← next edge from ET in BFS order3. Assign minimum time slot to e respecting
adjacency constraint4. ET ← ET \ {e}5. end while
1 2
3 4 5
6 7 8
9
10 11
12 13 15 1614
s
1
1
1
1
1
2 3
2
2
2 2
2
2
3 3
3
f1
f4
f4
f1
f2f2 f3
f3
f3
Theorem
BFS-TimeSlotAssignment gives a schedule of minimum length ∆(T ), where∆(T ): maximum node degree in T .
11 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Maximizing Convergecast Throughput
Scheduling with Unlimited Frequencies
Evaluation
Parameters:
Number of nodes: n : 200, square region of area A : 200× 200
Shortest Path Tree, Unit Disk Graph
0 0.1 0.2 0.3 0.4 0.50
5
10
15
20
25
30
Network Density
Num
ber
of F
requ
enci
es o
n S
PT
Largest Degree FirstUpper Bound ∆(G
C) + 1
Figure: Number of frequencies required to remove all secondary conflicts as a functionof network density n/A on SPT.
12 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Maximizing Convergecast Throughput
Scheduling with Unlimited Frequencies
So far...
Joint Frequency and Time Slot Assignment is
NP-hard with limited (constant number) frequencies on arbitrary graphs.
Polynomial with unlimited (at most ∆(GC) + 1) frequencies for arbitrarygraphs.
Next...
Approximation algorithms for Random Geometric Graphs with limited (constantnumber) frequencies.
Uniform transmission range (Unit Disk Graphs)
Non-uniform transmission range (General Disk Graphs)
13 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Maximizing Convergecast Throughput
Approximations on Unit Disk Graphs
Approximation on UDG
Algorithm AUDG Overview
Phase 1: Assign frequencies to thereceivers in each cell such that themaximum number of nodes transmittingon the same frequency (load) is minimized.(Load-Balanced Frequency Assignment)
Phase 2: Assign time slot greedily to eachedge.
e e�f f
g1 g1
g1 g1
g2 g2
g2 g2
g3 g3
g3 g3
g4 g4
g4 g4
a>2R
u u�v v�
e e�f f
u u�v v�
14 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Maximizing Convergecast Throughput
Approximations on Unit Disk Graphs
Approximation on UDG
Algorithm AUDG Overview
Phase 1: Assign frequencies to thereceivers in each cell such that themaximum number of nodes transmittingon the same frequency (load) is minimized.(Load-Balanced Frequency Assignment)
Phase 2: Assign time slot greedily to eachedge.
e e�f f
g1 g1
g1 g1
g2 g2
g2 g2
g3 g3
g3 g3
g4 g4
g4 g4
a>2R
u u�v v�
e e�f f
u u�v v�
Theorem
Algorithm AUDG achieves a constant factor 8µα ·(
43− 1
3K
)
approximation onthe optimal schedule length, where µα > 0 is a constant for any given cell sizeα ≥ 2R, and given node deployment, whereµα : maximum number of edges on the same frequency in any cell that can bescheduled simultaneously by any algorithm.
14 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Maximizing Convergecast Throughput
Approximations on Unit Disk Graphs
Evaluation
Parameters
UDG on a region of size 200× 200 with transmission radius R = 25
Shortest Path Tree
100 200 300 400 500 600 700 8005
10
15
20
25
30
35
40
45
50
Number of Nodes
Sch
edul
e Le
ngth
SPT, K=1SPT, K=3SPT, K=5
Observations
Schedule length decreases withincreasing number of frequencies
However, with diminishing returns
Structure of the routing tree (highnode degree) may lead tobottlenecks
15 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Maximizing Convergecast Throughput
Approximations on General Disk Graphs
Approximation on Disk Graphs
16 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Maximizing Convergecast Throughput
Approximations on General Disk Graphs
Approximation on Disk Graphs
Notations
r(u): Transmission range of node u
ℓ(e): Length of edge e
I(e): Set of edges that are either adjacent to e or form a secondaryconflict with e
I≥(e): Subset of I(e) that have larger disks than e
I≥(e) = {e′ = (u′, v′) : e′ ∈ I(e), ℓ(e′) ≥ ℓ(e)}
euv
16 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Maximizing Convergecast Throughput
Approximations on General Disk Graphs
Approximation on Disk Graphs
Algorithm ADG Overview
Phase 1: Assign frequencies such that the maximum number of edgesinterfering with any given edge is minimized.
0 − 1 Integer Linear Program: Define indicator variables Xvk for edgee = (u, v) as:
Xvk =
{
1, if receiver v is assigned frequency fk
0, otherwise
A frequency assignment is a 0 − 1 assignment to Xvk, ∀e, ∀fk.
17 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Maximizing Convergecast Throughput
Approximations on General Disk Graphs
Approximation on Disk Graphs
Algorithm ADG Overview
Phase 1: Assign frequencies such that the maximum number of edgesinterfering with any given edge is minimized.
0 − 1 Integer Linear Program: Define indicator variables Xvk for edgee = (u, v) as:
Xvk =
{
1, if receiver v is assigned frequency fk
0, otherwise
A frequency assignment is a 0 − 1 assignment to Xvk, ∀e, ∀fk.
Phase 2: Sort edges in non-increasing order of lengths, and assign smallestavailable slots starting from the largest.
17 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Maximizing Convergecast Throughput
Approximations on General Disk Graphs
Approximation on Disk Graphs
0-1 Integer Linear Program of Frequency Assignment
Minimize λ
subject to :
∀e = (u, v), ∀fk :∑
v′
n(e, v′)Xvk ≤ λ (1)
∀e = (u, v) :∑
fk
Xvk = 1, (2)
∀e = (u, v), ∀fk : Xvk ∈ {0, 1} (3)
Solve the Linear Relaxation (LP) by modifying constraint (3)
Randomized Rounding: Assign Yvk = 1 with probability X∗vk, where X∗
vk
are the optimal fractional LP solutions, and Yvk are the new integralrandom variables.
18 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Maximizing Convergecast Throughput
Approximations on General Disk Graphs
Approximation on Disk Graphs
Lemma
Let Yvk be the rounded solution, as described above. Then,
maxe,fk
∑
e′=(u′,v′)∈I≥(e)
Yv′k
= O (∆(T ) log n · λ∗) ,
with probability at least (1− 1/n).Proof: Apply Chernoff bound.
19 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Maximizing Convergecast Throughput
Approximations on General Disk Graphs
Approximation on Disk Graphs
Lemma
Let Yvk be the rounded solution, as described above. Then,
maxe,fk
∑
e′=(u′,v′)∈I≥(e)
Yv′k
= O (∆(T ) log n · λ∗) ,
with probability at least (1− 1/n).Proof: Apply Chernoff bound.
Theorem
The schedule constructed by time slot assignmentt strategy along with the fre-quency assignment using the above randomized rounding procedure results in aschedule of length O(∆(T ) · log n) times the optimum.
19 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Multi-Channel Scheduling in SINR
Outline
1 Maximizing Convergecast ThroughputMotivation and PreliminariesScheduling with Unlimited FrequenciesApproximations on Unit Disk GraphsApproximations on General Disk Graphs
2 Multi-Channel Scheduling in SINRMotivation and Interference ModelsProblem Formulation and ApproachApproximations in SINR
3 Throughput-Delay Trade-offMotivationBicriteria FormulationOptimal Spanning Trees
4 Conclusions
20 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Multi-Channel Scheduling in SINR
Motivation and Interference Models
Realistic Interference Model
Limitations of a Graph-Based Model (Protocol Model)
Interference is binary and stops abruptly beyond a distance - idealizesphysical laws
Fails to capture cumulative interference from (many) far-away transmitters
Sometimes pessimistic decisions and conflicting schedules
21 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Multi-Channel Scheduling in SINR
Motivation and Interference Models
Realistic Interference Model
SINR Model (Physical Model)
For a given frequency f , the received signal power from sender si to its intendedreceiver ri is:
Pri(si) =
P
d(si, ri)α,
where α ∈ [2, 6] is called the path-loss exponent; value depends on externalconditions.
22 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Multi-Channel Scheduling in SINR
Motivation and Interference Models
Realistic Interference Model
SINR Model (Physical Model)
For a given frequency f , the received signal power from sender si to its intendedreceiver ri is:
Pri(si) =
P
d(si, ri)α,
where α ∈ [2, 6] is called the path-loss exponent; value depends on externalconditions.
A transmission from sender si to its intended receiver ri is successful if the ratio,SINRri
, of the received signal power to the cumulative interference plus noiseN at ri is more than a hardware-dependent threshold β:
SINRri=
Pd(si,ri)α
Iri+N =
Pd(si,ri)α
∑
j 6=iP
d(sj ,ri)α +N≥ β
22 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Multi-Channel Scheduling in SINR
Motivation and Interference Models
Realistic Interference Model
Parameters
CC2420 radio parameters with receiver sensitivity −95 dB
Path-loss exponent α = 3.5, transmit power −6 dB
Percentage of Conflicting Schedules:SINR vs. Graph-based Model
100 200 300 400 500 600 700 8000
5
10
15
20
25
30
35
40
Number of Nodes
Con
flict
Lin
ks (
%)
SPT, K=1SPT, K=3SPT, K=5
Observations
For a given frequency, the amountof conflict increases as the networkgets denser, reaching almost 40%for densest deployment with asingle frequency
With multiple frequencies, theamount of conflict is much less
23 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Multi-Channel Scheduling in SINR
Problem Formulation and Approach
Problem Formulation in SINR
Joint Frequency and Time Slot Assignment
Given a connected network of arbitrarily deployed nodes,K orthogonal frequencies, and a routing tree T =(V, ET ) rooted at the sink s ∈ V
Assign a frequency to each receiver in T
Assign a time slot to each edge in T
such that the schedule length is minimized, subject to
SINRri≥ β at every receiver ri Cumulative interference
24 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Multi-Channel Scheduling in SINR
Problem Formulation and Approach
Approach
Link Diversity
Classify the edges in ET based on their lengths ℓi = d(si, ri). Link diversity isthe number of magnitude of different lengths
g(ET ) = |{m | ∃ ei, ej ∈ ET : ⌊log (ℓi/ℓj)⌋ = m}|
Usually a small constant in practice, but theoretically can be equal to the numberof nodes
25 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Multi-Channel Scheduling in SINR
Problem Formulation and Approach
Approach
Link Diversity
Classify the edges in ET based on their lengths ℓi = d(si, ri). Link diversity isthe number of magnitude of different lengths
g(ET ) = |{m | ∃ ei, ej ∈ ET : ⌊log (ℓi/ℓj)⌋ = m}|
Usually a small constant in practice, but theoretically can be equal to the numberof nodes
Overall Idea
Normalize (w.l.o.g) the minimum edge length to one
LetE =
{
E0, . . . , Elog(emax)
}
denote the set of non-empty length classes, where Ek is the set of edgesof lengths in
[
2k, 2k+1)
, and emax is the longest edge
Partition the problem into disjoint length classes, and process edges ineach class separately
25 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Multi-Channel Scheduling in SINR
Problem Formulation and Approach
Approach
For Length Class Ek
Divide the 2-D region into a setCk of square grids of lengthηk = δ · 2k, for some constant δ(will choose δ later)
Frequency Assignment:For each cell in Ck, runFrequencyGreedy on thereceivers
Time Slot Assignment:Four color the cells in Ck
Can we still reuse time slotsacross non-adjacent cells of thesame color? If so, how many?
g1 g1
g1 g1
g2 g2
g2 g2
g3 g3
g3 g3
g4 g4
g4 g4
hk=d.2k
g1 g1g2 g2
g1
g1
g3
g3
g1
e e�f f
u u�v v�
e e�f f
u u�v v�
26 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Multi-Channel Scheduling in SINR
Approximations in SINR
Approximation Algorithm in SINR
Theorem
For cells of the same color in Ck, we can simultaneously schedule at most oneedge in Ek from every non-adjacent cell that lies on the same frequency, so longas the cell size satisfies:
ηk = δ · 2k, for δ = 4
[
8β · α− 1
α− 2
] 1
α
,
where β ≥ 1 is the SINR threshold, and α > 2 is the path-loss exponent.
27 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Multi-Channel Scheduling in SINR
Approximations in SINR
Approximation Algorithm in SINR
Theorem
For cells of the same color in Ck, we can simultaneously schedule at most oneedge in Ek from every non-adjacent cell that lies on the same frequency, so longas the cell size satisfies:
ηk = δ · 2k, for δ = 4
[
8β · α− 1
α− 2
] 1
α
,
where β ≥ 1 is the SINR threshold, and α > 2 is the path-loss exponent.
Proof Sketch
Consider one specific edge ei = (si, ri) ∈ Ek in any cell c ∈ Ck
Received power at ri
Pri(si) =
P
ℓαi
≥ P
2α(k+1), ∵ 2k ≤ ℓi < 2k+1
Show that the cumulative interference Iricaused by concurrent
transmissions from all non-adjacent cells is still within β27 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Multi-Channel Scheduling in SINR
Approximations in SINR
Approximation Algorithm in SINR
g1 g1
g1 g1
g2 g2
g2 g2
g3 g3
g3 g3
g4 g4
g4 g4
g1 g1g2 g2
g1
g1
g3
g3
g1
g2
g2
g4
g4
g2
g2
g2
g4
g4
g2
g3 g3g4 g4 g3 g4g4
g3 g3g4 g4 g3 g4g4
g1 g1g2 g2 g1 g2g2
g1 g1g2 g2 g1 g2g2
g1
g1
g3
g3
g1
g3
g3
g1
g1
g1
g1
g3
g3
g1
g3
g3
g1
g1
ei
hk=d.2
k
Consider edge ei = (si, ri) ∈ Ek
28 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Multi-Channel Scheduling in SINR
Approximations in SINR
Approximation Algorithm in SINR
g1 g1
g1 g1
g2 g2
g2 g2
g3 g3
g3 g3
g4 g4
g4 g4
g1 g1g2 g2
g1
g1
g3
g3
g1
g2
g2
g4
g4
g2
g2
g2
g4
g4
g2
g3 g3g4 g4 g3 g4g4
g3 g3g4 g4 g3 g4g4
g1 g1g2 g2 g1 g2g2
g1 g1g2 g2 g1 g2g2
g1
g1
g3
g3
g1
g3
g3
g1
g1
g1
g1
g3
g3
g1
g3
g3
g1
g1
ei
ej ej ej
ej ej ej
ejej
hk=d.2
k
Consider edge ei = (si, ri) ∈ Ek
At most 8 from layer 1
29 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Multi-Channel Scheduling in SINR
Approximations in SINR
Approximation Algorithm in SINR
g1 g1
g1 g1
g2 g2
g2 g2
g3 g3
g3 g3
g4 g4
g4 g4
g1 g1g2 g2
g1
g1
g3
g3
g1
g2
g2
g4
g4
g2
g2
g2
g4
g4
g2
g3 g3g4 g4 g3 g4g4
g3 g3g4 g4 g3 g4g4
g1 g1g2 g2 g1 g2g2
g1 g1g2 g2 g1 g2g2
g1
g1
g3
g3
g1
g3
g3
g1
g1
g1
g1
g3
g3
g1
g3
g3
g1
g1
ei
ej ej ej ej ej
ej ej ej ej ej
ej
ej
ej
ej
ej
ej
ej ej ej
ej ej ej
ejej
hk=d.2
k
Consider edge ei = (si, ri) ∈ Ek
At most 8 from layer 1
At most 16 from layer 2, ..., atmost 8q from layer q
Add up all the interferences up tolayer ∞
30 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Multi-Channel Scheduling in SINR
Approximations in SINR
Approximation Algorithm in SINR
g1
g1
g2
d.2k-2
k+1ri
rj
si
sj
d.2k
Computing Cumulative Interference Iri
From transmitter sj in a layer 1 non-adjacent cell of the same color
Iri(sj) =
P
ℓαij
≤ P
[2k(δ − 2)]α, ∵ ℓij ≥ δ · 2k − 2k+1
31 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Multi-Channel Scheduling in SINR
Approximations in SINR
Approximation Algorithm in SINR
g1
g1
g2
d.2k-2
k+1ri
rj
si
sj
d.2k
Computing Cumulative Interference Iri
From transmitter sj in a layer 1 non-adjacent cell of the same color
Iri(sj) =
P
ℓαij
≤ P
[2k(δ − 2)]α, ∵ ℓij ≥ δ · 2k − 2k+1
In general, ℓij ≥ 2k {(2q − 1)δ − 2} in layer q
31 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Multi-Channel Scheduling in SINR
Approximations in SINR
Approximation Algorithm in SINR
g1
g1
g2
d.2k-2
k+1ri
rj
si
sj
d.2k
Computing Cumulative Interference Iri
From transmitter sj in a layer 1 non-adjacent cell of the same color
Iri(sj) =
P
ℓαij
≤ P
[2k(δ − 2)]α, ∵ ℓij ≥ δ · 2k − 2k+1
In general, ℓij ≥ 2k {(2q − 1)δ − 2} in layer q
Adding up
Iri≤
∞∑
q=1
8qP
[2k {(2q − 1)δ − 2}]α =8P
2(k−1)αδα· α− 1
α− 2≤ Pri
(si)
β
31 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Multi-Channel Scheduling in SINR
Approximations in SINR
Approximation Algorithm in SINR
Theorem :O(g(ET )) Approximation
Given a routing tree T on an arbitrarily deployed connected network in 2-D, andK orthogonal frequencies, the algorithm gives a O (g(ET )) approximation onthe optimal schedule length.
32 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Multi-Channel Scheduling in SINR
Approximations in SINR
Approximation Algorithm in SINR
Theorem :O(g(ET )) Approximation
Given a routing tree T on an arbitrarily deployed connected network in 2-D, andK orthogonal frequencies, the algorithm gives a O (g(ET )) approximation onthe optimal schedule length.
Proof Sketch
Choice of the critical cell c in which the number of edges on the samefrequency Lm∗
c is maximum across all length classes
Show that OPT will take at least Lm∗
c /µ slots, for
µ =
[
2(√
2 · δ + 1)]α
β
Show by contradiction that OPT cannot schedule µ + 1 edges on the samefrequency in any cell
32 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Multi-Channel Scheduling in SINR
Approximations in SINR
Approximation Algorithm in SINR
Proof Sketch (cobmining together)
Γ ≤ 4 ·maxk,c{|γc∈Ck
|} · g(ET )
≤ 8 ·maxk,c
{
Lφc∈Ck
}
· g(ET )
≤ 8 ·maxk,c
{(
4
3− 1
3K
)
· Lm∗
c∈Ck
}
· g(ET )
≤ 8µ ·(
4
3− 1
3K
)
· g(ET ) ·OPT
= O (g(ET ))
33 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Throughput-Delay Trade-off
Outline
1 Maximizing Convergecast ThroughputMotivation and PreliminariesScheduling with Unlimited FrequenciesApproximations on Unit Disk GraphsApproximations on General Disk Graphs
2 Multi-Channel Scheduling in SINRMotivation and Interference ModelsProblem Formulation and ApproachApproximations in SINR
3 Throughput-Delay Trade-offMotivationBicriteria FormulationOptimal Spanning Trees
4 Conclusions
34 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Throughput-Delay Trade-off
Motivation
Properties of Spanning Trees
0 50 100 150 2000
20
40
60
80
100
120
140
160
180
200
x−axis
y−ax
is
Shortest Path Tree (shallow and fat)
High node degree ⇒ low throughput
Few hops ⇒ low delay
35 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Throughput-Delay Trade-off
Motivation
Properties of Spanning Trees
0 50 100 150 2000
20
40
60
80
100
120
140
160
180
200
x−axis
y−ax
is
Shortest Path Tree (shallow and fat)
High node degree ⇒ low throughput
Few hops ⇒ low delay
0 50 100 150 2000
20
40
60
80
100
120
140
160
180
200
x−axis
y−ax
is
Minimum Interference Tree (weighted MST)[Mobihoc ’04] (deep and skinny)w(u, v): no. of nodes covered by the union of twodisks centered at u and v, each of radius |uv|
Low node degree ⇒ high throughput
More hops ⇒ high delay
35 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Throughput-Delay Trade-off
Motivation
Evaluation of AUDG on SPT & MIT
Parameters
Random geometric graph on a region of size 200× 200 with R = 25
Shortest Path Tree
100 200 300 400 500 600 700 8005
10
15
20
25
30
35
40
45
50
Number of Nodes
Sch
edul
e Le
ngth
SPT, K=1SPT, K=3SPT, K=5
Minimum Interference Tree
100 200 300 400 500 600 700 8005
10
15
20
25
30
35
40
45
50
Number of Nodes
Sch
edul
e Le
ngth
MIT, K=1MIT, K=3MIT, K=5
36 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Throughput-Delay Trade-off
Motivation
Tree Property for Bounded Throughput
Theorem
If the maximum node degree of the routing tree is bounded by a constant∆C > 0, then there exists an algorithm that gives a constant factor 8µα · ∆C
approximation on the optimal schedule length, where µα > 0 is a constant for agiven cell size α ≥ 2R and given node deployment.
37 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Throughput-Delay Trade-off
Motivation
Tree Property for Bounded Throughput
Theorem
If the maximum node degree of the routing tree is bounded by a constant∆C > 0, then there exists an algorithm that gives a constant factor 8µα · ∆C
approximation on the optimal schedule length, where µα > 0 is a constant for agiven cell size α ≥ 2R and given node deployment.
Properties
Degree-bounded spanning trees? Minimum Degree Spanning Tree isNP-hard [SODA’92]
Is constant factor approximation on schedule length and delay achievable?
37 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Throughput-Delay Trade-off
Bicriteria Formulation
Bicriteria Formulation
Bounded-Degree Minimum-Radius Spanning Tree
P = (Degree, Radius, Spanning Tree)Radius: R(T ): Maximum hop distance from any node to sink in T∆∗: Budget on max degreeGoal: Find a spanning tree of min radius subject to max degree ∆∗
38 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Throughput-Delay Trade-off
Bicriteria Formulation
Bicriteria Formulation
Bounded-Degree Minimum-Radius Spanning Tree
P = (Degree, Radius, Spanning Tree)Radius: R(T ): Maximum hop distance from any node to sink in T∆∗: Budget on max degreeGoal: Find a spanning tree of min radius subject to max degree ∆∗
Approximation Definition
Algorithm A is an (α, β)-approximation to a bicriteria optimization problem P =(Degree, Radius, Spanning Tree) if:
∆(T ) ≤ α + ∆∗
R(T ) ≤ β ·R∗(T )
where R∗(T ) is the min radius of any spanning tree T whose max degree is ∆∗.
38 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Throughput-Delay Trade-off
Optimal Spanning Trees
Algorithm Overview
Algorithm ABDMRST
Three phases:
1 Construct a GlobalBackbone Tree of lowradius.
2 Constructdegree-bounded LocalSpanning Trees.
3 Merge the localspanning trees and theglobal backbone tree.
R/2
Completegraph
39 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Throughput-Delay Trade-off
Optimal Spanning Trees
Algorithm Overview
Global Backbone Tree
r1
r3
w1
r4
r2
s
w2
r5
r6
R
R/2
r7
R
R
r8
r9
r10
40 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Throughput-Delay Trade-off
Optimal Spanning Trees
Algorithm Overview
Global Backbone Tree
r1
r3
w1
r4
r2
s
w2
r5
r6
R
R/2
r7
R
R
r8
r9
r10
Local Spanning Tree
wk
rj
v1
R/2
40 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Throughput-Delay Trade-off
Optimal Spanning Trees
Properties of Spanning Trees
0 50 100 150 2000
20
40
60
80
100
120
140
160
180
200
0 50 100 150 2000
20
40
60
80
100
120
140
160
180
200
0 50 100 150 2000
20
40
60
80
100
120
140
160
180
200
Bounded-Degree Minimum-RadiusSpanning Tree, ∆∗ = 4
41 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Throughput-Delay Trade-off
Optimal Spanning Trees
Approximation Result & Evaluation
Theorem
Algorithm ABDMRST gives a constant factor (α, β) bicriteria approximation forthe Bounded-Degree Minimum-Radius Spanning Tree, where α = 10 and β = 5.
42 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Throughput-Delay Trade-off
Optimal Spanning Trees
Approximation Result & Evaluation
Theorem
Algorithm ABDMRST gives a constant factor (α, β) bicriteria approximation forthe Bounded-Degree Minimum-Radius Spanning Tree, where α = 10 and β = 5.
Schedule Length
100 200 300 400 500 600 700 8005
10
15
20
25
30
35
40
45
50
Number of Nodes
Sch
edul
e Le
ngth
BDMRSTSPTMIT
Max Delay
100 200 300 400 500 600 700 8000
20
40
60
80
100
120
Number of Nodes
Max
imum
Del
ay (
Tre
e R
adiu
s)
BDMRSTSPTMIT
42 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Conclusions
Outline
1 Maximizing Convergecast ThroughputMotivation and PreliminariesScheduling with Unlimited FrequenciesApproximations on Unit Disk GraphsApproximations on General Disk Graphs
2 Multi-Channel Scheduling in SINRMotivation and Interference ModelsProblem Formulation and ApproachApproximations in SINR
3 Throughput-Delay Trade-offMotivationBicriteria FormulationOptimal Spanning Trees
4 Conclusions
43 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Conclusions
Contributions
Maximizing aggregated convergecast throughput using multi-channelscheduling
Constant factor approximation on UDG∆(T ) · log(n) approximation on DGO(g(ET )) approximation in SINR
Throughput-delay trade-off in the same optimization frameworkBicriteria formulationConstant factor approximation on constructing bounded-degreeminimum-radius spanning trees
Efficient, distributed topology control in 3-DSimple orthographic projection based approach extending 2-D resultsRobust spherical Delaunay triangulation based algorithm
44 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Conclusions
Publications
Book Chapters
1 O.D. Incel, A. Ghosh, and B. Krishnamachari, Scheduling Algorithms for Tree-Based Data Collection in Wireless Sensor Networks,Theoretical Aspects of Distributed Computing in Sensor Networks, Springer, 2010.
2 A. Ghosh and S.K. Das, Coverage and Connectivity Issues in Wireless Sensor Networks, Mobile, Wireless and Sensor Networks:Technology, Applications and Future Directions, John Wiley & Sons, 2006.
Journal Papers
1 A. Ghosh, O.D. Incel, V.S. Anil Kumar, and B. Krishnamachari, Multi-Channel Scheduling and Spanning Tree Algorithms:Throughput-Delay Trade-off for Fast Convergecast in Sensor Networks, ACM/IEEE Transactions on Networking, Feb 2010 (sub).
2 O.D. Incel, A. Ghosh, B. Krishnamachari, and K. Chintalapudi, Fast Data Collection in Tree-Based Wireless Sensor Networks,IEEE Transactions on Mobile Computing, Dec 2009 (rev).
3 A. Viswanath, P. Dutta, M. Chetlur, P. Gupta, S. Kalyanaraman, and A. Ghosh, Perspectives on Quality of Experience for VideoStreaming over WiMAX, ACM SIGMOBILE Mobile Computing and Communications Review, Oct, 2009.
4 A. Ghosh and S.K. Das, Coverage Connectivity Issues in Wireless Sensor Networks: A Survey, Journal of Pervasive and MobileComputing, 4(3):303-334, June 2008.
45 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Conclusions
Publications
Conference Papers (reverse chronologically)
1 A. Ghosh, R. Jana, V. Ramaswami, J. Rowland, and N.K. Shankar, Modeling and Characterization of Large-Scale Wi-Fi Traffic inPublic Hot-Spots, IEEE INFOCOM, 2011 (sub).
2 A. Ghosh, O.D. Incel, V.S. Anil Kumar, and B. Krishnamachari, Optimal Spanning Trees for Fast Data Collection in SensorNetworks, IEEE INFOCOM Student Workshop, Mar 2010.
3 A. Ghosh, O.D. Incel, V.S. Anil Kumar, and B. Krishnamachari, Multi-Channel Scheduling Algorithms for Fast AggregatedConvergecast in Sensor Networks, IEEE MASS, October 2009.
4 S.W. Lee, D. Yoon, and A. Ghosh, Intelligent Parking Lot Application using Wireless Sensor Networks, International Symposiumon Collaborative Technologies and Systems (CTS), May 2008.
5 A. Ghosh, L. Pereira, T. Yan, and H. Cao, Modeling Wireless Sensor Network Architectures using AADL, European Congress onEmbedded Real Time Software (ERTS), Jan 2008.
6 A. Ghosh, Y. Wang, and B. Krishnamachari, Efficient Distributed Topology Control in 3 Dimensional Wireless Networks, IEEESECON, Jun 2007.
7 A. Ghosh and S.K. Das, A Distributed Greedy Algorithm for Connected Sensor Cover in Dense Sensor Networks, IEEE/ACMDCOSS, Jul 2005.
8 A. Ghosh, Estimating Coverage Holes and Enhancing Coverage in Mixed Sensor Networks, IEEE LCN, Nov 2004.
9 S.K. Ghosh and A. Ghosh, A Plan-Commit-Prove Protocol for Secure Verification of Traversal Path, IEEE InternationalConference on Networks (ICON), Nov 2004.
Manuscripts
1 T. Fu, A. Ghosh, E.A. Johnson, and B. Krishnamachari, Energy-Efficient Deployment Strategies in Structural Health Monitoringusing Wireless Sensor Networks, Structural Control and Health Monitoring, 2009 (manuscript).
2 A. Ghosh, A.H. Patel, C.R. Sastry, Radio Frequency-Based Localization in Wireless Sensor Networks, Siemens Corporate ResearchTechnical Report, Aug 2009.
46 / 47
Algorithmic Aspects of Throughput-Delay Performance for Fast Data Collection in Sensor Networks
Conclusions
Acknowledgments
Collaborators
USC (EE, Civil): Bhaskar Krishnamachari, Erik Johnson, Yi Wang, Tat Fu
Bogazici University: Ozlem Durmaz Incel
Virginia Tech: Anil Vullikanti
University of Texas: Sajal K. Das
IBM Research: Shivkumar Kalyanaraman, Partha Dutta, Malolan Chetlur
Siemens Corporate Research: Chellury Ram Sastry
Microsoft Research: Krishnakant Chintalapudi
Eaton Corporation: Ting Yan, Luis Pereira, Hui Cao
AT&T Labs Research: Vaidyanathan Ramaswami, Rittwik Jana, N.K.Shankar
Thank You!
47 / 47