Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                

p132 Li PDF

Download as pdf or txt
Download as pdf or txt
You are on page 1of 9

Throughput and Energy Efficiency in Topology-Controlled

Multi-hop Wireless Sensor Networks


Li (Erran) Li

Prasun Sinha

Bell Labs, Lucent Technologies


Holmdel, NJ 07733

Dept of CIS, Ohio State University


Columbus, OH 43210

erranlli@dnrc.bell-labs.com

prasun@cis.ohio-state.edu

ABSTRACT

Keywords

In the context of multi-hop wireless networks, various topology control algorithms have been proposed to adapt the
transmission range of nodes based on local information while
maintaining a connected topology. These algorithms are
particularly suited for deployment in sensor networks which
typically consist of energy constrained sensors. Sensor nodes
should support power adaptation in order to use the benets
of topology control for energy conservation. In this paper,
we design a framework for evaluating the performance of
topology control algorithms using overall network throughput, and total energy consumption per packet delivered, as
the metrics. Our goal is to identify the scenarios in which
topology control improves the network performance. We
supplement our analysis with ns2 simulations using the conebased topology control algorithm [10, 19].
Based on our analysis and simulations, we nd that link
layer retransmissions are essential with topology control to
avoid throughput degradation due to increase in number of
hops in lightly loaded networks. In heavily loaded networks,
the throughput can be improved by a factor up to k2 , where
k is the average factor of reduction in transmission range
using topology control. Studies of energy consumption reveal that improvements of up to k4 can be obtained using
topology control. However, these improvements decrease as
the trac pattern shifts from local (few hop connections) to
non-local (hop lengths of the order of the diameter of the
network). These results can be used to guide the deployment
of topology control algorithms in sensor networks.

Sensor Networks, Ad-hoc Networks, Wireless Networks, Topology Control

1. INTRODUCTION
Multi-hop wireless networks, such as radio networks [7],
ad-hoc networks [14], and sensor networks [4, 15], are networks where communication between two devices may go
through multiple consecutive wireless links. Unlike wired
networks for which the topology is xed (except in cases
of link failures), these networks have dynamic topologies
based on the location of the devices and their range of transmission. For suciently dense networks, nodes can restrict
communication only to nearby nodes with reduced range
of transmission and yet maintain global network connectivity. The mechanism for computing the sucient transmission ranges for each node is called topology control, and its
main goal is to optimize performance metrics such as network lifetime and throughput while maintaining a connected
network.
Reduction in transmission range reduces the per packet
energy consumption and also promotes spatial reuse. However lengthening of routes eects the end-to-end reliability of
ows. Taking these factors into account, it becomes critical
to understand the applicability and usefulness of topology
control in various networking scenarios. To this end, the
goal in this paper is to design a framework for analyzing
the performance of networks with topology control and to
identify scenarios where topology control is benecial to use.
Our study is focused on UDP ows and the metrics used are
throughput improvement, and energy eciency. We dene
throughput to be the total number of bits received per second by the destinations of all the multi-hop ows in the
network. For purposes of energy consumption comparisons,
the metric we use is the total transmission and reception
energy consumed per successfully received bit.
This paper makes the following two contributions to the
understanding of the performance of topology control. First,
we present an analytical framework for characterizing the
performance of topology control under dierent trac patterns, packet error rates, and maximum number of link layer
retransmissions. Second, using the cone-based topology control algorithm [10] as an example, we show that the results
from realistic simulation setting closely follow our analysis.
Based on our studies, the performance of topology control
for large networks can be characterized as follows:

Categories and Subject Descriptors


C.2.1 [Network Architecture and Design]: Network
Communications, Network Topology, Wireless Communication

General Terms
Performance, Experimentation

Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
WSNA03, September 19, 2003, San Diego, California, USA.
Copyright 2003 ACM 1-58113-764-8/03/0009 ...$5.00.

In heavily loaded scenarios, topology control can im-

132

prove the throughput by a factor up to k2 , where k is


the average factor of reduction in transmission range
obtained by using topology control.

to the scheme where no topology control is used as Plain.


Let Rtc and rtc be the average eective interference range
and transmission range respectively, when topology control
is used and let Rpl and rpl be the corresponding notations
for Plain. We assume the network is dense enough such that
the network is connected if every node uses rpl as its transR
r
and k = rpl
. If the network is
mission range. Let ke = Rpl
tc
tc
suciently large, ke = k as typically the interference range
scales with the transmission range. Throughout this paper,
the subscript tc is used to refer to a parameter associated
with the network using Topology Control, and the subscript
pl to refer to the parameter associated with the network
without using Topology Control. We drop the subscript tc
and pl when our discussion applies to both. If Rm is the
maximum transmission range, then Rpl = Rm . We assume
that the network nodes are uniformly distributed in a square
area of size A.
The bit error rate is a function of various parameters including the modulation scheme, the received power, the ambient noise and the raw channel bit-rate. For our analysis
we make the simplifying assumption that the channel error
rate is independent of the received power. However, in our
simulations we consider both, a realistic channel error model
based on received power and ambient noise (see Section 3.3),
and the error model independent of received power. The former model considers our results in a more realistic setting
while the latter provides insights into our study as it corresponds to the model used in our analysis.
We assume that all packets are of same length. Our analysis assumes that the transmission range is uniform across all
the nodes in the network2 . This average transmission range
will be smaller for Topology Control than for Plain. We
do not model the possible eects of routing protocols and
MAC layer collision avoidance mechanisms. We leave it to
the simulations to explore these dimensions of the problem.
We use the power law trac pattern as our trac model [8].
It allows modeling of a variety of trac patterns ranging
from local trac to trac spanning the diameter of the
network. Under the power law distribution of network trafc, the probability that a node communicates with a destination at a distance of x, is proportional to x , for a constant
. So, the expression for the probability is given by

In lightly loaded scenarios, channel error causes reduction in throughput. Increase in link layer persistence
lessens the severity of throughput reduction.
Studies of energy consumption reveal that improvements of up to k4 can be obtained using topology control for local unicast trac. For an unreliable link
layer, Topology Control increases the energy consumption, if the trac pattern is non-local. However, this
can be eectively prevented by the use of link layer
retransmissions.
While the results in this paper apply to wireless ad hoc
networks in general, the following two main aspects of the
framework are specically targeted for sensor networks. First,
we assume a network of relatively large number of nodes that
are randomly placed in a region. The random distribution
can be a result of sensor deployment from moving aircrafts
or vehicles. Second, we consider energy eciency as one of
the study parameters, which is vital for sensor networks, as
replacing or recharging batteries is dicult, if not impossible.
The rest of the paper is organized as follows. Section 2
characterizes the benets of using topology control in terms
of end-to-end throughput improvements and energy eciency.
Section 3 presents results from simulation on the ns2 simulator using an idealized MAC layer and a simple xed error
rate channel model as well as a more realistic ambient noise
dependent channel model. In Section 4, we present related
work in the eld of topology control for multi-hop wireless
networks. Section 5 points to areas which are part of our
future work. Section 6 concludes the paper.

2.

ANALYSIS OF NETWORK THROUGHPUT AND ENERGY EFFICIENCY

In this section we analyze the performance of networks


with topology control in terms of network throughput and
energy eciency. We rst state our assumptions and describe our trac model. We then analyze the performance of
networks using topology control under dierent trac patterns, network loads, link layer reliability and channel error
rates.

q(x) =

(1)

A
t dt

is introduced to make the probability density function


well-dened for 1. It is a non-zero minimum distance between any source-destination pair. The power law
distribution captures the overall locality of network trac.
For large negative , destinations are clustered close to the
sender. For large positive , destinations are dispersed to
the periphery of the network. = 0 corresponds to a trafc pattern where the distance between the source and the
destination is picked from a uniform distribution.
For this trac pattern, the average path length L is given
by

2.1 Assumptions and Traffic Model


The range of interference for packet transmissions is larger
than the transmission range. For our analysis, we assume
that within the range of interference, no other transmissions
or receptions may occur simultaneously1 . Nodes closer to
the boundary of the network may have part of their interference region outside the network boundaries. In order to
properly apply our analytical results to the simulation scenarios, we dene the notion of eective interference range.
For a node with its interference region completely inside
the network boundary, the eective interference range is the
same as the interference range. Otherwise, it is dened as
the radius of a disk whose area is equal to the area of the
interference region inside the network boundaries. We refer

1
A more realistic interference model is used in our simulations.

L=

x
A
t dt

dx

(2)

2
This assumption holds well when the node density is uniform.

133

throughput of Topology Control will be less than Plain. We


plot the throughput ratio in Figures 1(a) and 1(b). The parameters used are as follows: rpl = 250m, rtc = 92m and
= 1.56 104 nodes per square meter. The transmission
range rpl is based on Lucent WaveLANs transmission range
that is also used in our simulations. The transmission range
for Topology Control rtc , is based on the average transmission range observed in our simulations with a node density
of . Figure 1(a) shows that the throughput of Topology
Control will decrease drastically as and p increases. Since
link layer retransmissions increase S(x), the throughput ratio will decrease gradually with increasing if link layer
retransmits as shown in Figure 1(b).
Intuitively, in an unsaturated network, the reduction in
spectrum utilization does not improve the throughput as the
spectrum is not the bottleneck. The end-to-end reliability
of the packets impact the total throughput.

Closed form expression of L with respect to dierent


can be found in [8]. Given this trac pattern with various
values of , we derive expressions for the throughput ratio
and energy eciency ratio of Topology Control over Plain.
We present the expressions graphically based on numerical
solutions obtained using the Matlab toolkit.

2.2 Throughput Analysis


We study two dierent types of network scenarios, namely
saturated and unsaturated. The distinction is based on the
total MAC layer trac load in the network. In a saturated
network, the MAC layer can not handle any extra transmissions beyond the current load. For such scenarios, besides
other factors, the total end-to-end throughput is limited by
the hop-by-hop throughput, commonly referred to as the
capacity of the network. In an unsaturated network, the
raw channel capacity is not fully utilized. For such scenarios, the limiting network capacity will not play any role.
However, other factors such as channel error rate and maximum number of allowed link layer retransmissions impact
the throughput ratio. We therefore study the two cases separately.

2.2.2 Saturated Networks


The throughput in this case will be the same as in Equation 3 except that will no longer be same for Topology
Control and Plain. is a function of p, number of link
layer transmission attempts mtx, and transmission range
r. The proportion of trac that can go into the air for
a destination distance x away is q(x). Let G(x) denote
the expected number of hop-by-hop transmissions that occur when a packet begins its journey to a destination at a
distance x away (See appendix for the expression of G(x)).
The total
one hop transmissions per unit time will then be

A
= q(x)G(x)dx. If the network is saturated, then
the maximum number of possible one hop transmissions per
WA
unit time is R
2 where W is the raw bandwidth of the channel, R is the eective sensing range. Let be the eciency
WA
of the MAC layer. Then = R
2 . Therefore,

2.2.1 Unsaturated Networks


In this scenario, the shared media has very few packets to
send. Let be the total end-to-end trac that gets injected
into the network per unit time. This does not include packets that have not left the senders buer. Let S(x) be the
probability that a packet will be delivered to a destination at
a distance of x units. Clearly S(x) = (1 pmtx )h(x) where
p is the probability of transmission failure on a link, mtx
is the maximum number of transmission attempts made by
the link layer, and h(x) is the number of hops needed for
the packet.
In a highly dense network, h(x) can be approximated by
 xr  (r is the average transmission range). If the network
density is not very high, the expression for h(x) is a lower
bound on the hop length. Since the total trac for distance
x is q(x), the total trac successfully delivered to distance
x is q(x)S(x). The overall throughput integrated over all
distances is thus,

q(x)S(x)dx

Tratio

(3)

In unsaturated scenarios, the total trac that the senders


are generating is exactly the trac that gets injected into
the network (). Therefore will drop out of the equation
when we consider the throughput ratio, which is given by
=

Tratio

A
1(1p
) rtc

A
r
1(1pmtx ) pl
mtx

if p=0
, if =0

W A
A

(5)

q(x)G(x)dx

Ttc
Tpl

8
<k
:

max(1, rL )
2
pl
e max(1, L ) ,
rtc

if p=0

(6)

From Equation 6 and Figure 2, we see that Topology Control can achieve a throughput ke2 times more than Plain if
the trac pattern is local. As the trac pattern becomes
more and more non-local, the throughput ratio drops.
Note that, in the above analysis we assume that for Topology Control, the variance of the transmission and interference ranges across nodes in the network is small. Otherwise,
Tratio can be much smaller.
Intuitively, in a saturated network, the extra available
spectrum obtained due to topology control, is used up by
other ows as the spectrum is the bottleneck. As a result, for saturated networks, topology control improves the
throughput, specially for local trac. For non-local trac,
the end-to-end reliability decreases sharply for the case of
topology control, and may reduce the total throughput.

Ttc
Tpl

8> 1,
><
>>:

Thus, the throughput ratio is,

T =

R2

(4)

Since the closed form solution for general p and is cumbersome and unwieldy, we only give the closed form solution
for specic cases. We resort to plots from numerical solutions to illustrate the value of Tratio for dierent p and .
Clearly, the overall throughput will be the same for both
Plain and Topology Control provided that the link is reliable. Since normally rtc < rpl , Equation 4 shows that the

134

Unsaturated Network: mtx=1, r =92 and r =250

Unsaturated Network: mtx=3, r =92 and r =250

pl

tc

0.8

0.8

0.6

0.6

pl

ratio

Throughput Ratio

tc

0.4

0.4

0.2

0.2

0
3

p=0
p=0.1
p=0.2
p=0.5

p=0
p=0.1
p=0.2
p=0.5

0
3

(a) No link layer retransmission

(b) Up to two retransmissions

Figure 1: Unsaturated networks: throughput ratio Tratio with respect to and packet dropping probability
p.
Saturated Network: mtx=1, r =92 and r =250
tc

Saturated Network: mtx=3, r =92 and r =250

pl

tc

p=0
p=0.1
p=0.2
p=0.5

3.5
3

ratio

2.5

ratio

p=0
p=0.1
p=0.2
p=0.5

3.5

2.5

1.5

1.5

0.5

0.5

0
3

pl

0
3

(a) No link layer retransmission

(b) Up to two retransmissions

Figure 2: Saturated networks: throughput ratio Tratio with respect to and packet dropping probability p.

2.3 Energy Efficiency

to

Recall that energy eciency is dened to be the ratio of


the amount of energy consumed per successfully delivered
packet for Plain over that of Topology Control. Therefore,
by denition, energy eciency is independent of network
load.
tx
is proportional to
For Plain, the transmission energy Epl

R , where both R, the maximum transmission range, and


tx
, the power loss factor are constants. So, Epl
is a contx
stant. However, for Topology Control, Etc
is proportional
to r , where the range r is discovered as a result of the
topology control algorithm. We use = 4 for our studies. Now we compute the energy consumed for receiving
packets. For Topology Control, every packet transmission
is received by R2 1 nodes, where is the density of
nodes in the network. Each packet reception consumes a
constant energy E rx . The sum of transmission energy and
receiving energy associated with one transmission is then
E = E tx + (R2 1)E rx . The total energy consumption for a packet trying to reach a destination at a distance
of x is G(x)E.
Therefore, the total energy consumption is

Etotal
T

Eratio

. Thus the energy eciency ratio is

ef
Epl
ef
Etc

R q(x)S
R q(x)S
8
<
:

A
A

R
R
(x)dx

tc (x)dx
pl

max(1, rL )Epl
pl

max(1, rL )Etc
tc

A
A

, if p=0

q(x)Gpl (x)Epl dx
q(x)Gtc (x)Etc dx
(7)

If we consider transmission power alone, then Eratio can


be up to k4 . If we consider reception power only, then Eratio
can be up to ke2 . In order to illustrate the energy eciency,
we plot it using the following parameters, rpl = 250, rtc = 92
rx
rx
= Epl
=
= 1.56 104 nodes per square meter, and Etc
1 tx
E
.
2 pl
From Figure 3, we see that Topology Control saves energy
when the trac pattern is local. Link layer retransmission
improves only a little bit.
We observe that for trac pattern ranging from local to
random ( = 0 corresponds to trac where the distance
to the destination is picked uniformly), Topology Control

Etotal = q(x)G(x)dx. Since the throughput T measures the number of packets delivered per unit time, the energy consumption per packet delivered E ef is proportional

135

mtx=1, r =92 and r =250


tc

mtx=3, r =92 and r =250

pl

tc

p=0
p=0.1
p=0.2
p=0.5

3.5
3

ratio

2.5

ratio

p=0
p=0.1
p=0.2
p=0.5

3.5

2.5

1.5

1.5

0.5

0.5

0
3

pl

0
3

(a) No link layer retransmission

(b) Up to two retransmissions

Figure 3: Energy eciency Eratio with respect to and packet dropping probability p.
is always better in terms of energy eciency. However, for
unreliable channel model, energy eciency of Topology Control is lower when the trac pattern is to destinations far
away ( > 0). The energy eciency can be improved by
link layer retransmissions.

3.

anism in IEEE 802.11 that deals with the hidden-terminal


problem (RTS/CTS/ACK) assumes that every node has the
same transmission range. This is not the case for networks
using topology control, as each node determines its own
transmission range. Several power ecient protocols [20,
13] that work in the presence of non uniform transmission
ranges of nodes have been proposed. When used in conjunction with topology control, these MAC layer solutions
will further improve energy eciency and throughput. However, to keep our study independent of any MAC layer behavior we use an idealized version of a CSMA/CA protocol using knowledge of transmissions in the nearby region.
For this purpose, we choose to modify the WaveLAN-I [18]
CSMA/CA MAC protocol. Specically, a transmission will
not proceed if it can collide with any receiving packet at
other nodes (i.e. the media is considered busy in that case)
or if the packet will collide at the receiver. We refer to this
MAC layer as the IDEAL MAC layer. Note that, the busy
tone in PCMA [13] can achieve this eect. But the use
of IDEAL MAC may underestimate the collision rate that
would occur on a real MAC.
We assume that the received power drops with distance
with an exponent of 4. That is, for the same received power,
the transmit power is proportional to d4 , where d is the
transmission distance. We assume that the power required
to receive a transmission at the receiver is a constant and
is one-half the maximum transmit power. The carrier frequency is 914MHz, and the transmission raw bandwidth
2MHz. We assume omni-directional antennas with 0dB gain,
and the antenna is placed 1.5 meters above a node. The receive threshold is 94 dBW. The carrier sense threshold
is 108 dBW and the capture threshold is 10 dB. These
parameters simulate the 914 MHz Lucent WaveLAN DSSS
radio interface. The receiving power is 0.1409 W.
Each node has a maximum transmission range of 250 meters which corresponds to a transmission power of 0.2818
W. There have been studies on network lifetime by assuming that each node has a xed amount of energy supply [10,
3]. We focus on the performance of the network before any
node runs out of energy. The nodes are placed uniformly at
random in a square region of 800 by 800 meters.
For our simulations we have used two dierent link error
models. The rst model is a simple error model that corresponds to the one used in our analysis. In this model, the

SIMULATION RESULTS

In our analysis in Section 2, we made some simplifying


assumptions (Section 2.1). The goal of our simulations is
to study the performance of Topology Control and Plain
in realistic settings. We nd that the conclusions from our
analysis are similar to those from the simulations. And, in
most cases the results from the simulations match closely
with the results from the analysis.

3.1 Simulation Setup


In order for topology control to be practical, the algorithm
must be distributed and use only local information, such
as the algorithms in [10, 17, 9] ( [9] is an improvement
over [17]). Li et al. show that [10] and [9] produce similar
topologies. We expect the results to be similar for the two
algorithms [10, 9], and chose to simulate the algorithm in
[10].
The algorithm is implemented in ns-2 network simulator
[16], using the wireless extension developed at Carnegie Mellon [6]. We implement the CBTC algorithm proposed in [10]
with = 2/3 and its optimizations. According to the algorithm, each node independently picks the minimum transmission range such that for every cone of degree from that
node, there is a neighbor in the reduced transmission range.
In practice, frequent message exchanges are needed to maintain these neighbors which can incur a signicant overhead.
We implemented asymmetric edge removal, shrink-back and
pair-wise edge removal optimizations as discussed in [10].
Since our goal is not to study a specic topology control
algorithm but to characterize the performance of topology
control algorithms, we assume perfect knowledge of the location (x-y coordinates) of neighbors. This corresponds to a
long-lived static network where neighbors do not go down or
go away. For the same reason, we also pre-compute shortest
routes based on global information.
The standard IEEE 802.11 does not work correctly with
topology control algorithms. The reason is that the mech-

136

p=0
p=0.1
p=0.2
p=0.5

1.5

Tratio

1.5

Tratio

p=0
p=0.1
p=0.2
p=0.5

0.5

0.5

0
-3

-2

-1

-3

(a) Tratio with no retx.


5

-1

(b) Tratio with max. retx. = 2


5

p=0
p=0.1
p=0.2
p=0.5

-2

p=0
p=0.1
p=0.2
p=0.5

Eratio

Eratio

3
2

0
-3

-2

-1

-3

-2

-1

(c) Eratio with no retx.

(d) Eratio with max. retx. = 2

Figure 4: 100 node unsaturated network: performance with channel error independent of transmission power
(Comparable analytical results shown in Figures 1 and 3)
obtained from analysis, the pattern of the corresponding
graphs are very similar. Similarly for energy eciency, the
simulation graphs of Figures 4 (c) and 4(d) correspond to
the analysis graphs of Figures 3(a) and 3(b). Once again the
patterns from the analysis match very well with the simulations. However, for values of 0, the energy eciency observed in simulations is lower than the results obtained from
analysis. The main reason for this discrepancy is the approximate model used in our analysis that assumes uniform
transmission power across all nodes in the network when
topology control is used. In the simulations, the nodes have
varying transmission ranges depending on the result of applying the topology control algorithm on the local topology.
We observe several interesting facts from these graphs: (1)
Even modest retransmissions such as 2 times signicantly
increases the throughput of Topology Control. (2) Energy
eciency ratio decreases drastically as increases. Energy
eciency ratio is only slightly aected by p and link layer
retransmissions.
We present the results of the saturated case in Figure 5.
As our analysis predicts, neither the packet error rate nor
the number of link layer retransmissions aect the throughput ratio very much. Note that the throughput ratio at
= 3 and p = 0 is around 2.8 (Figure 5(a)). which is
2
less than the ke1
= 4. This is the eect of the variance
of transmission range of Topology Control. Also observe
that the pattern of the graphs in Figure 5 corresponds very
closely to our analysis graphs in Figure 2 and Figure 3. Like

channel error rate is represented by a xed probability. In


our second model, the bit error rate is a function of several
factors such as the ambient noise, the received power and
the raw bit rate. The model also depends on the modulation scheme. This model helps in studying the performance
of Topology Control more realistically. In the remaining section, we present the results based on the two error models.

3.2 Simulation Results: fixed link error model


We vary the trac pattern and show how it aects the
performance of Topology Control over Plain. We consider
the class of trac patterns with power law distance distribution.
We simulate a 100 node network deployed in the 800m
x 800m square region. The measured average transmission
range of Topology Control is rtc1 = 92m. Since rpl is 250m,
r
= 2.72. The
the transmission range ratio is k1 = rtcpl = 250
92
1
interference range of Plain is 550m. For Plain, the measured average eective interference range Rpl is 390m. For
Topology Control, the measured average eective interference range Rtc1 = 187.7. Therefore, ke1 = 2.04, ke2 = 2.80.
First, we present results for the unsaturated scenario. We
plot the simulation results on throughput ratio and energy
eciency ratio with respect to dierent and p in Figure 4. For throughput ratio, the simulation graphs of Figures 4 (a) and 4(b) correspond to the analysis graphs of
Figures 1(a) and 1(b). We observe that although the simulation results are slightly lower than the throughput ratios

137

p=0
p=0.1
p=0.2
p=0.5

2.5

2
Tratio

2
Tratio

p=0
p=0.1
p=0.2
p=0.5

2.5

1.5

1.5

0.5

0.5

0
-3

-2

-1

-3

(a) Tratio with no retx.


5

-1

(b) Tratio with max. retx. = 2


5

p=0
p=0.1
p=0.2
p=0.5

-2

p=0
p=0.1
p=0.2
p=0.5

Eratio

Eratio

3
2

0
-3

-2

-1

-3

-2

-1

(c) Eratio with no retx.

(d) Eratio with max. retx. = 2

Figure 5: 100 node saturated network: performance with channel error independent of transmission power
(Comparable analytical results shown in Figures 2 and 3)
in the unsaturated case, the energy eciency graph shows
that Topology Control does not save energy if the trac
pattern is non-local.

put ratio and energy eciency ratio, as observed in our analysis and previous simulations with simplied received power
independent channel error model. We present a few results
for this channel model and show conformance with previous
observations.
We present the throughput ratios for both saturated and
unsaturated networks with 100 nodes. Figure 6(a) shows
the unsaturated scenario without any retransmissions. As
observed in our analysis and simulation with simple channel model, under low network load, the throughput drops
signicantly with increase in the ambient noise level N .
The ambient noise level eects the packet drop probability.
By adding retransmissions the throughput ratio can be improved. Figure 6(b) shows the unsaturated scenario without
any retransmissions. Once again our observations are very
similar to the simple channel model and analysis. The same
holds for energy eciency (not presented here due to lack
of space). Thus we observe that the performance characterization of topology control using the simple channel model
is applicable to the more realistic channel model as well.

3.3 Simulation Results: ambient noise based


link error model
Our analysis as well as the simulations presented thus far
assumed that the channel error rate is independent of the
received power. However, in reality the bit error rate (pb )
is a function of the modulation scheme, the received power,
the ambient noise and the raw channel bit rate. Assuming
BPSK (binary-phase-shift-keying) modulation 3 , the expression for bit error rate pb , is given by [5]:
pb = 0.5 erf c(

Pr
)
N f

(8)

where N is the noise spectral density (noise power per Hz),


f is the raw channel bit rate which for our simulations is 2
Mbps, and erf c(x) is given by:
2
erf c(x) = 1

et dt

(9)

4. RELATED WORK

We have studied our results under this channel model as


well, and have found the same pattern in terms of through-

Recent work on topology control has concentrated on nding algorithms to improve throughput and energy eciency
while guaranteeing global connectivity using only local position or direction information. In the cone-based topology
control algorithm (CBTC) [19, 10] Li Li et al. have studied,

For modulation schemes OOK and M-FSK an extra constant inside the square root is needed for the expression of
pb

138

N = 2.0e-17 Watt/Hz
N = 2.5e-17 Watt/Hz
N = 3.0e-17 Watt/Hz
N = 3.5e-17 Watt/Hz

1.5

N = 2.0e-17 Watt/Hz
N = 2.5e-17 Watt/Hz
N = 3.0e-17 Watt/Hz
N = 3.5e-17 Watt/Hz

2.5

Tratio

Tratio

2
1

1.5
1

0.5
0.5
0

0
-3

-2

-1

-3

(a) Tratio with no retx. (Unsaturated Network)

-2

-1

(b) Tratio with no retx. (Saturated Network)

Figure 6: With noise dependent channel error model: 100 node network
We would like to incorporate the distribution of the
transmission range of Topology Control into our analysis. Our current analysis assumes that the transmission range of all nodes is uniform, albeit reduced, even
when Topology Control is used.

a node u transmits with the minimum power pu, required


to ensure that in every cone of degree around u, there is
some node that u can reach with power pu, . They showed
that taking = 5/6 is a necessary and sucient condition
to guarantee that network connectivity is preserved. More
precisely, if there is a path from s to t when every node
communicates at maximum power then, if 5/6, there
is still a path in the smallest symmetric graph G containing
all edges (u, v) such that u can communicate with v using
power pu, . On the other hand, if > 5/6, connectivity is
not necessarily preserved. Rodoplu and Meng [17] propose a
distributed position-based topology control algorithm that
preserves connectivity and minimum energy path. Their algorithm is improved by Li Li and Halpern [9]. Ning Li et
al. [11] proposed a minimum spanning tree based topology
control algorithm, where each node builds its local minimum
spanning tree independently and only keeps on-tree nodes
that are one-hop away as its neighbors in the nal topology.
Fault tolerance issues have also been considered recently.
Bahramgiri et al. [1] have shown that, for the CBTC algorithm, it is necessary to have 2/3(kk mod 2) in order
to tolerate k node failures or to have k node-disjoint paths
between any two nodes in the network. XiangYang Li et
al. [12] have shown that a variant of Yao graph can also preserve k node-disjoint paths. None of these algorithms have
characterized the operating region of topology control under dierent trac pattern, load condition and packet error
rate.
Bansal et al. [2] have studied the eect of dierent xed
transmission radius on the throughput of TCP trac and
energy eciency. They have shown that the total energy
consumption is a convex function of the transmission range
due to tradeos between individual packet transmission energy and the likelihood of retransmission. The TCP session
throughput decreases supra-linearly with a decrease in the
transmission range. However, they do not consider other
trac patterns or the eect of topology control.

5.

We have assumed that a transmission reserves an area


within the sensing range. However, in protocols with
RTS/CTS/ACK, the reservation is the union of areas
within the sensing range of source and destination. We
would like to model this type of MAC as well.
Although our simulations have shown that the simplied link error model suces for the purposes of characterizing Topology Control, a more accurate analysis
will be based on a realistic channel error model such
as the one used in part of our simulations.
We would also like to enhance our simulation studies by
incorporating the following:
We have used an idealized version of CSMA protocol
for the MAC layer. We would like to investigate the
eect of a real MAC protocol such as PCMA [13].
Mobility has not been considered in our study and
it remains to be investigated how the relative performance of topology control gets aected when nodes
are mobile.
Our simulation focuses only on UDP trac. The effect of dierent transmission ranges on TCP throughput has been studied in [2]. However, they have also
assumed that all nodes have the same transmission
range. It would be interesting to study how TCP
throughput gets aected by variable transmission ranges
if topology control is used.

6. CONCLUSION
In this paper, we study the performance gains of topology
control using analysis and simulations in terms of throughput and energy eciency under dierent trac patterns,
network load conditions, and packet error rates.
Using the power law trac pattern for UDP ows, we
have shown that topology control improves the overall endto-end network throughput up to a factor of ke2 (where ke is

DISCUSSION AND FUTURE WORK

We have made some simplifying assumptions in our analysis. As part of our future work, we are interested in rening
the presented analytical framework by taking the following
factors into account:

139

[10] L. Li, J. Y. Halpern, P. Bahl, Y. M. Wang, and


R. Wattenhofer. Analysis of distributed topology
control algorithms for wireless multi-hop networks. In
Proc. ACM Symposium on Principle of Distributed
Computing (PODC), pages 264273, August 2001.
[11] N. Li, J. Hou, and L. Sha. Design and analysis of an
mst-based topology control algorithm. In Proc. IEEE
Infocom, pages 17021712, 2003.
[12] X. Li, P.-J. Wan, Y. Wang, and C. W. Yi. Fault
tolerant deployment and topology control in wireless
networks. In Proc. IEEE MobiHoc, pages 117128,
2003.
[13] J. Monks, V. Bharghavan, and W. Hwu. A power
controlled multiple access protocol for wireless packet
networks. In Proc. IEEE Infocom, pages 219228,
April 2001.
[14] C. E. Perkins. Ad Hoc Networking. Addison-Wesley,
Reading, MA, 2001.
[15] G. J. Pottie and W. J. Kaiser. Wireless integrated
network sensors. Communications of the ACM,
43(5):5158, May 2000.
[16] V. Project. The UCB/LBNL/VINT network
simulator-ns (Version 2).
http://www.isi.edu/nsnam/ns.
[17] V. Rodoplu and T. H. Meng. Minimum energy mobile
wireless networks. IEEE J. Selected Areas in
Communications, 17(8):13331344, August 1999.
[18] B. Tuch. Development of WaveLAN, an ISM band
wireless LAN. AT&T Technical Journal, 72(4):2733,
1993.
[19] R. Wattenhofer, L. Li, P. Bahl, and Y. M. Wang.
Distributed topology control for power ecient
operation in multihop wireless ad hoc networks. In
Proc. IEEE Infocom, pages 13881397, April 2001.
[20] S. L. Wu, Y. C. Tseng, and J. P. Sheu. Intelligent
medium access control for mobile ad hoc networks
with busy tones and power control. IEEE Journal on
Selected Areas in Communications, 18(9):16471657,
2000.

the ratio of eective interference range) in saturated scenarios when the trac pattern is local. Beyond random trafc pattern, i.e., when trac becomes more and more nonlocal, topology control degrades the performance in terms
of throughput. Although link layer retransmissions can improve the throughput for topology control signicantly for
local trac, its eect on non-local trac is minimal. The
energy eciency of topology control is independent of network load, and can be up to k4 where k is the average factor
of reduction in transmission range by the use of topology
control. It is not eected much by link layer retransmissions. The energy eciency of topology control decreases as
the trac pattern changes from local to non-local.
Based on our analysis and simulations, we conclude that
sensor networks can benet from the use of topology control
both in terms of throughput and energy consumption, if
the trac pattern in the network is anywhere from local to
random.

7.

REFERENCES

[1] M. Bahrangiri, M. T. Hajiaghayi, and V. S. Mirrokni.


Fault-tolerant and 3-dimensinoal distributed topology
control algorithms in wireless multi-hop networks. In
Proc. the 11th IEEE International Conference on
Computer Communications and Networks (IC3N),
pages 392397, 2002.
[2] S. Bansal, R. Gupta, R. Shorey, I. Ali, A. Razdan, and
A. Misra. Energy eciency and throughput for TCP
trac in multi-hop wireless networks. In Proc. IEEE
Infocom, pages 210219, June 2002.
[3] B. Chen, K. Jamieson, H. Balakrishnan, and
R. Morris. Span: an energy-ecient coordination
algorithm for topology maintenance in ad hoc wireless
networks. In Proc. seventh Annual ACM/IEEE
International Conference on Mobile Computing and
Networking (MobiCom), pages 8596, 2001.
[4] L. P. Clare, G. J. Pottie, and J. R. Agre.
Self-organizing distributed sensor networks. In
Proc. SPIE Conf. on Unattended Ground Sensor
Technologies and Applications, pages 229237, April
1999.
[5] V. K. Garg and J. E. Wilkes. Wireless and Personal
Communications Systems. Prentice Hall, 1996.
[6] C. M. Group. Wireless and mobility extensions to
ns-2. http://www.monarch.cs.cmu.edu/cmu-ns.html,
October 1999.
[7] R. E. Kahn. The organization of computer resources
into a packet radio network. IEEE Transactions on
Communications, COM-25(1):169178, January 1977.
[8] J. Li, C. Blake, D. S. J. D. Couto, H. I. Lee, and
R. Morris. Capacity of ad hoc wireless networks. In
Proc. seventh Annual ACM/IEEE International
Conference on Mobile Computing and Networking
(MobiCom), pages 6169, 2001.
[9] L. Li and J. Y. Halpern. Minimum energy mobile
wireless networks revisited. In Proc. IEEE
International Conference on Communications (ICC),
pages 278283, June 2001.

APPENDIX
In this appendix, we derive the expected number of hop-byhop transmissions G(x) for a packet destined for a node at
a distance of x. Let p be the packet error rate for each hop.
Let mtx be the maximum number of transmission attempts
of the link layer.
Since a transmission fails over a given hop only when all
mtx transmissions fail, the probability of packet transmission failure over a link is p = pmtx . Given that the transmission of a single hop is successful (event ), the conditional
probability that the i-th attempt is successful is given by
i1
P rob(i|event ) = p 1p(1p)
mtx . Therefore, the conditional expected number of transmissions over a single link given event
mtxpmtx
1
is m1 = mtx
i=1 iP rob(i|event ) = 1p 1pmtx . Therefore the expected number of transmissions over a h hop path
i p + h m1 (1 p)
h.
is g(h) = h1
i=0 (m1 i + mtx) (1 p)
x
The expression G(x) is simply g( r ) where r is the transmission range.

P
P

140

You might also like