Você está na página 1de 8

A Comparison of Lifetime-Efficient Forwarding Strategies

for Wireless Sensor Networks

Marcel Busse, Thomas Haenselmann, and Wolfgang Effelsberg


Computer Science IV, University of Mannheim
Seminargebäude A5, D-68159 Mannheim, Germany
{busse, haenselmann, effelsberg}@informatik.uni-mannheim.de

ABSTRACT they cannot communicate with the sink node directly; intermediate
As shown in previous work, energy-efficient forwarding is a very nodes are used for the message forwarding process.
promising forwarding strategy since it trades off the delivery rate For such many-to-one communication, many routing algorithms
of nodes sending data to a sink node and the required energy in were proposed [15, 22, 16, 23] that are based on different routing
the network very well. However, although forwarding paths might metrics. One of the most important challenges is energy efficiency.
be energy-efficient, this does not have to extend the lifetime of the It is expected that most of the sensor nodes carry a limited and ir-
network. Some nodes may consume more energy than others, in- replaceable power supply. As the radio communication consumes
creasing the probability of network partitions. Thus, algorithms a large fraction of energy, the costs of data transmissions are most
maximizing the lifetime were proposed. However, only consider- critical and should be kept to a minimum. Energy-efficient routing
ing the lifetime must not achieve good delivery rates and a low algorithms address this issue, with the goal of either minimizing
energy consumption. The contribution of this paper is a forward- the energy consumption [12, 31, 30], or maximizing the network
ing metric that incorporates the end-to-end delivery, the required lifetime [27, 18, 7]. Generally, routing algorithms that aim at min-
energy costs, and in addition the residual energy of nodes being on imizing the energy costs use shortest path algorithms, with the en-
the forwarding path. Using simulations, we compare our scheme ergy required for the transmission between two adjacent nodes used
with forwarding strategies based on other metrics and give insights as the edge cost. However, just minimizing the energy costs might
on its performance characteristics. have a negative effect on the data delivery rate. A forwarding path
with low energy costs but high loss rates might not be very effi-
cient. On this account, the energy efficiency should incorporate
Categories and Subject Descriptors both energy consumption and the delivery quality.
C.2.1 [Network Architecture and Design]: Wireless Communica- However, although most of the routing algorithms are very ef-
tion; I.6 [Computing Methodologies]: Simulations and Modeling fective, e.g., in reducing the energy consumption or maximizing
the energy efficiency, the overall network lifetime will not neces-
General Terms sarily be maximized. Once the forwarding paths are determined,
they remain stable as long as the costs do not change. Thus, it is
Performance, Design, Algorithms
likely that nodes on the optimal paths are used quite often and run
out of energy quickly. This problem is addressed by maximum-
Keywords lifetime algorithms that aim at preventing the early “burn out” of
Lifetime-Efficient Forwarding, Energy Efficiency, Wireless Sensor such paths. The network lifetime is commonly defined as the time
Networks until the first node runs out of energy [4, 5] since this might cause
malfunctions or network partitions. However, there are also defini-
1. INTRODUCTION tions referring to the time where at least a fraction of α nodes are
alive or, concerning the area covered by the network, at least an α
Consider a network of wireless sensor nodes that are capable of
coverage is maintained [36].
sensing, processing, and communicating [1]. Due to advances in
In comparison to existing work focusing on the maximum-
micro-sensor and radio technologies, sensor nodes use low-cost and
lifetime problem, we emphasize the existence of lossy links. Most
low-power hardware that allows for densely populated networks.
of the work in the literature assume that two nodes are able to com-
Typical applications include habitat or environmental monitoring,
municate with each other without any packet loss as long as their
disaster detection, or military applications. In most cases, sensed
distance is lower than the radio range of the transceiver. However,
events are processed by a node and then forwarded to one or sev-
as many recent measurements have shown [38, 37, 28], most of the
eral sink nodes. Since all nodes are scattered over a wide area,
links in a network experience a high variation in the packet recep-
tion rate. Thus, considering link losses either increases the energy
required for a packet to be delivered due to retransmissions or de-
Permission to make digital or hard copies of all or part of this work for creases the delivery rate in case the number of retransmissions is
personal or classroom use is granted without fee provided that copies are not sufficient. In any case, taking the link loss into account may
not made or distributed for profit or commercial advantage and that copies have a significant impact on the energy efficiency of a routing al-
bear this notice and the full citation on the first page. To copy otherwise, to gorithm and the overall network lifetime.
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
The remainder of this paper is organized as follows. First, we
PE-WASUN’06, October 6, 2006, Torremolinos, Malaga, Spain. outline related work in the next section. Section 3 then describes
Copyright 2006 ACM 1-59593-487-1/06/0010 ...$5.00.

33
the maximum-lifetime problem in conjunction with a linear pro- Tassiulas [9] present a heuristic algorithm to solve the linear pro-
gramming formulation. Based on the Energy-Efficient Forwarding gram approximately. They consider two different models for the
strategy [6] which we review in Section 4, Lifetime-Efficient For- information-generating process, the first assumes constant rates and
warding is presented in Section 5. Models and assumptions made the second assumes an arbitrary generating process. An approx-
in this paper are contained in Section 6. Section 7 shortly describes imation based on the Garg-Konemann algorithm [11] achieves a
other forwarding strategies used in our simulations for compari- (1 − 3)-approximation for any  > 0 [8]. While the heuristic
sion. The simulation results are presented in Section 8. Finally, algorithm of Kalpakis et al. [17] improves this approximation to
Section 9 concludes the paper and gives an outlook on future work. (1 − ), it does not scale with the network size very well. A better
algorithm running time is achieved by Xue et al. [35]. They con-
2. RELATED WORK sider the possibility of data aggregation, and in doing so improve
the running time by K, with K being the number of commodities,
Addressing energy consumption in wireless networks has been
where a commodity represents the data generated by a sensor node
an active research, considering several issues when designing
and delivered to a destination. Sankar and Liu [27] finally formu-
energy-aware networks [21, 24]. There exists much work on power
late the problem as a maximum flow problem and adapt distributed
control algorithms where the goal is to find an optimal transmission
flow algorithms [2, 3].
power for each node without the loss of connectivity [10, 14, 20,
25, 26, 33]. While still maintaining connectivity, the topology that
is formed may aim at improving the network throughput [10, 14], 3. THE MAXIMUM-LIFETIME PROBLEM
to bound the number of one-hop neighbors [25], or to minimize the Consider a directed graph G(N, A) with N being the set of
overall power consumption required for routing [20, 26, 33]. In nodes and A being the set of edges. If there exists a link be-
[20], Lloyd et al. also study algorithmic aspects for minimizing the tween two nodes i and j, then (i, j) ∈ A with prrij denoting
total energy. Distributed algorithms are presented in [26] and [33]. the packet reception rate on that link. Furthermore, Ni defines the
A good survey on topology control issues can be found in [19]. set of neighbors of i, and s refers to the sink node. Let erx and
On the other hand, several approaches are proposed that are not etx be the amount of energy units required to receive resp. send a
designed for topology control but for power-aware or energy-aware packet and Ei be the initial energy of a node i. Then, T refers to
routing, aiming at either minimum-energy routing or maximum- the maximum-lifetime of the entire network, defined as the time in
lifetime routing. The former approach tries to minimize the to- rounds until the first node has no energy left.
tal energy required for a packet to reach its destination. How- If we assume that all nodes are connected and each node sends a
ever, nodes along the optimal path are used very frequently and data packet to the sink every lifetime unit (round), we can formulate
soon run out of energy, likely causing network partitions although the maximum-lifetime problem as a linear programming problem.
many nodes still have enough energy [29]. Furthermore, minimum- Let tij be the number of packets a node i sends to a node j over its
energy routing causes an unbalanced power consumption. This is- entire lifetime. Then, we are searching for optimal values t∗ij that
sue is addressed by algorithms that are designed to maximize the maximize T . Assuming infinite retransmissions of packets, the LP
network lifetime. is of the following form:
While classic routing protocols are based on minimum-hop rout- T → max

X X
LP 1:
ing [15, 16, 22, 23], power-aware algorithms are rather based on
power-based metrics that are used in conjunction with a shortest s.t.
tji τji erx + tij τij etx ≤ Ei , ∀i ∈ N \{s}

X X
path algorithm [30, 31, 32, 18]. Singh et al. propose several met- (1)
rics for power-aware routing in [30], including cost per packet and j:i∈Nj j∈Ni
node, time until the network might get partitioned, consumed en-
ergy per packet and variance in residual energy. In [31], a localized tij = tji + T, ∀i ∈ N \{s} (2)
j∈Ni j:i∈Nj
algorithm is proposed by Stojmenovic and Lin. Based on a node’s
lifetime and distance-based power metrics, they aim to extend a ts,s = T (|N | − 1) (3)
node’s worst-case lifetime. In [32], Tho presents a conditional with τij = 1
being the number of transmissions until a packet
max-min battery capacity routing algorithm that combines the min- prri,j

imum total energy routing and max-min residual energy routing. If is received correctly.
the minimum residual energy along a path is higher than a given Equation 1 covers the energy constraint. Due to simplicity, we
threshold, minimum total energy routing is performed. Otherwise, just consider the energy required to receive and transmit a packet.
the forwarding path is selected such that the minimum residual en- The sink node is assumed to have infinite energy. Equation 2 refers
ergy of a node on the path is maximized. In [18], Li et al. study to the forwarding constraint that specifies that for each node the
the problem of the maximum network lifetime, assuming the mes- number of outgoing packets is equal to the number of incoming
sage flow is not known a priori. They propose an approximation packets plus the number of packets originated at this node. In order
algorithm called max − min zPmin that also tries to trade off be- the make sure that all packets finally reach the sink, Equation 3 is
tween minimum transmission energy and max-min residual energy necessary, too.
routing. First, the algorithm determines the path with the small-
est transmission costs, with Pmin being its required energy and r 4. ENERGY-EFFICIENT FORWARDING
being the minimum residual node’s energy on this path. Then, all In [6], we proposed two forwarding strategies called Single-
edges whose residual energy is smaller or equal to r are removed Link Energy-Efficient Forwarding (SEEF) and Multi-Link Energy-
from the graph. This procedure is repeated until the total energy Efficient Forwarding (MEEF). Forwarding paths are selected based
of the minimum transmission energy path exceeds z times Pmin , on an energy efficiency metric where the efficiency is defined as the
where z ≥ 1. Li et al. have shown that their approximation is very ratio between the end-to-end delivery rate and the energy required
close to the optimal solution they obtained by linear programming. for a message to reach the sink.
The works in [9, 17, 27, 35] also formulate the problem of the With MEEF, we further introduced the concept of multi-link for-
maximum network lifetime using linear programming. Chang and warding. In general, nodes that are not addressed in a packet header

34
as the destination temporarily turn their radio unit off to save en- a packet to a neighbor j, which in turn forwards the packet towards
ergy. However, it might be more efficient if some nodes stay awake the sink.
and overhear the packet, even if this requires more energy. In case For the multi-link case the calculation is quite similar but now
of packet loss, retransmissions of the entire packet might then be considers the fact that before a packet is retransmitted it is possible
prevented. If the actual receiver does not acknowledge the packet, that another node is able to forward it. Thus, ÊlR+1 is iteratively
another node may do so instead. That is, a node first determines computed by
a sorted set of potential forwarders, the forwarding set, which is X
added to the packet header. All nodes contained in this set stay ÊlR+1 = ai,α(j)−1 prrij L1ij + ai,n ÊlR
awake and receive the entire packet. If the first node of the sorted j∈Ωi

forwarding set does not acknowledge the packet, the sending node ..
tries to poll the next node of the set. If there is no node remaining, .
X (7)
the entire packet is retransmitted. The decision which nodes should Ê1 = ai,α(j)−1 prrij L1ij + ai,n L2ij
operate in such a backup mode is due to the MEEF strategy.
Q
j∈Ωi
Based on the definitions already used in the previous section, we
now define the energy efficiency. Let Eir be the end-to-end delivery with ai,k = j∈Ωi ,α(j)≤k (1 − prri,j ).
rate of node i, Eie be the overall required energy to deliver a packet Then, Eil is
to the sink, and Eief f be its efficiency defined as Eir /Eie . Then, for P ai,α(j)−1 prrij L2ij
a sorted set Ωi of n potential forwarding nodes, the MEEF strategy j∈Ωi 2
Eil = + aR+1
i,n Lij (8)
P (1 − ai,n )/(1 − aR+1
i,n )
locally maximizes

Eief f = P j∈Ωi ai,α(j)−1 prri,j Ejr


ai,α(j)−1 prri,j (Eje + b) + ai,n b
(4) for the multi-link case.
With these calculation, the SLEF and MLEF strategies select the
j∈Ωi
forwarder resp. forwarding set of a node i that maximizes Eief f ×
Q
for each node i by examining each neighbor j, with α(j) being the
position of node j in Ωi , ai,k = j∈Ωi ,α(j)≤k (1 − prri,j ), and Eil , with Eief f being obtained from Equation 4.
b = etx + nerx [6].
Intuitively speaking, a node selects an optimal set of helping 6. MODELS AND ASSUMPTIONS
hands from its direct neighborhood with a small amount of effort The loss model used in our simulations captures the packet re-
(the denominator of Expression 4) coupled with a large delivery ception rate (PRR) on a link between two nodes as follows. Below
rate (the numerator). Those neighbors that maximize Eief f are fi- a distance D1 , nodes exhibit full connectivity, i.e., PRR is equal
nally chosen as forwarders. In case of n = 1, we get the SEEF to 1. Nodes are disconnected if they are at least distance D2 away
strategy since only a single neighbor is considered. from each other. In the transitional region between D1 and D2 , the
expected reception rate decreases smoothly with some variation.
5. LIFETIME-EFFICIENT FORWARDING We model this behavior by
While the efficiency of forwarding paths is one aim of a sensor 8>
network, the entire lifetime is another but orthogonal one. Even if <1h i 1
d < D1

>:
D2 −d
all nodes always forward packets along the most efficient paths, P RR(d) = D2 −D1
+X D1 ≤ d ≤ D2 (9)
0
the network lifetime will not necessarily be maximized. More-
0 d > D2
over, one problem that occurs during the lifetime of the network is
that some paths are used more frequently than others. Nodes along with [·]ba = max{a, min{b, ·}} and X ∼ N (0, σ) being a Gaus-
these paths will spent more energy and die pretty soon. We account sian variable with variance σ 2 . For parameters set to D1 = 10,
for that with the Multi-Link and Single-Link Lifetime-Efficient For- D2 = 30, and σ = 0.3, the model shows similar characteristics as
warding (MLEF resp. SLEF) strategies which try to maximize the the link loss model, e.g., presented in [38].
energy efficiency as well as the network lifetime. The simulations are carried out on a stationary network without
Let Êlk be a measure for the expected minimum residual node’s any mobility where nodes are uniformly distributed over a 100 ×
energy on a forwarding path towards the sink if up to k − 1 retrans- 100 m2 field with a maximum radio range of 30 m. We assume
missions are used. ÊlR+1 is then calculated iteratively by that each node already knows the reception rate of all its neighbors,
ÊlR+1 = prrij L1ij + aij ÊlR e.g., through packet reception measurements. Link estimators, as
analyzed in [34], could be used to provide this information.
ÊlR = prrij L1ij + aij ÊlR−1 We consider many-to-one communication with one sink and
.. several sensor nodes reporting data to the sink node periodically.
. (5) MAC-related issues are beyond the scope of this paper, but nodes
Ê1 = prrij L1ij + aij L2ij not participating in packet reception should be able to turn their
communication radio off to save energy. For each packet transmis-
with aij = 1 − prrij , L1ij = min{ei , Ejl }, L2ij = min{ei , ej }, sion, the required energy could be calculated as
and ei being the residual energy of a node i.
With Eil being equal to ÊlR+1 , the lifetime measure of a node i e = etx + n · erx + (N − n) · ehrx (10)
for the single-link case is then
where etx and erx are the amount of energy for packet transmission
prrij L1ij resp. receiving, n the number of addressed receivers, and N the
Eil = + aR+1
ij L2ij . (6) number of neighbors within communication range. ehrx quantifies
(1 − aij )/(1 − aR+1
ij )
the amount of energy required only for decoding the packet header.
In other words, Eil refers to the expectation of the minimum We assume that nodes who are not receivers of a packet turn their
residual energy of nodes that are affected if the node i forwards radios off as soon as they have heard the header.

35
According to the first order model described in [13], etx and erx retransmissions2 . In doing so, the solution can be considered
are defined by as a relaxation of the optimal solution.
etx (d) = eelec + eamp · dα (11)
erx (d) = eelec (12)
8. SIMULATION
We simulated all forwarding strategies for different network den-
for a distance d and a loss exponent α. We set eelec = 50 nJ/bit, sities (number of nodes per radio range), with one sink node and a
eamp = 100 pJ/bit/m2 , and a path loss exponent α of 21 . The fixed number of source nodes that issue data packets. For each
packet size is assumed to be 512 bit. The initial amount of a node’s setting, the simulation was repeated 100 times such that each data
energy is set to 0.1 J. Due to simplicity, we neglect the energy point in the graphs indicates the average value over 100 runs. The
of detecting the packet header as well as the energy required for a number of retransmissions was set to three in all scenarios. Further-
node to stay awake. more, simulation runs last for 1000 rounds, even if nodes run out of
The overall network lifetime is subdivided into rounds and de- energy. We believe that this simulation setup reflects a more real-
fined as the number of rounds until the first node runs out of energy. istic scenario than stopping the simulation as soon as the first node
In each round, a certain amount of nodes issues a data packet that consumes all its energy. Even if some nodes are dead and thus un-
is then transferred along the forwarding path towards the sink. The able to generate or forward data, the rest of the network might still
propagation time on a link and the processing time at a node are be useable. Thus, we think that the total output of a network after
neglected. After each round, the forwarding paths may change due a fixed period is a more realistic performance measure.
to decreasing energy levels. Thus, for all forwarding strategies, the
paths are recomputed periodically. 8.1 Network Performance over Time
Concerning the network lifetime, it is expected that the LP
7. OTHER FORWARDING STRATEGIES achieves the longest lifetime until the first node runs out of en-
ergy. According to the LP 1 formulation, the amount of data sent
We compare SLEF and MLEF with four other strategies that are
by a node i to one of its neighbors j is expressed by tij over the
extended by a lifetime component which incorporates the residual
entire lifetime. Since the order in time packets are sent to neigh-
energy levels of forwarding nodes: (i) Hop-based Forwarding, (ii)
bors does neither change the total amount of consumed energy nor
Er -based Forwarding, (iii) Minimum Transmission (MT) Forward-
the network lifetime itself (as long as each node sends tij packets
ing, (iv) and a globally optimal LP relaxation. Except for the LP
to its neighbor j) we can neglect it. Thus, wePcompute the ratio of
solution, all forwarding strategies are based on a shortest path algo-
packets sent to a neighbor i by t̂ij = tij / k∈Ni tik . Note that
rithm. Therefor, let L = min(j,∗)∈φ {ej } be the minimum residual
then forwarding paths do not change over time. Nevertheless, us-
energy on a forwarding path φ consisting of the set of forwarding
ing several forwarding paths with only the fraction t̂ij of packets
links towards the sink.
(and thus of required energy) leads to the same lifetime and total
consumed energy. However, since all simulations are carried out
• Hop-based Forwarding: Based on a neighbors’ hop counter for 1000 rounds, the LP is recalculated every time a node runs out
h, the node with the smallest h × L−1 value becomes the of energy. Although this approach further relaxes the optimal so-
forwarder. In case of equal values, the node with the best lution we still believe that the LP solution gives us a good upper
reception rate is selected. performance bound.
Figure 1 summarizes several performance metrics that are plot-
• Er -based Forwarding: Defining
Q the the end-to-end delivery
 ted over time. The simulations were carried out using a density of
rate of a node i by Eir = (j,k)∈φ 1 − (1 − prrjk )R+1 ,
20 nodes with 10% of all nodes acting as source nodes that generate
the node that maximizes Eir × L becomes the forwarder of a data packet in each round. Instead of describing each graph on
i. In case of equal values, the node with the smallest hop its own, we rather consider how each forwarding strategy performs
counter is selected. concerning the different metrics. Beginning with the LP solution,
Figure 1(a) shows its significant amount of consumed energy over
• MT Forwarding: MT Forwarding attempts to minimize the
time. However, the minimum residual energy in the network, de-
number of overall transmissions required to successfully de-
picted in Figure 1(b), is always higher than for all other forwarding
liver a packet. For the
P case of infinite retransmissions, the strategies. Thus, it surely achieves a higher network lifetime until
node that minimizes (j,k)∈φ (prrjk ×L)−1 becomes a for-
the first node dies. Considering the amount of nodes being alive
warder. In case of finite retransmissions, MT Forwarding
P 1−(1−prrjk )R+1
over time, as depicted in Figure 1(f), also shows that nodes run
minimizes (j,k)∈φ prrjk
× L−1 . out of energy later in time. Besides the network lifetime, the num-
ber of packets delivered to the sink is another important metric.
• LP Relaxation: In order to provide a globally optimal solu- Figure 1(c) shows the average delivery rate of a source node over
tion concerning the network lifetime, we use the linear pro- time. Although at the beginning the average delivery rate is only
gram described in Section 3. While for infinite retransmis- 80%, the LP solution performs quite well due to longer network
sions LP 1 is sufficient, finite retransmissions requires an in- lifetimes. The minimum delivery rate depicted in Figure 1(d) also
teger program (IP) formulation in order to prevent routing shows that even the worst connected source node in the network
cycles. However, such an IP is quite complex concerning achieves a moderate but sufficient end-to-end delivery rate. Note
its processing time. Thus, we just refer to the LP relaxation that the LP only maximizes the network lifetime and thus actually
achieved by LP 1. That is, although if finite retransmissions does not take the delivery rate of data packets into account at all.
are used during the data forwarding process, the paths are However, even if the LP shows promising results, it uses global
computed based on the solution of LP 1 that assumes infinite knowledge and is just shown for comparison.
1 2
Since we assume that all nodes use a fixed transmission power, we Note that even if infinite retransmissions are allowed, the expected
set d to the maximum radio range of 30 m. number of retransmissions is finite and much lower.

36
1 1
MLEF MLEF
0.95 SLEF 0.9 SLEF
HOP HOP
0.9 ER 0.8 ER
MT MT

Minimum Residual Energy


0.85 LP LP
Residual Energy 0.7
0.8
0.6
0.75
0.5
0.7
0.4
0.65
0.3
0.6
0.55 0.2

0.5 0.1

0.45 0
0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000
Time Time

(a) Residual Energy (b) Minimum Residual Energy

1 1
MLEF
0.9 0.9 SLEF
HOP
0.8 0.8 ER
MT
LP

Minimum Delivery Rate


0.7 0.7
Delivery Rate

0.6 0.6

0.5 0.5

0.4 0.4

0.3 MLEF 0.3


SLEF
0.2 HOP 0.2
ER
0.1 MT 0.1
LP
0 0
0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000
Time Time

(c) Delivery Rate (d) Minimum Delivery Rate

9 1
MLEF
SLEF
8 HOP 0.98
ER
7 MT 0.96
LP
Forwarding Path Length

Fraction of Alive Nodes

6
0.94
5
0.92
4
0.9
3
MLEF
0.88 SLEF
2
HOP
1 0.86 ER
MT
LP
0 0.84
0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000
Time Time

(e) Forwarding Path Length (f) Fraction of Alive Nodes

Figure 1: Forwarding Performance over Time (Density = 20, R = 3)

Hop-based Forwarding shows good results concerning the resid- that most of these nodes die pretty soon which is indicated by Fig-
ual energy, achieving the highest average value at the end of the ure 1(f). Dying nodes are likely close to the sink, thus source nodes
simulation time. Since this forwarding strategy is based on the might get partitioned from the sink, resulting in decreasing deliv-
shortest path measured in hops, it minimizes the forwarding path ery rates. Furthermore, in case of partitions, fewer nodes receive
length that is shown in Figure 1(e). Simultaneously, the number forwarding packets and consume less energy. Due to this, the av-
of nodes forwarding data packets is minimized. Thus, it is as- erage residual energy remains almost constant for simulation times
sumed that only a small fraction of nodes spends energy on packet greater than 700 (see Figure 1(a)). Concerning the average deliv-
forwarding. However, although Hop-based Forwarding consumes ery rate (Figure 1(c) and 1(d)), Hop-based Forwarding achieves the
the least total energy, some nodes consume significant more energy poorest results since in the majority of cases long-distance links
than others as shown by the minimum residual energy depicted in are quite lossy. Due to the same reason, the minimum delivery rate
Figure 1(b). That is due to the fact that most of the links on a for- during the lifetime of the network is very low and drops to zero
warding path experience a high packet loss due to long-distance pretty soon, likely because of network partitions or dying source
links and thus require many retransmissions. Hence, it is likely nodes. The best results concerning the delivery rate are achieved

37
by Er -based Forwarding at the beginning of the simulation time. ergy, both strategies change their forwarding paths, increasing the
However, as it is shown in Figure 1(e), the forwarding paths are number of used hops. Although this affects more nodes, it avoids
very long since short-distance links show better reception rates. forwarding packets over low-energy nodes. Due to this, the fraction
Thus, many nodes are affected by packet forwarding, leading to of nodes being alive over time is better, as depicted in Figure 1(f).
the highest energy consumption in the network (see Figure 1(a)). Again, MLEF benefits from its multi-link concept and further im-
Due to the quickly decreasing amount of residual energy, other for- proves the results achieved by SLEF. It sometimes even performs
warding paths are selected that might not be optimal with respect better than the LP solution. However, note that we used an LP re-
to end-to-end delivery. Interestingly, the forwarding path lengths laxation that does not necessarily represent the optimal case.
become shorter over time due to the high energy consumption on
the optimal path. At the same time, the end-to-end delivery rate 8.2 Node Density
becomes worse. However, the fraction of alive nodes of Er -based We now consider the impact of node density on the network per-
Forwarding performs poorly due to its high total energy costs. formance. The simulation setup is the same as in the last section,
With respect to the average end-to-end delivery, MT Forwarding with a varying node density between 10 and 50 nodes per radio
achieves similar results as Er -based Forwarding. However, con- range. Note that the number of source nodes also varies since the
cerning the worst-connected source node, it performs worse. This fraction of source nodes remains constant.
is actually not surprising since MT Forwarding tries to minimize Figure 2 summarizes the most relevant performance characteris-
the number of packet transmissions regardless of the end-to-end tics of all forwarding strategies. The network lifetime, i.e., the time
delivery rate. Figure 1(a) shows this difference on the basis of con- until the first node runs out of energy, is depicted in Figure 2(a). As
sumed energy where MT Forwarding performs best, except for the expected, the LP solution gives an upper bound on the network life-
Hop-based Forwarding strategy. However, Hop-based Forwarding time which increases for an increasing node density3 . Hop-based
causes network partitions more frequently and thus prevents that Forwarding achieves the lowest network lifetime since individual
some transmissions take place at all. The reduced energy con- nodes spent significant energy on lossy links as we have already
sumption of MT Forwarding is also shown by the forwarding path seen in Figure 1(b). Er -based as well as MT Forwarding improve
lengths depicted in Figure 1(e). Compared to Hop-based Forward- the lifetime of Hop-based Forwarding and almost achieve similar
ing, MT Forwarding uses longer paths in order to avoid lossy links. results. However, SLEF and MLEF consider the residual energy
At the same time, it avoids paths that are too long as they are used of individual nodes in a more efficient way and further extend the
by Er -based Forwarding in order to save energy. Since MT For- time until the first node has consumed all its energy.
warding nevertheless achieves similar delivery rates as Er -based However, extending the lifetime causes higher energy costs as
Forwarding but with significantly less energy, it is more preferable. shown in Figure 2(b). For low node densities, the amount of resid-
Also concerning the number of nodes being alive, MT Forwarding ual energy is still high due to the loss of connectivity after a shorter
performs better due to its lower energy consumption. time period. Thus, disconnected source nodes do not issue any data
At last, we consider the performance of SLEF and MLEF that packets that must be forwarded by other nodes. If the node den-
are proposed in this paper. As it is shown in Figure 1(a), both ap- sity increases, the network will be better connected allowing source
proaches consume more energy than Hop-based and MT Forward- nodes to generate more data packets but at the same time consume
ing. However, until the middle of the simulation time, the con- more energy. Further increasing the node density finally reduces
sumed energy of all four strategies is the same. Then, Hop-based the energy consumption again since eventually a source node it-
and MT Forwarding start to save energy due to dying source nodes self becomes the bottleneck of the forwarding path and runs out of
and network partitions. On the other hand, SLEF as well as MLEF energy.
further consume the same amount of energy per round for a longer Figure 2(c) shows the average number of delivered packets to
time. At the end of the simulation, the total energy consumption of the sink after 1000 rounds. Except for the Hop-based Forwarding
MLEF is higher than the one of SLEF due to the multi-link concept strategy, all other strategies benefit from an increasing node density.
employed by MLEF that uses energy in a more efficient way, thus In contrast, Hop-based Forwarding shows a constant delivery rate
causes network partitions later in time. Concerning the minimum for node densities larger than 15. This is mainly due to the short
residual energy depicted in Figure 1(b), SLEF and MLEF perform lifetime of individual nodes and the usage of long-distance links.
similar to Er -based and MT Forwarding. At first, Er -based For- Furthermore, for higher densities, the source nodes themselves run
warding performs better since it prevents single nodes from spend- out of energy first. For lower densities, the network is not con-
ing too much energy due to long forwarding paths with few trans- nected very well such that the network likely gets partitioned if
missions per link. Later it is slightly outperformed by SLEF and some nodes die. Thus, the delivery rate is significantly lower. Er -
MLEF due to changing forwarding paths used by both approaches. based Forwarding suffers from its high total energy costs, resulting
However, the LP solution still achieves significant better results. in forwarding paths that are no longer optimized according to the
As Figure 1(c) and 1(d) suggests, the best delivery rates over time end-to-end delivery rate. It does not even achieve higher delivery
(average and minimum) are achieved by SLEF and MLEF (maybe rates than MT Forwarding that on the other hand shows a better
except for the LP solution). Almost over the entire simulation time, trade-off of consumed energy and delivered packets. The LP so-
both strategies significantly perform better than Er -based and MT lution again benefits from a longer network lifetime and is able to
Forwarding. The LP solution achieves a lower delivery rate at the deliver more packets than Hop-based, Er -based, and MT Forward-
beginning, but benefits from longer network lifetimes. Due to its ing. However, SLEF performs better for high node densities since
global knowledge, the data traffic is spread throughout the network it considers in addition to the minimum residual energy in the net-
in a more optimal fashion, resulting in a lower delivery rate but work the node’s end-to-end delivery rate, too. Due to the multi-link
a better traffic and energy balancing. Thus, at the end, the total concept, MLEF is able to further improve these delivery rates and
number of delivered packets might be higher than for SLEF and outperforms all other strategies.
MLEF. Regarding the average length of a forwarding path, Fig- 3
ure 1(e) shows that SLEF and MLEF use a similar path length as Since for high node densities the network lifetime of the LP so-
MT Forwarding. However, with a decreasing residual network en- lution exceeds 1000 rounds, Figure 2(a) also shows the time when
the first node would actually be dead.

38
1200 0.8

0.7
1000
0.6
800
Network Lifetime

Residual Energy
0.5

600 0.4

0.3
400
MLEF MLEF
SLEF 0.2 SLEF
200 HOP HOP
ER 0.1 ER
MT MT
LP LP
0 0
10 20 30 40 50 10 20 30 40 50
Node Density Node Density

(a) Network Lifetime (b) Residual Energy

1000 0.18

900 0.16
Average Number of Delivered Packets

800 0.14
700
0.12

Energy Efficiency
600
0.1
500
0.08
400
0.06
300 MLEF MLEF
SLEF 0.04 SLEF
200 HOP HOP
ER 0.02 ER
100 MT MT
LP LP
0 0
10 20 30 40 50 10 20 30 40 50
Node Density Node Density

(c) Delivery Rate (d) Energy Efficiency

Figure 2: Forwarding Performance for Different Node Densities

However, the trade-off between delivery rates and consumed en- tends the network lifetime significantly. For future work we plan to
ergy should not be ignored. We account for this by the energy integrate the approach into our existing Energy-Efficient Forward-
efficiency (ratio of delivered packets and consumed energy units) ing implementation on real sensor node platforms and study the
that is depicted in Figure 2(d). Er -based Forwarding shows the forwarding performance in real-world experiments.
worst energy efficiency due to its significantly higher energy costs.
While Hop-based Forwarding achieves very low delivery rates, it
also saves much energy, leading to a similar energy efficiency as
10. REFERENCES
the LP solution that is able to deliver more data packets but with [1] I. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci.
higher energy costs. This trade-off is best considered by MT For- A Survey on Sensor Networks. IEEE Communications
warding, SLEF, and MLEF. Although SLEF and MLEF both con- Magazine, 40(8):102–114, Aug. 2002.
sume more energy, they still show a slightly better efficiency than [2] B. Awerbuch and T. Leighton. A Simple Local-Control
MT Forwarding. However, the energy efficiency should not be con- Approximation Algorithm for Multicommodity Flow. In
sidered on its own since it gives no information on how much data Proc. of IEEE Symposium on Foundations of Computer
and how long the network was able to deliver data packets to the Science, Palo Alto, CA, Nov. 1993.
sink node. Thus, SLEF and MLEF are preferable due to their high [3] B. Awerbuch and T. Leighton. Improved Approximation
delivery rates and good energy efficiencies. Algorithms for the Multicommodity Flow Problem and
Local Competitive Routing in Dynamic Networks. In Proc.
9. CONCLUSIONS of ACM Symposium on Theory of Computing, Montral,
In this paper, we have introduced Single-Link and Multi-Link Qubec, Canada, May 1994.
Lifetime-Efficient Forwarding which extend our previously pro- [4] M. Bhardwaj, T. Garnett, and A. P. Chandrakasan. Upper
posed Energy-Efficient Forwarding strategies by a lifetime com- Bounds on the Lifetime of Sensor Networks. In Proc. of
ponent that takes the residual energy of forwarding nodes into ac- IEEE International Conference on Communications,
count. We have shown by means of simulations that SLEF as well Helsinki, Finland, June 2001.
as MLEF trade off the entire lifetime of the network, the amount of [5] D. Blough and S. Paolo. Investigating Upper Bounds on
data delivered to the sink node, and the amount of consumed energy Network Lifetime Extension for Cell-Based Energy
very well. Both strategies outperform other forwarding strategies Conservation Techniques in Stationary Ad Hoc Networks. In
that are either based on the number of transmissions, the end-to-end Proc. of ACM Mobicom, Atlanta, GA, Sep. 2002.
delivery rate, or the number of hops on the forwarding path. Even [6] M. Busse, T. Haenselmann, and W. Effelsberg.
the used LP relaxation achieves lower delivery rates although it ex- Energy-Efficient Forwarding Schemes for Wireless Sensor

39
Networks. In Proc. of IEEE International Symposium on a [23] C. Perkins and P. Bhagwat. Highly Dynamic
World of Wireless, Mobile and Multimedia Networks, Destination-Sequenced Distance-Vector Routing (DSDV) for
Niagara-Falls, Buffalo-NY, June 2006. Mobile Computers. In Proc. of ACM SIGCOMM, London,
[7] J. Chang and L. Tassiulas. Energy Conserving Routing in UK, Aug. 1994.
Wireless Ad-Hoc Networks. In Proc. of IEEE INFOCOM, [24] V. Raghunathan, C. Schurgers, S. Park, and M. B. Srivastava.
Tel-Aviv, Israel, Mar. 2000. Energy-Aware Wireless Microsensor Networks. IEEE Signal
[8] J. Chang and L. Tassiulas. Fast Approximate Algorithms for Processing Magazine, 19(2):40–50, Mar. 2002.
Maximum Lifetime Routing in Wireless Ad-Hoc Networks. [25] R. Ramanathan and R. Rosales-Hain. Topology Control of
In Proc. of Networking, Paris, France, May 2000. Multihop Wireless Networks using Transmit Power
[9] J. Chang and L. Tassiulas. Maximum Lifetime Routing in Adjustment. In Proc. of IEEE INFOCOM, Tel-Aviv, Israel,
Wireless Sensor Networks. IEEE/ACM Transactions on Mar. 2002.
Networking, 12(4):609–619, Aug. 2004. [26] V. Rodoplu and T. H. Meng. Minimum Energy Mobile
[10] T. A. ElBatt, S. V. Krishnamurthy, D. Connors, and S. K. Wireless Networks. IEEE Journal on Selected Areas in
Dao. Power Management for Throughput Enhancement in Communications, 17(8):1333–1344, Aug. 1999.
Wireless Ad-Hoc Networks. In Proc. of IEEE International [27] A. Sankar and Z. Liu. Maximum Lifetime Routing in
Conference on Communications, New Orleans, LA, June Wireless Ad-hoc Networks. In Proc. of IEEE INFOCOM,
2000. Hong Kong, China, Mar. 2004.
[11] N. Garg and J. Konemann. Faster and Simpler Algorithms [28] R. Schmitz, M. Torrent-Moreno, H. Hartenstein, and
for Multicommodity Flow and Other Frational Packing W. Effelsberg. The Impact of Wireless Radio Fluctuations on
Problems. In Proc. of IEEE Symposium on Foundations of Ad Hoc Network Performance. In Proc. of IEEE Workshop
Computer Science, Palo Alto, CA, Nov. 1998. on Wireless Local Networks, Tampa, FL, Nov. 2004.
[12] J. Gomez, A. Z. Campbell, M. Naghshineh, and C. Bisdikian. [29] R. C. Shah and J. M. Rabaey. Energy-Aware Routing in
Conserving Transmission Power in Wireless Ad Hoc Mobile Ad Hoc Sensor Networks. In Proc. of IEEE Wireless
Networks. In Proc. of IEEE International Conference on Communications and Networking Conference, Orlando, FL,
Network Protocols, Riverside, CA, Nov. 2001. Mar. 2002.
[13] W. R. Heinzelman, A. Chandrakasan, and H. Balakrishnan. [30] S. Singh, M. Woo, and C. S. Raghavendra. Power-Aware
Energy-Efficient Communication Protocols for Wireless Routing in Mobile Ad Hoc Networks. In Proc. of ACM
Microsensor Networks. In Proc. of the International MobiCom, Dallas, TX, Oct. 1998.
Conference on System Sciences, Maui, Hawaii, Jan. 2000. [31] I. Stojmenovic and X. Lin. Power-Aware Localized Routing
[14] L. Hu. Topology Control for Multihop Packet Radio in Wireless Networks. IEEE Transactions on Parallel and
Networks. IEEE Transactions on Communications, Distributed Systems, 12(11):1122–1133, Nov. 2001.
41(10):1474–1481, Oct. 1993. [32] C.-K. Toh. Maximum Battery Life Routing to Support
[15] C. Intanagonwiwat, R. Govindan, and D. Estrin. Directed Ubiquitous Mobile Computing in Wireless Ad Hoc
Diffusion: A Scalable and Robust Communication Paradigm Networks. IEEE Communications Magazine, 39(6):138–147,
for Sensor Networks. In Proc. of ACM Mobicom, Boston, June 2001.
MA, Aug. 2000. [33] R. Wattenhofer, L. Li, P. Bahl, and Y.-M. Wang. Distributed
[16] D. B. Johnson and D. A. Maltz. Dynamic Source Routing in Topology Control for Power Efficient Operations in
Ad Hoc Wireless Networks. In Imielinski and Korth, editors, Multihop Wireless Ad Hoc Networks. In Proc. of IEEE
Mobile Computing, volume 353. Kluwer Academic INFOCOM, Anchorage, AK, Apr. 2001.
Publishers, 1996. [34] A. Woo, T. Tong, and D. Culler. Taming the Underlying
[17] K. Kalpakis, K. Dasgupta, and P. Namjoshi. Efficient Issues for Reliable Multihop Routing in Sensor Networks. In
Algorithms for Maximum Lifetime Data Gathering and Proc. of ACM SenSys, Los Angeles, CA, Nov. 2003.
Aggregation in Wireless Sensor Networks. Computer [35] Y. Xue, Y. Cui, and K. Nahrstedt. Maximizing Lifetime for
Networks, 42(6):697–716, Aug. 2003. Data Aggregation in Wireless Sensor Networks. Mobile
[18] Q. Li, J. Aslam, and D. Rus. Online Power-Aware Routing in Networks and Applications, 10(6):853–864, Dec. 2005.
Wireless Ad-Hoc Networks. In Proc. of ACM MobiCom, [36] H. Zhang and J. C. Hou. Maximizing α-Lifetime for
Rome, Italy, July 2001. Wireless Sensor Networks. In Proc. of International
[19] X.-Y. Li. Algorithmic, Geometric and Graphs Issues in Workshop on Measurement, Modeling, and Performance
Wireless Networks. Wireless Communications and Mobile Analysis of Wireless Sensor Networks, San Diego, CA, July
Computing, 3(2):119–140, Mar. 2003. 2005.
[20] E. L. Lloyd, R. Liu, and M. V. Marathe. Algorithmic Aspects [37] J. Zhao and R. Govindan. Understanding Packet Delivery
of Topology Control Problems for Ad Hoc Networks. In Performance in Dense Wireless Sensor Networks. In Proc. of
Proc. of ACM MobiHoc, Lausanne, Switzerland, June 2002. ACM SenSys, Los Angeles, CA, Nov. 2003.
[21] R. Min, M. Bhardwaj, S. H. Cho, N. Ickes, E. Shil, A. Sinha, [38] M. Zuniga and B. Krishnamachari. Analyzing the
A. Wang, and A. Chandrakasan. Energy-Centric Enabling Transitional Region in Low Power Wireless Links. In Proc.
Technologies for Wireless Sensor Networks. IEEE Wireless of IEEE Secon, Santa Clara, CA, Oct. 2004.
Communications Magazine, 9(4):28–39, Aug. 2002.
[22] V. D. Park and M. S. Corson. A Highly Distributed Routing
Algorithm for Mobile Wireless Networks. In Proc. of IEEE
INFOCOM, Kobe, Japan, Apr. 1997.

40

Você também pode gostar