Available online at www.sciencedirect.

com

Computer Communications 31 (2008) 17631776 www.elsevier.com/locate/comcom

The tradeo between maximizing the sensor network lifetime and the fastest way to report reliably an event using reporting nodes selection
Fatma Bouabdallah *, Nizar Bouabdallah
INRIA-IRISA, Campus de Beaulieu, 35042 Rennes, France Received 19 March 2007; received in revised form 21 November 2007; accepted 25 November 2007 Available online 14 December 2007

Abstract Energy-eciency is one of the major concerns in wireless sensor networks since it impacts the network lifetime. In this paper, we investigate the relationship between sensor networks performance, particularly its lifetime, and the number of reporting nodes N by using both analytical and simulation approaches. We rst show that the network lifetime and the number of correctly received reports increase when N decreases. Moreover, we demonstrate that the average time required to report an event is a convex function of N. Based on these results, and as a main contribution, we prove that the optimal number of reporting nodes minimizing the energy consumption in the network does not correspond to the optimal number of reporting nodes allowing the fastest way to report an event. The tradeo between these two requirements is therefore specic to each sensor application, depending on its particular needs. In this paper, we provide a simple methodology to achieve this tradeo. 2007 Elsevier B.V. All rights reserved.
Keywords: Energy-eciency; WSN performance; Network lifetime

1. Introduction Energy-eciency is a critical issue in wireless sensor networks (WSNs) due to the limited capacity of the sensor nodes batteries [1]. Indeed, once a WSN is in place, its lifetime must last as long as possible based on the initially provided amount of energy. In view of this, techniques minimizing energy consumption are required to improve the network lifetime. A widely employed mechanism is to schedule sensor nodes activity so that redundant nodes enter the sleep mode as often as possible [2,3]. Based on this concept, several energy-ecient MAC protocols [46] and energy-ecient routing protocols [7,8] have been proposed. Additional solutions, based on congestion control, to
Corresponding author. Tel.: +33 6 79 84 33 78. E-mail addresses: fatma.bouabdallah_othman@irisa.fr (F. Bouabdallah), nizar.bouabdallah@inria.fr (N. Bouabdallah).
*

reduce energy consumption are also proposed in [9,10]. These mechanisms aim at achieving further energy conservation by reducing the energy wastage due to the frequently occurring collisions in WSN networks. The majority of previous works focused mainly on the energy minimization problem. However, minimizing the energy consumption must be achieved while respecting the specic QoS requirements of sensor applications, such as the maximum tolerable time to report an event, and the required event reliability, etc. Indeed, the key performance metrics in WSN networks are both the network lifetime and the average time required to report reliably an event. The optimal solution must therefore take into account these two metrics. In view of this, in this paper, we focus on the analysis of this tradeo. Moreover, the current studies handled the energy optimization issue without paying attention to the impact of the number of reporting nodes on the WSN performances. In other words, given a reporting frequency, how the

0140-3664/$ - see front matter 2007 Elsevier B.V. All rights reserved. doi:10.1016/j.comcom.2007.11.020

1764

F. Bouabdallah, N. Bouabdallah / Computer Communications 31 (2008) 17631776

network lifetime and the reporting latency evolve with respect to the number of active reporting nodes? Our work is motivated by the results in [11,12], which highlight the signicant energy conservation that could be achieved when spatial and temporal correlation is exploited to reduce the number of redundant packet transmission in the network. Specically, a new MAC protocol is proposed in [12] to regulate the network access by limiting the reporting tasks of an observed event to a minimum number of sensor nodes subject to reliability constraints. In this paper, we present an in-depth analysis of the impact of the number of selected reporting nodes on the WSN performances (i.e., network lifetime, reporting latency). Our ultimate goal is to determine the optimal number of reporting nodes that both minimizes the energy required to report reliably an event and respect the latency constraints. The obtained result can be thus used to parameterize the MAC protocol proposed in [12], which can be seen as the feasibility demonstration of the basic access nodes selection method, to achieve the energy-reliabilitylatency tradeos. Moreover, in our study, we relax the minimum boundary constraint on the number of reporting nodes entailed by Vuran and Akyildiz [12], by allowing the selected reporting nodes to transmit redundant information (multiple reports). Doing so, further exibility and thus additional energy conservation could be achieved by our proposal. To the best of our knowledge, we are the rst to tackle the energy optimization problem from this perspective while considering the energy-reliability-latency tradeos. To achieve this, we develop new analytical models to explore the relationship between the WSN performances (i.e., network lifetime, event reporting time) and the number of active reporting nodes, given a predened network reporting frequency. Specically, we analyze the basic access mechanism of IEEE 802.11 DCF (distributed coordination function) with its optional request-to-send/clearto-send (RTS/CTS) scheme [13]. This basic protocol is used by the currently deployed WSNs to arbitrate access, among multiple sensor nodes, to the shared medium in order to communicate with the sink node. We rst derive the expression of the collision probability and the average energy required to report reliably an event as functions of the number of reporting nodes and the reporting frequency. Based on these results, and as a rst main contribution of this paper, we prove that the network lifetime increases when decreasing the number of active reporting nodes. We show that the maximal network lifetime is achieved when only one reporting node is activated while the remaining nodes undergo the sleep mode. Indeed, in doing so, collisions among reporting nodes is avoided, eliminating thus unnecessary energy consumption. We then show analytically that the time required to report an event is a convex function of the number of active reporting nodes N, where the minimum is obtained for N opt > 1. Consequently and as a second main contribution, we dem-

onstrate that the fastest way to report reliably an event does not necessarily lead to the most ecient energy consumption. The tradeo between these two requirements (i.e., energy consumption and reporting time) depends mainly on the specic QoS needs of the sensor application. Finally, in order to assess the accuracy of our proposed analytical model, we conduct simulations. To do so, we develop our own evaluation environment using NS-2 [14]. The results show a good match between simulations and the analytical expressions, which conrms the accuracy of our models. The rest of the paper is organized as follows. Section 2 discusses the related works and Section 3 presents the general problem statement. Communications among sensor nodes will be outlined in Section 4. In Section 5, we introduce the mathematical models to evaluate the impact of the number of reporting nodes on the WSN performances. Analytical and simulation results are discussed in Section 6. In Section 7, we provide a simple methodology to achieve the tradeo between energy consumption and event reporting latency. Finally, Section 8 concludes this paper. 2. Related work As stated before, in order to minimize the energy consumption in WSNs, several energy-ecient MAC protocols [46] and energy-ecient routing protocols [7,8] have been proposed in the literature. These schemes aim to decrease the energy consumption by using sleep schedules. The key idea behind this concept is to turn o completely some parts of the sensor circuitry (e.g., microprocessor, memory, radio) when it does not receive or transmit data, instead of keeping the sensor node in the idle mode. This scheme simply attempts to reduce wasted energy due to idle listening, i.e., lost energy while listening to receive possible trac that is not sent. To do so, works in [46] suggest wake-up scheduling schemes at the MAC layer to activate sleeping nodes when it is needed. On the other hand, works in [7,8] address the problem at the network layer by proposing new routing solutions that take into account the sleep state of some network nodes. To maximize the network lifetime, works in [15,16] try to use a cross-layer-based solution. Their aim is to nd the optimal setting parameters in physical, MAC and routing layers to maximize the network lifetime in the context of collision-free wireless sensor networks. In fact, in [15], the authors try to investigate the inherent tradeo between the application layer performance and the network lifetime by developing a theoretical framework based on an optimization problem. The proposed problem aims at optimizing both the application performance and the network lifetime in a wireless sensor networks with orthogonal link transmissions. To do so, Nama et al. dene the application layer performance as a network utility function which is the sum of individual node utilities. In fact, the utility function of each node mono-

F. Bouabdallah, N. Bouabdallah / Computer Communications 31 (2008) 17631776

1765

tonically increases with the intrinsic source rate of that node. Resolving this cross-layer optimization problem leads to compute the optimal set of source rates, network ows and radio resources at the transport, network and radio resource layers, respectively. Moreover, the authors in [16] consider the joint optimal design of physical, medium access control (MAC), and routing layers to maximize the lifetime of energy constrained wireless sensor networks. The problem of computing a lifetime-optimal routing ow, link schedule and link transmission powers is formulated as a non-linear optimization problem. In their study, the authors consider both orthogonal and non-orthogonal link schedules. In the case of general link schedules, the authors proposed an iterative algorithm to approximate the optimal solution. This algorithm was shown to be quite accurate in linear and rhombus topology but not in large topologies. Although the signicant energy saving achieved by such schemes, the WSN keeps always sending redundant data. Typically, WSNs rely on the cooperative eort of the densely deployed sensor nodes to report detected events. As a result, multiple sensor nodes may report the same event. To further decrease energy consumption, several works are now focusing on the elimination of the useless redundant information [11,12,17,18]. The reduction of the number of redundant packets can be achieved either at the data originator level (i.e., sensor nodes that detect the event) [11,12], by regulating their access, or at the intermediate sensor nodes routing the information to the sink, by means of aggregation mechanisms [17,18]. In the latter case, paths from dierent sources to the sink form an aggregate tree, where the redundant data at the branching nodes are replaced by a single message. In doing so, the number of packets traversing the network is considerably reduced, which leads to signicant energy conservation [17,18]. However, such schemes aect the reliability of the information transmitted to the sink. The aggregation process at intermediate nodes must therefore be aware of these reliability constraints [10], which may become in case of challenging multiple aggregation points in the route to the sink. Reducing the redundant information is more ecient when it is realized at the source nodes [12]. This is achieved by limiting the number of access reporting nodes. Specically, [12] shows that using a small subset of the nodes (called representative nodes) rather than all the sensor nodes in the event area, to report the detected event reduces considerably the energy consumption. Indeed, limiting the number of reporting nodes alleviates the energy wastage caused by collisions, idle listening and redundant packet transmission. In the optimal case, only one node will be allowed to report the detected event. In such case, collisions, idle listening and redundant packet transmission are totally eliminated. But, such choice may not guarantee the required reliability since only one report is received by the sink regarding the observed event.

In this paper, we analyze the impact of the number of selected reporting nodes on the WSN performances (i.e., network lifetime, reporting latency). We aim at deriving the optimal numbers of reporting nodes that minimizes the energy consumption while fullling the reporting latency constraint. In this regard, as a second stage in our optimization process, we have to make sure that the adopted number of reporting nodes ensures the maximum tolerated latency. To the best of our knowledge, this is the rst eort that considers the triplet energy-reliabilitylatency constraints. It is worth noting that we will adapt the MAC proposed in [12] to meet our specic requirements. To summarize, our study enables to derive the optimal setting parameters (number of reporting nodes N and the reporting frequency f), to be used later by protocols such as [12], in order to achieve the energy-reliability-latency tradeos. 3. Problem statement Let us consider the WSN as depicted in Fig. 1. In essence, a WSN ensures the supervision of a given area by the use of a sink node, which collects reports from the network. In this analysis we consider event detection driven wireless sensor applications. In other words, communications are triggered by the occurrence of a pre-specied type of events. Once an event occurs, it has to be reported to the sink by the sensor nodes. In such network, sensor nodes, within an event radius Rc , are the sources (i.e., reporting nodes) for the detected event. Recall that sensor nodes are characterized by their coverage range Rc (i.e., sensing range) and transmission range Rt as shown in Fig. 1. Henceforth, we denote by N the number of reporting nodes for a detected event. Moreover, we denote by f the network reporting frequency. The network reporting frequency is dened as the number of packets generated per unit of time by the network to report an event. Hence,

Fig. 1. Example of a sensor network.

1766

F. Bouabdallah, N. Bouabdallah / Computer Communications 31 (2008) 17631776

given N reporting nodes, the reporting frequency of each sensor node must be set equal to fs f =N to get the predened network reporting frequency. This parameter f is generally xed by the network administrator in order to achieve required event detection reliability, R. The desired event reliability, R, is the number of data packets required by the sink to consider the event as reliable [10]. Once the sink node receives R reports, it instructs the sensor nodes to stop the event reporting. In this study, we aim at analyzing the impact of the number of reporting nodes N on the WSN performance. The basic idea is to let some potential reporting nodes to enter a sleep mode. In the extreme case, we only let one sensor node (N 1) to report a detected event with a reporting frequency fs f . Furthermore, we evaluate the collision probability, the average time required to report an event and the network lifetime as a function of the number of active reporting nodes N.

RTC
Contention window DIFS source Backoff Slot destination SIFS CTS SIFS ACK RTS SIFS DATA

Fig. 2. Basic access mechanism of IEEE 802.11 DCF.

CW value is reset to CW min . Fig. 2 illustrates the IEEE 802.11 DCF access mechanism. 5. Performance analysis In this section, we present mathematical models to derive both the WSN lifetime and the average time required to report an event, as functions of the number of reporting nodes N and the reporting frequency f. To achieve this, we rst calculate the collision probability in such networks caused by the multiple reporting nodes. Then, we derive the average time required to report reliably an event (i.e., to transmit R reports). Based on this result, we can easily obtain the associated consumed energy. Finally, using the analytic formula given in [19], we obtain the expression of the WSN lifetime. In this study, we distinguish between two modes of functioning according to the network reporting frequency f: the saturated and unsaturated regimes. The rst mode is obtained when f is high enough. In this case, each time the channel is free for transmission, each station among the N reporting ones has at least one report to transmit. In other words, for each new transmission cycle, all the reporting nodes compete to access the common channel. In contrast, in the unsaturated regime, it may happen that the channel remains free. This is the case when f is chosen to be relatively low. Generally, for the sake of simplicity, previous works limited their studies to the saturated mode. The unsaturated mode was considered in few paper, such as in [10], using simulations. In this work, we develop an analytical model for each mode. To our knowledge, we are the rst to study analytically the WSN performance under the unsaturated regime. Our aim is to analyze the impact of the reporting frequency f , in addition to N, on the WSN performances. 5.1. Probability of collision in IEEE 802.11 DCF-based MAC protocols In this subsection, we derive the collision probability as a function of the number of reporting nodes N and the reporting frequency f considering both the saturated and unsaturated regimes. We note that collision is a key factor that impacts the total energy consumption as well as the time required to report an event. In fact, the more frequent the collisions are, the more time and energy are spent to report an event.

4. Wireless sensor networks As stated before, communications in currently deployed WSN are carried using the basic IEEE 802.11 DCF protocol and its optional RTS/CTS mechanism. Specically, once an event is detected, the N active reporting nodes compete to access the common data channel to report the event to the sink. The IEEE 802.11 DCF access method is based on the CSMA/CA technique. Accordingly, a host, wishing to transmit a frame, rst senses the channel activity until an idle period equal to Distributed Inter Frame Space (DIFS) is detected. Then, the station waits for a random backo interval before transmitting. The backo time counter is decremented in terms of time slots as long as the channel is sensed free. The counter is suspended once a transmission is detected on the channel. It resumes with the old remaining backo interval when the channel is sensed idle again for a DIFS period. The station transmits its frame when the backo time becomes zero. In this case, the host starts the process by sending a RTS frame. If the frame is correctly received, the receiving host sends a CTS frame after a Short Inter Frame Space (SIFS). Once the CTS frame is received, the sending host transmits its data frame. If the sending host does not receive the CTS frame, a collision is assumed to have occurred. In this case, the sending host attempts to send the RTS frame again when the channel is free for a DIFS period augmented by the new backo, which is calculated as follows. For each new transmission attempt, the backo interval is uniformly chosen from the range 0; CW in terms of time slots. At the rst transmission attempt of a frame, CW equals the initial backo window size CW min 31. Following to each unsuccessful transmission, CW is doubled until a maximum backo window size value CW max 1023 is reached. Once the frame is successfully transmitted, the

F. Bouabdallah, N. Bouabdallah / Computer Communications 31 (2008) 17631776

1767

5.1.1. Probability of collision in the saturated regime Assume N reporting stations contending to access the common channel. In saturation conditions, each station has always a report to transmit. Thus, when we reach the saturated regime the collision probability becomes insensitive to the reporting frequency f. In other words, 8f P fsaturation the collision probability remains the same. In this case, a collision occurs when two or more backo counters Bi i 1; . . . ; N of dierent stations expire at the same time. Hereafter, we assume that the number of transmissions that are subject to multiple successive collisions is negligible. This assumption, denoted henceforth by assumption 1, is widely used in the literature to simplify the analytical models. Accordingly, following to a successful transmission, we can also assume that the backo Bi i 1; . . . ; N of each reporting station takes a value in 0; CW min . This second assumption (assumption 2) holds since we omit successive collisions occurrence as explained in [20]. The accuracy of these approximations is justied, as it will be demonstrated in Section 6, through the perfect match between the analytical and simulation results. Let us now calculate the probability of collision occurrence P col sat N when reporting an event, that is during a Reporting Transmission Cycle (RTC). The RTC is dened as the time spent between two successive acknowledgment transmissions by the sink node. Recall that the sink node sends an acknowledgment frame after the reception of each report. In other words, RTC is the time required by the WSN to report an event to the sink. Before we delve in calculations, it is important to note that our model gives simple expression and more accurate results of the collision probability than [20]. As we neglect multiple successive collisions occurrence, during an RTC cycle, a report can be either successfully transmitted from the rst attempt (Fig. 3a), or following to a rst collision (Fig. 3b). Hence, a collision can only occur at the beginning of the RTC cycle with a probability P col sat N P c N , where P c N is the probability of collision among N competing access nodes with their associated backos Bi i 1; . . . ; N ranging between 0; CW min . A collision occurs when several backo counters expire at the same time. Hence, the probability of collision P c N can be written as follows: P c N PrfU g
CW min X k 0

PrfX k ; U g

 N  X N CW min k N i
i2

CW min 1

5.1.2. Probability of collision in the unsaturated regime In the unsaturated regime, the reporting frequency of each station is relatively low. Specically, a reporting node may have no report to transmit at the beginning of a new RTC cycle. Recall that previous works limited their study to the saturated regime omitting the unsaturated one. In other words, collision probability is always calculated considering the saturated regime. So, to the best of our knowledge, we are the rst to derive this parameter in the unsaturated regime. In order to compute the probability of collision, we assume that all the reporting nodes detect an occurring event exactly at the same time. Then, they will try to send new reports each T 1=fs units of time until the desired event reliability R is attained. In the unsaturated regime, we deal with successive cycles of T units of time. We suppose that during each cycle, N reports are transmitted to the sink (see Fig. 4). Each cycle of T units of time is thus composed of N successive RTC cycles, corresponding to the N reports transmission, followed by an idle period. This idle period is interrupted, and thus the next cycle T begins, as soon as the reporting nodes generate their new reports. In this regard, T can be expressed as follows: T
N X i1

RTCi idle period

where RTCi corresponds to the time required by the WSN to report an event to the sink when the number of active reporting nodes (i.e., nodes that have not yet transmitted their reports) is i. Specically, at the beginning of a cycle T, all the N reporting nodes generate new reports to transmit to the sink. Immediately, the reporting nodes leave their idle states (see Fig. 4) and proceed according the DCF algorithm, described in Section 4, to transmit their reports. Let us now calculate the collision probability among access nodes trying to report the detected event. It is dened as the probability of collision when there is at least one packet to be sent by the N reporting nodes. Hence, the probability of collision can be written as follows: P col
unsat N

PrfX k ; U g

PrfCollision occursjY P 1g

where the random variable X denotes mini2h1;N i Bi and the event U is dened as follows: U f9i; j 2 h1; N i; i 6 j; Bi Bj X g fCollided transmissiong 2

where Y denotes the stationary state of the stochastic process fY t; t P 0g; which represents the number of reporting nodes still having a packet to transmit. By conditioning on the stationary state Y , we get:
P col unsat N
N X i1

The event fX k ; U g simply implies that the backo counter becomes zero for the rst time in k slots for at least two stations, which leads to a collision occurrence. Thus, PrfX k ; U g can be derived as follows:

PrfCollision occursjY ig PrfY ijY P 1g

where PrfCollision occursjY ig can be derived simply using (1) as follows:

1768

F. Bouabdallah, N. Bouabdallah / Computer Communications 31 (2008) 17631776

T(N) = tov + t1cont + ttr ttr

ACK

DIFS t1cont source

RTS

SIFS

DATA

SIFS destination

CTS

SIFS ACK

ACK

DIFS remaining nodes

BUSY MEDIUM

DIFS

RTS

T(N)

t' 2

t'' 2-twait t RTS DIFS RTS DATA ACK

ACK

DIFS

twait

Collided node 1 sent RTS

ACK

DIFS

RTS

twait

DIFS

BUSY MEDIUM

Collided node 2

ACK

DIFS

RTS

EIFS = twait2 slots

DIFS

BUSY MEDIUM

received Non Collided node collided RTS

Fig. 3. The reporting transmission cycle RTC.

N 1 X P c i N i 1

t0 N reports

t0+T
N reports

t0+2T
N reports

t0+3T

t0+1/M

P col

unsat N

RTC(N)

idle

RTC(i)

RTC(1)

5.2. Average time to report an event

cycle2 cycle3

cycle1

Fig. 4. Transmission cycles in the unsaturated regime.

PrfCollision occursjY ig P c i

On the other side, the expression of PrfY ijY P 1g, is simply given by: RTCi 1 6 PrfY ijY P 1g PN N j1 RTCj Finally, we derive the expression of the collision probability as follows:

In this section, we evaluate the average time of an RTC in both saturated and unsaturated regimes. It is the mean time required by the WSN to successfully transmit a report when the number of reporting nodes equals to N. 5.2.1. Average time to report an event in the saturated regime Considering the network under the saturation conditions, the overall transmission time of a report (i.e., RTC) can be written as follows: T sat N ttr tov tcont N 8

F. Bouabdallah, N. Bouabdallah / Computer Communications 31 (2008) 17631776

1769

where ttr is the transmission time of the data packet and tov is a constant overhead, which can be simply deduced from Fig. 2, and thus given by: tov DIFS tRTS 3 SIFS tCTS tACK 9

Moreover, tcont N represents the average time spent in contention procedure when N reporting nodes compete for the medium access with their associated backos Bi i 1; . . . ; N ranging between 0; CW . In other words, it is the extra time lost due to the collision occurrence. Hereafter, we focus on tcont N calculations. As stated before, we neglect in our study the successive collisions occurrence. Doing so, we distinguish between two cases: Case 1: The report is transmitted successfully by one of the reporting nodes from the rst attempt (i.e., following a successfully transmitted report) (See Fig. 3a). Case 2: The report is transmitted successfully by one of the reporting hosts following to a rst collision occurrence on the medium (See Fig. 3b). (a) Case 1: This case happens with a probability 1 P col sat N . In this case, tcont N t1cont N is simply the average backo time spent by the transmitting node, denoted by node j, before accessing to the data channel (See Fig. 3a). According to assumption 2 (Refer to Section 5.1.1), all the reporting nodes backo counters take values in 0; CW min at the beginning of an RTC cycle. Moreover, as the report is successfully transmitted, the transmitting node j has certainly the minimum backo value among the N competing access nodes (i.e., X Bj ). In addition 8i 6 j, we have Bi > Bj . Let U denote that event: U f9!j 2 h1; N i; Bj X g fSuccessful transmissiong 10 Note that PrfU g 1 PrfU g Doing so, t1cont N can be expressed as follows: t1cont N EX jU Slots where EX jU EX ; U =PrfU g Moreover, EX ; U can be written as follows: EX ; U
CW min X k 0

(b) Case 2: In this case, the report is successfully transmitted by one of the reporting nodes after a rst failed attempt. Such case happens with a probability P col sat N . tcont N t2cont N is therefore the sum of the time spent from the beginning of the RTC cycle until the end of the transmission of the collided RTS frame t02 N and the average backo time required by the new transmitting node j to access to the channel in order to transmit correctly another RTS frame t00 2 N (See Fig. 3b). Hence, we get: t2cont N t02 N t00 2 N And we have: t02 N DIFS tRTS EX jU Slots 17 16

where EX j U is the average backo time of the collided stations. It can be simply derived using the fact that EX j U EX ; U =PrfU g: Doing so, we have: Bc N E X ; U
CW min X k 0

k PrfX k ; U g

18

11

12

13

k PrfX k ; U g

14

where PrfX k ; U g is given by (3). Let us now focus on the calculation of t00 2 N . As we mentioned before, a collision can occur only when N c N c P 2 stations send RTS requests at the same time. The N c collided stations perceive the collision as they do not receive the CTS frame from the sink after tCTS SIFS twait units of time. On the other side, the remaining N N c nodes, which did not participate in the collision, detect immediately the collision occurrence as they receive a collided RTS frame and they will wait for a period of time equal to EIFS twait2 Slots before attempting again to access the channel. In this case, starting from the collision occurrence backo counters of these N N c nodes take values in twait2 ; twait2 CW min . On the other hand, the backo windows of the N c collided stations double. Accordingly, the backo counters of the collided stations take values in 0; 2 CW min . However, these stations have to wait for a period of time approximately equal to 11 slots corresponding to twait tCTS SIFS before they try again to access to the data channel. Hence, starting from the collision occurrence, the backo counters of the N c collided stations vary between twait ; twait 2 CW min , whereas the remaining nodes backo counters vary between twait2 ; twait2 CW min . Let the random variable X 0 denote mini2h1;N i Bi and U 0 be the following event: U 0 f9!j 2 h1; N i; Bj X 0 g fSuccessful transmissiong 19 We recall that we aim at calculating t00 2 N , which is the average backo time required by the WSN to successfully transmit a new report after a rst failed attempt. t00 2 N can be therefore written as:

where PrfX k ; U g can be simply derived based on (3): PrfX k ; U g PrfX k g PrfX k ; U g   N CW min k N 1 1 CW min 1N

15

1770
0 0 t00 2 N E X ; U j U

F. Bouabdallah, N. Bouabdallah / Computer Communications 31 (2008) 17631776

20

which leads to: t00 2 N

twait 2X CW min 1 k twait

k PrfX 0 k ; U 0 jU g t00 2 N , 0

21

And PrfN c n j U g is already given by (25). Sub-case (b.2): The transmitting host j did not participate in the rst collision. Such event is denoted by C . In this case, we have: PrfX 0 k ; U 0 ; C jU g
N X n2

PrfX 0 k ; U 0 ; C jN c n; U g 29

In order to calculate we have rst to derive the expression of PrfX 0 k ; U j U g. To achieve this, three cases are to be distinguished according to the value of X 0 (in term of time slots). (i) twait 6 X 0 k < twait2 In this case, the host j that accesses the medium is one of the N c collided stations. Using the theorem of total probability, we get: PrfX 0 k ; U 0 jU g This yields to
PrfX k ; U jU g
0 0 N X n2

PrfN c njU g where

PrfX 0 k ; U 0 ; C jN c n; U g   N n CW min twait2 k N n1 2 CW min twait k n 1 CW min 1N n 2 CW min 1n 30

N X n2

PrfX 0 k ; U 0 ; N c njU g

22

Putting both sub-cases together, we get the expression of PrfX 0 k ; U 0 j U g when twait 6 X 0 k 6 CW min as follows:
PrfX 0 k ; U 0 jU g PrfX 0 k ; U 0 ; C jU g PrfX 0 k ; U 0 ; C jU g 31

PrfX k ; U jN c n; U g PrfN c njU g 23

Since the transmitting node j participate in the previous collision, we have: PrfX 0 k ; U 0 jN c n; U g   n 2 CW min twait k N n1 1 2 CW min 1N n Moreover, we have: PrfN c njU g where PrfN c ng
CW min X k 0

(iii) CW min twait2 < X 0 k < twait 2 CW min This case happens only when all the N reporting nodes participated in the rst collision (i.e., N c N ). Thus, we have: PrfX 0 k ; U 0 jU g PrfX 0 k ; U 0 ; N c N jU g 32

24

This leads to PrfX 0 k ; U 0 jU g PrfX 0 k ; U 0 jN c N ; U g PrfN c N jU g 33

PrfN c n; U g PrfN c ng PrfU g PrfU g  

N n

25

where PrfX 0 k ; U 0 jN c N ; U g   N 2 CW min twait k N 1 N 1 2 CW min 1

N n

CW min k N CW min 1

26

34

(ii) twait2 6 X 0 k 6 twait2 CW min In this case the host j that accesses the channel may be either one of the N c stations, which already participated in the rst collision, or belongs to the N N c remaining ones. Accordingly, we distinguish between two sub-cases: Sub-case (b.1): The transmitting host j already participated in the rst collision. Such event is denoted by C. In this case, we have: PrfX 0 k ; U 0 ; C jU g
N X n 2

And PrfN c N jU g is already given by (25). Moreover, using (21), (22), (31) and (33), we simply derive t00 2 N and thus we get the expression of t 2cont N by means of (16). Doing so, we nally derive the expression of tcont N , which is given by: tcont N 1 P c N t1cont N P c N t2cont N 35 By substituting (35) in (8), we get the average time required to report an event in the saturated regime when the number of reporting nodes is N. Hence, the average time needed to report reliably an event is: T sat reliably N R T sat N :

PrfX 0 k ; U 0 ; C jN c n; U g 27

PrfN c njU g where

PrfX 0 k ; U 0 ; C jN c n; U g   n CW min twait2 k N n 2 CW min twait k n1 1 CW min 1N n 2 CW min 1n 28

5.2.2. Average time to report an event in the unsaturated regime In the unsaturated regime, the reporting frequency of each station is low enough so that the medium remains free for a period of time after the transmission of N reports. As 1 stated before, we deal with successive cycles of T f units

F. Bouabdallah, N. Bouabdallah / Computer Communications 31 (2008) 17631776

1771

of time. During each cycle N reports are transmitted to the sink (see Fig. 4). Specically, at the beginning of a cycle, all the N reporting nodes compete to access the channel, similarly to the saturated regime. Once the rst report is successfully transmitted, the remaining N 1 nodes compete again to access the data channel. Then, once the next report is transmitted successfully, the N 2 remaining nodes compete once more to access the medium, and so on and so forth until all the reports are transmitted. Hence we get: T unsat N
N 1 X T sat i N i1

our model, each sensor node can listen to all the other sensors. In view of this, ERTC sat N can be simply written as follows: ERTC
sat N

Etr N Eov N Econt

sat N

39

where Etr N is the amount of energy consumed during the transmission of a data packet (i.e., during ttr ) by the N active reporting nodes, Eov N is the amount of energy consumed during the constant overhead period of time tov by the N active reporting nodes, and Econt sat N is the amount of energy spent in contention procedure under the saturated regime by the N active reporting nodes. 8 Etr N tr Etr N 1Erx > > > > > Eov N Eidle N DIFS 3 SIFS > > > > > Erx N 1tRTS N tCTS N tACK > > > > < Etx tRTS > E N 1 P c N N Eidle t1cont N cont sat > > > > > P c N Eidle N t00 > 2 N DIFS > > P > N > Bc N Slots n2 PrfN c njU g > > > : nEtx tRTS N nErx tRTS 40 These quantities are expressed as shown in (40). Let us now calculate the continuous energy consumption Ec considering the desired event reliability R and the mean number of events occurring by unit of time M. We 1 assume M P R T sat N , that is the mean time between 1 two successive events occurrence i:e:; M is higher than the mean time required to report reliably an event i:e:; T sat reliably N . Hence, we get: Ec N 1 R M T sat N Eidle 41

36

The average time required to report reliably an event in the unsaturated regime can be therefore written as follows:   1 R R T unsat reliably N ; f b c R b cN T unsat N f N N 37 5.3. Sensor network lifetime In this paper, the network lifetime T network lifetime N , when considering N reporting nodes are activate, is dened as the time spent from the deployment until the network becomes unable to report events due to the lack of energy. Typically, T network lifetime N depends on the total initially provided amount of energy Einitial , the rate of event occurrence M, the reporting frequency f and the desired reliability RN . Based on [19], the average network lifetime can be expressed as follows: T network
lifetime N ; f

Einitial Ew kERTC N ; f Ec

38

where Ec is the constant continuous energy consumption per unit of time needed to sustain the network during its lifetime without data collection, Ew is the expected wasted energy (i.e., the total unused energy in the network when it dies), k is the average sensor reporting rate dened as the number of transmitted reports by the WSN per unit of time (i.e., R M ) and ERTC N ; f is the expected reporting energy consumed by all the sensors to report an event. In the remainder of this paper, we ignore Ew : Indeed, Ew is negligible when we achieve balanced energy consumption across the network. Hence, to derive the network lifetime, we only need to calculate ERTC N ; f and Ec : 5.3.1. Sensor network lifetime in the saturated regime Hereafter, we rst evaluate the average energy ERTC sat N required by the WSN to successfully transmit a report under the saturated regime (i.e., during an RTC cycle) when the number of active reporting nodes equals to N. To achieve this, we take into account transmitting, listening and idle energies. We denote by Eidle , Etr and Erx the consumed energy per unit of time during idle, transmitting and listening states, respectively. As shown in Fig. 1, in

The network lifetime in the saturated regime can be therefore expressed as follows: T network
lifetime N

Einitial M R ERTC
sat N

Ec

42

5.3.2. Sensor network lifetime in the unsaturated regime In the unsaturated regime we deal with successive cycles of T units of time. During each cycle N reports are transmitted (see Fig. 4). Specically, each time the ith i 0; . . . ; N report is successfully transmitted, the remaining N i nodes, which have not yet transmitted their packets, compete again to access the data channel. Hence, like in (39), the average energy required to report an event under the unsaturated regime with N reporting nodes is given by:

1772

F. Bouabdallah, N. Bouabdallah / Computer Communications 31 (2008) 17631776

ERTC unsat N Etr N Eov N Econt unsat N 8 N P > 1 > E N 1 P c i > cont unsat N > > i1 > > > > N Eidle t1cont i P c i > > > > > Eidle N t00 > 2 i DIFS < > > > > > > > > > > > > > > > : Bc i Slots i P PrfN c njU gnEtx tRTS
n 2

43

44

N nErx tRTS

where Etr N and Eov N are already given in (40), and Econt unsat N can be expressed as shown in (44). Let us now derive the continuous energy consumption Ec under the unsaturated regime. To do so, we assume that 1 R R > T b c R b c N T N , that is the mean unsat M N N 1 time between two successive events occurrence M is higher than the mean time required to report reliably an event R R T bN c R bN cN . So we have, Ec N 1 R M T unsat N Eidle 45 Using (38), the network lifetime in the unsaturated regime can be therefore expressed as follows: T network
lifetime N

Einitial M R ERTC unsat N Ec

46

6. Performance evaluation In this section, we evaluate the impact of the reporting nodes on the WSN performances using both analytical and simulation approaches. The simulations are run on NS-2 simulator [14]. In our simulations, we have not assumed the mobility of the sensor nodes. Therefore, the topology does not continuously vary with time during simulations. However, we note that the sensor nodes may die due to energy depletion leading to variation in overall topology. As stated before, the IEEE 802.11 DCF MAC protocol is used. The parameter settings in our experiments are listed in Table 1. Fig. 5a and b represent the probability of collision in the saturated and unsaturated regimes, respectively, for a varying number of reporting nodes. We can see that this probability increases, in both cases, with the increase of the reporting nodes. Indeed, collisions become more frequent
Table 1 Simulation parameters Communication range Sensing range Packet length IFQ length Transmit power Receive power Idle power Initial network energy 40 m 30 m 30 bytes 65 packets 0.660 W 0.395 W 0.035 W 100 J

Fig. 5. Probability of collision.

when the number of competing access nodes increases, which leads to increasingly extra energy expenditure and increases the average time to report an event (RTC). To alleviate these shortcomings, we have to reduce the number of reporting nodes. We note that Fig. 5 shows a good match between analytical and simulation results, which conrms the accuracy of our models. This also holds for the remaining simulations described in this section. Moreover, Fig. 5a shows that our analytical results match better the simulation results than those given in [20]. Fig. 6 plots the average backo time (i.e., t1cont N ) required by a host to access the medium in order to successfully report an event to the sink node from the rst attempt in the saturated regime. We can observe that this waiting time decreases when the number of reporting nodes N increases. Doing so, the overall time required to report an event (i.e., RTC cycle) may be reduced. According to Figs. 5a and 6, we can see that we have two opposite requirements to minimize the time required

F. Bouabdallah, N. Bouabdallah / Computer Communications 31 (2008) 17631776

1773

Fig. 6. Average backo time for a successful transmission in the saturated regime (i.e. t1cont ).

to report an event in the saturated regime. On one hand, increasing the number of reporting nodes enables a faster access to the medium during each RTC cycle and hence, the average RTC time decreases. On the other hand, rising the number of reporting nodes, increases the probability of collision which amplies the time lost in contention procedure during each RTC cycle. Hence, the optimal RTC is a tradeo between these two opposite requirements. Reconciling these requirements, the minimum RTC time is obtained for N opt latency 8. In fact, Fig. 7a shows that the RTC cycle in the saturated regime is a convex function of N where the minimum is obtained for N opt latency 8. Accordingly, the fastest way to report reliably an event in the saturated regime is to devise a network where the number of active reporting nodes is set equal to N opt latency . Similar behavior is also observed in the unsaturated regime (Fig. 7b) where the minimum RTC time is obtained in this case for N opt latency 9. In other words, considering a particular reporting node frequency f belonging to the unsaturated range, the fastest way to report reliably an event is to let only N opt latency nodes among the N potential ones to report a detected event. In this case, the remaining N N opt latency reporting nodes undergo a sleep mode. So far, we have presented the impact of N on the RTC time for each regime (saturated, unsaturated). In what follows, we are interested rather in understanding the impact of the reporting frequency f on the RTC time. Fig. 8 reports the average RTC time as a function of the network reporting frequency f. In this case, when f exceeds approximately 900 reports=s we deal with the saturated regime, otherwise we get the unsaturated regime. Fig. 8 shows that the RTC time decreases signicantly in the case of the saturated regime. Indeed, the saturated regime enables a faster access to the medium in order to report an event (i.e., min Bi ) compared to the unsaturated regime. According
ih1;N i

Fig. 7. Average time to report an event (i.e. RTC).

to this result, we can see that considering a reporting frequency in the saturated range allows faster event reporting

Fig. 8. Average time to report an event (i.e., RTC).

1774

F. Bouabdallah, N. Bouabdallah / Computer Communications 31 (2008) 17631776

compared with the unsaturated case. Combining, the results of Figs. 7 and 8, we can conclude that the fastest way to report reliably an event (i.e., transmit R reports to the sink) is to plan a network where the number of active reporting nodes is set equal to N opt latency 8, and to adopt the saturated reporting mode as depicted in Fig. 9. As shown in Fig. 9 the saturated regime is more ecient from the latency point of view than the unsaturated regime. Fig. 10a and b show the average amount of energy consumed during each RTC cycle (i.e., to report an event) for a varying number of reporting nodes N in the saturated and unsaturated regimes, respectively. Unlike, the RTC curves (i.e., Fig. 7), these gures show that the amounts of energy ERTC sat N and ERTC unsat N are monotonically rising with N. This monotonous increase is mainly due to two factors. First, increasing N amplies the wasted energy due to collisions. Moreover, increasing N means waking up more sensor nodes within the event radius Rc . Doing so, the total amount of energy consumed by the network in the reception of the signaling messages (RTS, ACK) increases considerably (see the hatched zones of Fig. 10). Finally, we underline that the impact of these two factors is dominant with respect to the slight idle energy reduction when N increases due to the faster access to the medium. According to these results, we can see that the optimal number of reporting nodes enabling the minimal energy consumption is N opt energy 1 for both saturated and unsaturated regimes. Hence, we can state that the fastest way to report an event does not correspond to the optimal manner to consume the network energy. In this regard, the choice of the number of active reporting nodes depends mainly on the specic QoS requirements of the WSN application. In the previous paragraph, we showed the impact of N on the energy consumption. Hereafter, we want to understand the impact of the reporting frequency f on the energy consumption. Fig. 11 plots the analytical results of the

Fig. 10. The amount of energy consumed to report an event.

Fig. T sat

9. Average

time

reliably N ; T unsat reliably N ; f :

needed to report case R 20.

reliably

an

event

amount of energy required to report reliably an event (i.e., transmit R reports to the sink) as function of the network reporting frequency f. We note that the simulation results practically overlap analytical ones, thus we represent only the analytical results. We can observe that the energy reduces with f and becomes constant once the saturated regime is reached. Indeed, increasing f in the unsaturated range, reduces the idle period that experience the reporting nodes each time they nish the transmission of their N reports and before generating the N new ones (see Fig. 4). Hence, increasing f reduces the time required to report reliably an event (Fig. 9) as well as the associated consumed energy. Doing so, we reduce the idle energy spent by the network to report reliably an event. This amount of energy is null when we reach the saturated regime. Combining the results of Figs. 10 and 11, we can conclude that the optimal way to consume energy is to activate only one reporting node that works in the saturated regime.

F. Bouabdallah, N. Bouabdallah / Computer Communications 31 (2008) 17631776

1775

Fig. 11. The amount of energy needed to report reliably an event: case R 20.

Fig. 12a and b plot the network lifetime evolution, in both saturated and unsaturated regimes, respectively, as a function of the number of reporting nodes N for a varying values of the desired event reliability R. In our simulations, we consider the rate of event occurrence M 5. In other words, 5 events occur per unit of time. Fig. 12 shows that the smaller N is, the longer the network lifetime becomes whatsoever the value of R. Thus, the maximal network lifetime is reached when N is equal to 1. Indeed, when N increases the probability of collision rises causing important energy depletion. In contrast, when N is set equal to 1, the probability of collision is null, avoiding thus extra energy expenditure due to collisions. 7. Tradeo between energy and latency Energy-eciency is a critical issue in wireless sensor networks. However, minimizing the energy consumption in such networks must be achieved while respecting the maximum tolerable time to report an event. The optimal solution must therefore take into account these two metrics. In this section, we propose a simple function fchoice to determine the optimal number of reporting nodes N opt that achieves the above-mentioned tradeo. For illustration purposes, we give hereafter the expression of fchoice in the saturated regime. Similar results can be easily obtained in the unsaturated regime. fchoice can be expressed as follows: fchoice N ; a a ERTC sat N T sat N 1 a Esat average T sat average 47
Fig. 12. Sensor network lifetime.

where a is a weight that ranges between 0; 1. It is xed by the network administrator depending on the particular needs of the sensor application (i.e., whether more emphasis is given to energy conservation or to reporting latency minimization), Esat average and T sat average are, respectively, the mean value of ERTC sat N and T sat N for the dierent number of reporting nodes N :

Fig. 13 plots the evolution of fchoice as a function of the number of reporting nodes N : We consider two values of the weight a 1=8 and a 1=4: Two main observations can be identied through Fig. 13. First, fchoice is a convex function of N where the minimum is obtained for N opt . In this regard, N opt minimizes the weighted function fchoice and achieves therefore the desired tradeo between the energy conservation and reporting latency minimization. Moreover, we can observe through Fig. 13 that the value of N opt depends of the weight a. Specically, if more emphasis is given to the energy conservation aspects (i.e., a is set close to 1), the value of N opt will be close to N opt energy (i.e., N opt 1). In contrast, if more priority is given to the latency minimization aspects (i.e., a is set close to 0), the value of N opt will be close to N opt latency (i.e., N opt 8). This result is shown in Fig. 14, where N opt decreases with a from N opt latency to N opt energy . It is worth noting that for the extreme cases where a 0 and a 1 we get, respectively, the same curves as Figs. 7a and 10a.

1776

F. Bouabdallah, N. Bouabdallah / Computer Communications 31 (2008) 17631776

References
[1] I. Akiyldiz, W. Su, Y. Sankarasubramaniam, E. Cayirci, A survey on sensor networks, IEEE Communication Magazine 40 (8) (2002) 102 114. [2] S. Singh, C.S. Raghavendra, PAMAS: power aware multi-access protocol with signaling for ad hoc networks, ACM Computer Communication. Review (1998) 526. [3] F. Dai, J. Wu, Distributed dominant pruning in ad hoc wireless networks, in: Proceedings of IEEE International Conference on Communications, Anchorage, USA, 2003. [4] M. Miller, N. Vaidya, A mac protocol to reduce sensor network energy consumption using a wake-up radio, IEEE Transactions on Mobile Computing 4 (3) (2005) 228242. [5] W. Ye, J. Heidemann, D. Estrin, Medium access control with coordinated adaptive sleeping for wireless sensor networks, IEEE/ ACM Transactions on Networking 12 (3) (2004) 493506. [6] T. van Dam, K. Langendoen, An adaptive energy-ecient MAC protocol for wireless sensor networks, in: Proceeding of ACMSenSys 2003, Los Angeles, CA, 2003, pp. 171180. [7] R.C. Shah, H.M. Rabaey, Energy aware routing for low energy ad hoc sensor networks, IEEE Wireless Communication and Networking Conference, Orlando, FL, USA, 2002. [8] J. Chang, L. Tassiulas, Maximum lifetime routing in wireless sensor networks, IEEE Transactions on Networking 12 (4) (2004) 609619. [9] S. Tilak, N.B Abu-Ghazaleh, W. Heinzelman, Infrastructure tradeos for sensor networks, in: Proceedings of ACM WSNA 2002, Atlanta, GA, USA, 2002, pp. 4958. [10] O.B. Akan, I.F. Akyildiz, Event-to-sink reliable transport for wireless sensor networks, IEEE/ACM Transactions on Networking 13 (5) (2005) 10031016. . B. Akan, I.F. Akyildiz, Spatio-temporal correlation: [11] M.C. Vuran, O theory and applications for wireless sensor networks, Computer Networks Journal (Elsevier) 45 (3) (2004) 245259. [12] M.C. Vuran, I.F. Akyildiz, Spatial correlation-based collaborative medium access control in wireless sensor networks, IEEE Transactions On Networking 14 (2) (2006) 316329. [13] Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specications, ISO/IEC IEEE 802.11 Standard, 1999. [14] The Network Simulator ns-2. [Online]. Available from: <http:// www.isi.edu/nsnam/ns>. [15] H. Nama, M. Chiang, N. Mandayam, Optimal utility-lifetime tradeo in self-regulating wireless sensor networks: a distributed approach, in: Proceedings 40th Annual Conference on Information Sciences and Systems, 2006, pp. 789794. [16] R. Madan, S. Cui, S. Lall, A. Goldsmith, Cross-layer design for lifetime maximization in interference-limited wireless sensor networks, in: Proceedings of IEEE INFOCOM, 2005. [17] C. Intanagonwiwat, R. Govindan, D. Estrin, Directed diusion: a scalable and robust communication paradigm for sensor networks (2000), in: Proceeding of Sixth Annual International Conference on Mobile Computing and Networks (MobiCom 2000), Massachusetts, Boston, 2000. [18] C. Intanagonwiwat, D. Estrin, R. Govindan, J. Heidemann, Impact of network density on data aggregation in wireless sensor networks, in: Proceedings 22nd International Conference on Distributed Computing Systems 2002, pp. 457458, 2002. [19] Y. Chen, Q. Zhao, On the lifetime of wireless Sensor Networks, IEEE Communications Letters 9 (2005). [20] M. Heusse, F. Rousseau, G. Berger-Sabbatel, A. Duda, Performance anomaly of 802.11b, in: Proceedings of IEEE INFOCOM 2003, San Francisco, USA, 2003.

Fig. 13. Tradeo between energy conservation and latency minimization.

Fig. 14. The optimal number of reporting nodes for a varying value of the weight a.

8. Conclusion In this paper, we explored the relationship between the wireless sensor network performance and the number of reporting nodes. To the best of our knowledge, we are the rst to investigate the energy optimization problem from this perspective. Accordingly, we demonstrated that the optimal number of reporting nodes that minimizes the energy expenditure in the sensor network does not correspond to the fastest way to report an event. Based on this result, we propose a simple methodology to achieve this tradeo, which depends on the specic requirements of each WSN application.