Available online at www.sciencedirect.

com

Computer Networks 52 (2008) 2543 www.elsevier.com/locate/comnet

Rate-lifetime tradeo for reliable communication in wireless sensor networks q

a,*

, Farid Nait-Abdesselam

Department of Computer Science and Engineering, The Hong Kong University of Science and Technology, University Road, Clear Water Bay, Kowloon, Hong Kong IRCICA/LIFL, CNRS UMR 8022, Batiment M3, University of Sciences and Technologies of Lille, 59655 Villeneuve dAscq Cedex, France Available online 29 September 2007

Abstract The network lifetime and application performance are two fundamental, yet conicting, design objectives in wireless sensor networks. There is an intrinsic tradeo between network lifetime maximization and application performance maximization, the latter being often correlated to the rate at which the application can send its data reliably in sensor networks. In this paper we study this tradeo by investigating the interactions between the network lifetime maximization problem and the rate allocation problem with a reliable data delivery requirement. Severe bias on the allocated rates of some sensor nodes may exist if only the total throughput of the sensor network is maximized, hence we enforce fairness on source rates of sensor nodes by invoking the network utility maximization (NUM) framework. To guarantee reliable communication, we adopt the hop-by-hop retransmission scheme. We formulate the network lifetime maximization and fair rate allocation both as constrained maximization problems. We characterize the tradeo between them, give the optimality condition, and derive a partially distributed algorithm to solve the problem. Furthermore, we propose an approximation of the tradeo problem using NUM framework, and derive a fully distributed algorithm to solve the problem. 2007 Elsevier B.V. All rights reserved.
Keywords: Wireless sensor networks; Network utility maximization; Rate allocation; Energy eciency; Reliability

1. Introduction Self-organized and -congurable wireless sensor networks have invaluable potential in military and civilian applications where distributed sensing, collection and dissemination of information are

This work is supported in part under grant Hong Kong RGC DAG05/06.EG42. * Corresponding author. E-mail addresses: csjhzhu@cse.ust.hk (J. Zhu), ether@ cse.ust.hk (K.-L. Hung), brahim@cse.ust.hk (B. Bensaou), farid.nait-abdesselam@li.fr (F. Nait-Abdesselam).

required. Examples of such applications include battleeld surveillance, environmental monitoring, home automation [1], and so on. Typically, sensor nodes are battery operated, and hence have to run on a limited energy budget. Further, battery replacement is impossible in many wireless sensor network applications. Thus, energy supply is limited, and sensor networks have a nite operational lifetime. Although substantial improvements have been achieved in the chip design for energy conservation, advances in battery design still lag behind, making energy eciency one of the fundamental challenges in sensor networks.

1389-1286/$ - see front matter 2007 Elsevier B.V. All rights reserved. doi:10.1016/j.comnet.2007.09.011

26

J. Zhu et al. / Computer Networks 52 (2008) 2543

Tremendous research eorts have been spent on the problem of energy conservation in wireless sensor networks. One important perspective is to maximize the network lifetime [2,1113,30], where network lifetime usually refers to the time interval between the initialization of the network and the exhaustion of the battery of the rst sensor node. The lifetime of a sensor node is generally inversely proportional to the average rate of information generated and the information relayed by itself, and hence network lifetime is determined by the source rates of sensor nodes in the network. The application performance is often evaluated by the accuracy and latency of information obtained in wireless sensor networks [3]. The precise semantics of accuracy and latency are application-dependent. For instance, in wireless video sensor networks, the accuracy of information can be quantized by rate distortion analysis [4]. Roughly, the larger the source rate, the smaller the distortion level. In general wireless sensor networks, both metrics depend on the rates of application data reliably delivered from the sensor nodes to the sink nodes. Thus, the application performance can be indirectly measured by the rate allocation in the network. Motivated by this reason, many works have been conducted on the rate allocation problem with energy constraint in wireless sensor networks [5,14,16]. These works try to strike a balance between fully utilizing the network resources and maintaining a certain fairness on source rates among sensor nodes. Since the application performance correlates to the rates of data obtained reliably in sensor networks, we need to consider the data delivery reliability requirement in sensor networks. Many denitions of reliability have been proposed by researchers, e.g., the end-to-end reliability, the event-to-sink reliability proposed by Akan and Akyildiz for event-based sensor networks [6], and so on. Here we focus on the rst one, that is, data from sensor nodes are delivered to the sink nodes without loss or error. Due to the error-prone nature of wireless channels, there are two approaches to improve the reliability: reducing the probability of data loss or error and retransmitting data once loss or error occurs. Here we adopt the retransmission scheme. Under this category, there exist two approaches: the end-to-end retransmission scheme and the hop-by-hop retransmission scheme. In the rst method, the source node will initialize a retransmission when loss or error is detected either

by the source or the destination, while in the second method, the intermediate node will start a retransmission as soon as loss or error is detected. The second method is well known to be more eective than the rst one in wireless networks since losses or errors are recovered locally, and resources used to deliver data to the current node will not be wasted. Thus we use the hop-by-hop retransmission scheme to guarantee the reliable data delivery from sensor nodes to sink nodes. From the above discussion, we notice that there are intrinsic tradeos between network lifetime maximization and rate allocation in wireless sensor networks. Although both problems have been extensively studied in recent years separately, few works consider these two goals together, and study the tradeo between them with a reliability requirement. In this paper, we address this problem with a multi-path routing approach and observe that, only maximizing the throughput of the sensor network, leads sensor nodes that are far away from the sink nodes to suer extremely small data rates, which makes the performance of these sensor nodes extremely bad. Thus, a certain fairness in rate allocation of sensor nodes is required, and should be achieved in a distributed way. This motivates the usage of a network utility maximization (NUM) framework [7] for our rate allocation problem, which has been proved ecient in such a case. In this paper, we rst consider network lifetime of wireless sensor networks as global information shared among all sensor nodes. We nd that the solution maximizing the network lifetime also minimizes the maximum normalized power dissipation of sensor nodes with the latter one being relatively more tractable. Thus, the problem of minimizing the maximum normalized power dissipation of sensor nodes is solved instead. For the sake of simplicity, we still call this problem the network lifetime maximization problem. Given a set of routes, the tradeo between network lifetime maximization and fair rate allocation is formulated as a multiple objective programming problem with a set of convex constraints. By introducing a system parameter c, we can combine these two objectives together as a single weighted objective. The tradeo between them can be characterized by the parameter c. Then we study the optimality condition, and derive a subgradient algorithm. Note that at this stage the information of network lifetime is shared among all sensor nodes, so global exchange of this information cannot be avoided in this algorithm; for example, it

J. Zhu et al. / Computer Networks 52 (2008) 2543

27

can be broadcast by the sink node to all the nodes in the network, however this can impose a large amount of communication overhead. Hence this algorithm is partially distributed, and is not ecient for large scale sensor networks. To overcome this, we draw the parallel between maxmin rate allocation problem in data networks [8] and our problem, and nd a close similarity between the two. Since maxmin rate allocation can be approximated with the NUM framework, here we adopt the same idea, and introduce a new utility for each sensor node as a function of its normalized power dissipation. From this, the studied tradeo problem is formulated in a unied NUM framework. Also, the optimality condition is identied, and an ecient fully distributed algorithm is derived. The impact of the system parameters on the tradeo, and the impact of utility functions on the approximation ratio of the fully distributed algorithm are studied via numerical examples. The remainder of this paper is organized as follows. In Section 2, we review some related work, and in Section 3, we describe the network model, and then study the network lifetime maximization problem and fair rate allocation problem with reliability requirement separately. In Sections 4 and 5, we study the tradeo between these two goals and present our partially distributed and fully distributed algorithms. Section 6 presents some numerical results, and Section 7 concludes this paper. 2. Related work Due to energy constraints in micro sensor nodes, performance limits of wireless sensor networks have been extensively studied. Bhardwaj et al. [9] derive the upper bound of network lifetime for energy-ecient collaborative data gathering with optimal role assignments. Hu et al. [10] give asymptotic analytic results on the relationship between the operational lifetime and number of nodes in the network given a xed node density. Our work diers from the above works since we focus on how to achieve the performance limits, not what the upper bound is. As explained above, two of the most important performance metrics are network lifetime [2,11 13,15,30] and rate allocation of sensor nodes [5,14,16]. In [11], Chang and Tassiulas formulate the problem of maximum lifetime routing as a linear programming problem, and propose a distributed heuristic to approximate the maximum lifetime. In [12], Kalpakis and Namjoshi present heuristics to

maximize system lifetime with data aggregation. In [13], Sankar and Liu propose a distributed ow algorithm with guaranteed approximation error. In [2,30], distributed approximation algorithms that maximize system lifetime are proposed by formulating the problem as a convex nonlinear programming problem. In [14], Srinivasan et al. apply a cooperative game framework to the problem of cooperation among the nodes to relay trac. They propose an algorithm that results in a Nash equilibrium and converges to the optimal throughput given a minimum lifetime requirement. In [15], Kar et al. present a heuristic based on a shortest path algorithm with guarantee of a logarithmic competitive ratio if admission control of messages is allowed. In [16], Hou et al. study the lexicographical maxmin rate allocation among all nodes with a system lifetime requirement. In [5], a fair data collection problem with NUM framework is studied. Srinivasan et al. [5] propose algorithms to split trac among multiple paths given the system lifetime requirement. Our work diers from the above works as we take into consideration both network lifetime maximization and fair rate allocation. Furthermore, these works implicitly assume a contention-free MAC layer and an error-free physical layer, which is hard to achieve in real sensor networks. In our paper, we address the data delivery reliability requirement explicitly. Data delivery reliability in sensor networks has been addressed in many works from dierent perspectives. Most works focus on tackling this problem from the transport layer with functionalities provided by lower layers [6,1720]. (A survey can be found in [21].) It can be achieved either by reducing the probability of loss or error, or retransmissions. In [19], Deb et al. propose two mechanisms based on MAC layer retransmissions to provide reliable data delivery: hop-by-hop reliability (HHR) and hop-by-hop reliability with acknowledgement (HHRA). Both rely on sending multiple copies of the same packet to improve the data delivery success probability. The required number of copies is computed based on a locally estimated packet error rate. In this paper, we use the same approach as HHR and HHRA, but without the retransmission limit. In this way, we can provide end-to-end guaranteed data delivery. A work similar to ours is [22]. Nama et al. study the same tradeo problem. They propose a general cross-layer optimization-based framework that takes into account radio resource allocation,

28

J. Zhu et al. / Computer Networks 52 (2008) 2543

routing and rate allocation. They obtain a distributed algorithm via dual decomposition of the original problem. In our paper, although we have the same objective function, we focus on addressing this problem at the transport layer, and hence have different set of constraints. Also, the derived algorithm is dierent.

Also the data rate on all routes should be non-negative, i.e., y r P 0 8 r 2 R: 2 To ensure data from sensor nodes to be delivered to sink nodes reliably, we use the hop-by-hop retransmission method. At each node, a packet will be retransmitted at the MAC layer until it is successfully received at the next hop node on the route. Assuming the network to be under-loaded, which is common in most sensor networks applications, most packet losses are due to link errors because of interference. Under this assumption, it is reasonable to assume that the frame retransmission probability is independent of the source rates vector y. A model that considers collision errors in a heavily loaded networks and their dependence on the rate vector is considered in [23]. To model this, for each link l, denoting the link error probability by sl, the average number of transmissions until successful 1 delivery of one packet is 1 . sl Since in most types of sensor nodes, communication modules dominate the energy consumption [24], we ignore energy consumed by other tasks such as sensing and data processing. The same simple energy consumption model as in [25] is used for the communication module of all nodes. Assuming all nodes have the power control functionality, the power dissipation at node i for transmission is determined by ptij es ij fij ; 3

3. Problem formulation We consider a wireless sensor network consisting of sensor nodes and sink nodes. Sensor nodes are battery driven, non-rechargeable and irreplaceable. They periodically perform some sensing tasks, collect data about phenomena of interest, and report the measurements to the sink nodes. Data fusion is not considered here, i.e., data will not be aggregated along the routes to the sink nodes. There are multiple sink nodes in the network, and each sensor node is assigned to only one sink node. The performance of the sensor network is characterized by two metrics, the network lifetime and the quality of information received at the sink nodes. The quality of information is roughly proportional to the allocated rate at the source node. The larger the source rate, the more ne grained and accurate the information, and vice versa. We assume that sink nodes accept information of any quality level. Thus, there is no source rate requirement for sensor nodes. The wireless sensor network is modeled as a directed graph GN; L, where N S [ T is the set of all sensor and sink nodes in the network; S denotes the set of sensor nodes; T denotes the set of sink nodes and L is the set of all directed links in the network. We dene Lout(i) to be the set of all outgoing links of node i, and Lin(i) to be the set of all incoming links of node i. In general, in wireless sensor networks, proactive routing is preferred since nodes in the network are static and the trac requirement is also xed. Thus, each sensor node will rst nd a set of routes to deliver its collected data to the sink node, and then determine the data rate on each of its routes. Let R be the set of all routes in the network, and Rs  R be the set of routes for sensor node s. Let yr be the data rate on route r. Denoting the source rate of sensor node s by xs, we have the following ow conservation constraint: xs X
r 2 R s

where ptij is the power dissipated at node i when it is transmitting to node j, fij is the rate of data transmitted at the physical layer, and es ij , the transmission energy consumption cost of link (i, j), is given by
m es ij l g d ij :

Here, l denotes the energy cost of the transmit electronics of node i, g is a coecient term corresponding to the energy cost of transmit ampliers, dij is the distance between node i and j, and m is the path loss exponent, with values ranging between 2 and 4. Typically, l = 50nJ/b, g = 0.0013pJ/b/m4 (for m = 4). The power dissipation at a receiver i can be expressed as X r pr fji ; 5 i e
j6i

yr

8s 2 S:

where e is theP energy consumption cost of the radio receiver, and j6i fji is the rate of data received at the physical layer of node i. Typically, er = 50nJ/b.

J. Zhu et al. / Computer Networks 52 (2008) 2543

29

Since we adopt a link layer retransmission scheme, for a link l, the actual average data rate at the physical layer is the product of the data rate at the network layer and the average number of transmissions at the MAC layer. Let R l be the P set of routes on link l, we have fl r2Rl y r = 1 sl . The total power dissipation at node i is the sum of total power dissipation by radio transmitter and receiver. Thus, the total power dissipation at node i is given by P P X X r 2 R l y r r 2 R l y r t s r pi el e : 6 1 sl 1 sl l2L i l2L i
out in

Assuming that each sensor node i has an initial energy budget Ei, the lifetime of sensor node i is given by Ti Ei : pti 7

3.1. Fair rate allocation problem Since the accuracy of information depends on the allocated source rate, only maximizing the total throughput of the network is not good. It is not sufcient to guarantee the application performance as this objective may be achieved at the expense of some source rates. For example, in a sensor network that tracks objects mobility in a large eld, smaller rates means lesser location tracking accuracy and larger rates equate to a ner grained accuracy. By maximizing the total throughput instead of considering the fairness issue among source rates, one can end up with a solution that shuts o many sources in the network and enables only those sources whose transport energy-cost to the sink is the smallest. Hence some fairness on the rates allocated to the dierent sensor nodes is desired. One good methodology to achieve this goal is to adopt the NUM framework, in which a concave, nondecreasing, and twice dierentiable utility function Us(xs) measures the satisfaction of sensor node s with the assigned rate xs, and the goal is to maximize the sum of the individual utilities. One class of utility functions that has been extensively used to achieve fair resource allocation [8] in economics and distributed computing is U a x  log x;
1 x1a ; 1a

When a = 1, the utility function leads to the socalled proportional fairness, whereas when a ! 1 the utility function leads to maxmin fairness. What class of fairness is required for rate allocation in a sensor network is determined by the application and is beyond the scope of this paper. While maxmin fairness maximizes the smallest rate in the network, allowing thus for the best possible accuracy in monitoring the whole network, proportional fairness may favor nodes that are nearer the sinks over those that are located further away. One can imagine many application scenarios, both military and civilian, that match these two requirements. Typically, the average data rate generated by each sensor node is small, so it is reasonable to assume that there is sucient bandwidth for each link in the network. That is, resource constraints below the network layer are not present. Under this assumption, the fair rate allocation problem given network lifetime requirements can be formulated as a NUM problem X ws U s xs maximize
s2 S

subject to T i P T net 8i 2 S; constraints 1; 2; 6; 7;

where ws is a weight associated with the utility function, and Tnet is the lifetime requirement of the sensor network. With the help of parameters w, further dierentiation among sensor nodes can be achieved. For example, sensor nodes deployed in the area that attracts most interest can get fairly large allocated rates by setting their weights relatively large. In this way, we can achieve weighted fairness on source rates of sensor nodes. 3.2. Network lifetime maximization problem All sensor nodes are assumed to be of equal importance, which is a reasonable assumption since the death of one sensor node may cause the network to be partitioned, or some area requiring monitoring to be uncovered. Thus the lifetime of a sensor network is dened as the time until the death of the rst sensor node, i.e., T net mini2S T i . Given source rate demands of all sensor nodes, the problem of maximizing the lifetime of a sensor network can be stated as maximize min
i2S

Ti 8s 2 S; 10

a 1; a > 1:

subject to xs x0 s

constraints 1; 2; 6; 7;

30

J. Zhu et al. / Computer Networks 52 (2008) 2543

where x0 s is the source rate requirement of sensor node s. This formulation is very similar to that of [11], and the heuristic in [11] can be used. Generally, constraints (7) are not convex, and hence the optimization problem (10) is not convex. This usually makes the problem dicult to solve. To overcome this, we introduce a new variable zi = 1/Ti, which is the inverse of the lifetime of sensor node i. Note that from (7), it is also the normalized power dissipation of node i with respect to initial energy Ei. It is easy to verify that maximizing the minimum lifetime of all sensor nodes is equivalent to minimizing the maximum normalized power dissipation of all sensor nodes. Hence the original problem (10) is reformulated as maximize max zi
i2S

ws U s xs 1 c max zi i2S P P X X y r 2 R l r r 2 R l y r s r subject to el e 1 sl 1 sl l 2 L i l2L i maximize c

Ei zi 8i 2 S; X y r xs 8s 2 S;
r 2 R s

yr P 0

8r 2 R:

subject to pti Ei zi xs x0 s

8i 2 S

11

8 s 2 S;

constraints 1; 2; 6: Notice that problem (11) is a linear programming problem, which is much easier to solve than the original problem in (10). For the sake of simplicity, we will still call this problem the network lifetime maximization problem. 4. Tradeo between network lifetime and fair rate allocation From problems (9) and (11), we observe that we have two important but conicting objectives when optimizing the sensor network performance, i.e., achieving fair rate allocation among sensor nodes (problem (9)) and maximizing the network lifetime (problem (11)). Both can be formulated as constrained maximization problems. Hence, the tradeo between them can be formulated as a multi-objective programming problem, and a simple and ecient way to achieve desired tradeo between them is the weighting method [26]. That is, we introduce a new system parameter c 2 [0, 1], and combine these two objective functions into a single objective function. Since c characterizes the tradeo, we call it tradeo factor. With the constraint that data rates on all routes should be non-negative, the tradeo between fair rate allocation and network lifetime maximization is formulated as a constrained optimization problem

12 The above weighted objective function is a concave function since U() is concave and the maximum of a set of linear functions is convex, and both x and z are linear functions of y. The constraint sets are also convex, thus it is a convex optimization problem, and there is only a unique optimal objective value, i.e., a locally optimal solution is also a globally optimal solution [27]. Denote the column vector of {yr}, {xi}, and {zi} by y, x, and z respectively. Note that z and x depend on y, and can be represented by functions of y. Therefore, x and z are dummy variables in (12), and the only optimization variables are y. Dene the weighted objective function as Qy P c s2S ws U s xs 1 cmaxi2S zi . We have the following optimality condition. Proposition 1. Let znet maxi2S fzi g, J(z) = {ijzi = znet}, and Ri be the set of routes on node i. then (a) a solution {x*, y*, z*} is globally optimal if and only if for all d 2 Rn, there holds ( X min cr y w s U s x s
s2 S

!0 dj i 2 J z

) 6 0; 13

1 cry z i

(b) if a solution {x*, y*, z*} is optimal, then for all i 2 J(z*), there holds X cr y r w s U s x s 1 cry r zi 6 0 8r 2 Ri :
s2 S

14 Proof. The proof of the proposition is given in Appendix A. Part (a) of Proposition 1 gives a necessary and sufcient condition for a solution to be optimal, however, it does not give us any insight on the structure of the optimal solution. From part (b), we nd that at the optimal point, all the routes on the sensor nodes with the minimum lifetime would not

J. Zhu et al. / Computer Networks 52 (2008) 2543

31

receive positive prot if considering the value of function (14) as its prot. Thus, a class of algorithms can be developed in which each route updates its data rate based on the prot received and certain update rule. Notice here that the objective function Q(y) is non-differentiable, hence the gradient projection algorithm cannot be applied, and a subgradient algorithm is developed instead [28]. Assume we know J(z), the set of sensor nodes with the maximum normalized power dissipation given the current rate allocation y. For each i 2 J(z), dene X Qi y c ws U s xs 1 czi :
s2 S

gradient. Assume that there is only one senor node with maximum normalized power dissipation; we nd the objective function Q(y) is indeed dierentiable, and problem formulation (12) is very similar to that of [7]. Thus we use the same method to update the data rates of routes. At each iteration t, the data rate of a route r is changed by Dy r t j y r t g t cU 0s xs t y r 1 c ry zp t: cU 0s xs t r 17

jy r tws

Then the gradient of Qi(y) with respect to data rate yr is given by ry r Qi y cws U 0s xs 1 cry r zi ; 15

where xs is the source rate of sensor node s that owns route r. For each route r 2 R, we compute ry r Qi y for all i 2 J(z), and choose the largest one as the direction to update data rate yr, i.e., gy r cws U 0s xs 1 cry r zp ; where p arg max ry r Qi y:
i2J z

16

Lemma 2. g(y) in (16) is a subgradient of Q(y) at point y. Proof. For any other point y1, the following inequalities hold by the denition of Q(y) and Qi(y) : Qi(y) = Q(y) "i 2 J(z), and Qi y1 6 Qy1 8i 2 S. Therefore, for any i 2 J(z), there holds: Q(y) Q(y1) 6 Qi(y) Qi(y1). Since function U() is concave, it can be veried that function Qi(y) is concave. That is, Qi(y) Qi(y1) 6 $yQi(y)T(y y1). Thus we have Qy Qy1 6 ry Qi y y y1 8i 2 J z: By the denition of g(y), we have Qy Qy1 6 gy y y1 ; which by the denition of a subgradient proves the lemma. h Note that when we approach the optimal solution y*, the subgradient g(y) may get very small, and hence the rate of convergence becomes very small. To solve this issue, we need to scale the subT T

We observe that when data rate yr becomes zero, it will never increase in the future. Thus, the algorithm may not converge. To avoid this, data rate yr should be bounded away from this absorbing state. Therefore, we introduce a new assumption: Assumption A1: Each route r 2 R has a small minimum data rate demand d. The details of the algorithm with data rate update formula (17) is described in Algorithm 1. Subgradient may not be a direction of ascent at a point y, but it makes an angle less than 90 degrees with all ascent directions at y. Thus the new point may not improve the objective value, but it gets close to the optimal value. There exist many results on the convergence property of subgradient algorithm [28]. Generally, for constant step size, the algorithm can converge to within a neighborhood of the optimal value, and for diminishing step size satisfying certain rules, the algorithm is guaranteed to converge to the optimal value. With constant step size, the algorithm can be implemented asynchronously and still guarantee to converge to an optimal solution. This reduces the implementation complexity in sensor networks, so we adopt constant step size in Algorithm 1. Noting that the optimal solution y* is not necessarily unique, we have the following convergence results. Proposition 3. Let Y* be the set of optimal solutions of problem (12). If constant step size j is positive and sufciently small, then Algorithm 1 can converge to within a small neighborhood of an optimal solution y* 2 Y *. Algorithm 1. Partially distributed algorithm to solve (12) At each iteration t Step 1: Identify the set of sensor nodes with the maximum normalized power dissipation.

32

J. Zhu et al. / Computer Networks 52 (2008) 2543

1. Sensor nodes compute their normalized power dissipation, respectively. 2. Sensor nodes exchange this information with other sensor nodes to determine the set of sensor nodes with the maximum normalized power dissipation. Denote this set by D. Step 2: Compute the costs of all routes. for each route r 2 R do Routes a message from source to destination, records the set of prices ry r zi ; i 2 D received, and at last deliver this information back to the source node s. endfor Step 3: Update the rate of all routes. for each sensor node s 2 S do for each route r 2 Rs do 1. Compute Dyr(t) with formula (16) and (17). 2. y(t + 1) = max(d,y(t) + Dyr(t)). end for end for The proof of convergence can be obtained by following the method in [29]. Algorithm 1 can be interpreted by the law of demand and supply. In (16), the rst item on the right-hand side represents the marginal increase of rate utility due to the increase of rate on route r, and the second term on the righthand side represents the marginal increase of the normalized power dissipation of sensor node p due to the increase of rate on route r. This can be interpreted as follows: each sensor node s pays for its route r according to its marginal rate utility cws U 0s xs , and at the same time, each sensor node i with maximum normalized power dissipation penalizes any route through it by price 1 c ry r zi . At each iteration, a route r delivers a message from sensor node s to its sink node, and records all the received prices, and then the sink node forward this message back to the sensor node s. After that, the sensor node s decides the new data rate of route r based on the revenue and prices that route r receives, using formulas (16) and (17). Comments: Algorithm 1 guarantees convergence to within a small range of the optimal value given appropriate constant step size. However, it is not fully distributed, as it requires knowing the sensor nodes with the maximum normalized power dissipation at each iteration. In sensor networks comprising of hundreds of thousands of sensor nodes, this procedure will create a large amount of message exchange

and slow down the time to converge to the optimal solution, and hence is not ecient. Such communication overhead can be saved in a fully distributed algorithm. Thus, in the next section, we develop a fully distributed algorithm for the tradeo problem. 5. Fully distributed algorithm Considering utility function Ua() dened in (8), it has been shown in [8] that when a ! 1, maximizing the aggregate utility leads to a maxmin fair rate allocation. We can adopt this idea to maximize the minimum lifetime of sensor nodes in distributed way. From this perspective, we simply regard the lifetime of a sensor node as a prot drawn by the node, and allocate lifetime to dierent sensor nodes in such a manner that the minimal node prot is maximized by invoking the NUM framework subject to the energy consumption constraints and the trac demands of sensor nodes, and using multi-path routing. For this purpose, a new utility of a sensor node i is introduced as a function of its lifetime Ua(Ti). The goal of maximizing the minimal lifetime is approximated by maximizing the aggregate utilities of sensor nodes with a suciently large parameter a. As indicated earlier, maximizing the minimum lifetime of the sensor nodes is equivalent to minimizing their maximum normalized power dissipation; the latter being easier to solve with the NUM framework. Dene a function V b i as Vb i zi 1 z b 1 b 1 i 18

and let V b i zi denote the utility of sensor node i with respect to its normalized power dissipation zi. For a set of predened routes R and a set of source rate demands x of sensor nodes, the NUM formulation of the network lifetime maximization problem is as follows: X b maximize V i zi
i2S

subject to

X
l2Lout i

es l

r 2 R l y r

1 sl

X
l2Lin i

P e
r

r 2 R l y r

1 sl

Ei zi 8i 2 S; X y r xs 8s 2 S;
r 2 R s

yr P 0

8r 2 R: 19

J. Zhu et al. / Computer Networks 52 (2008) 2543

33

Problem (19) is very similar to that of [30], and has the following property: Lemma 4. When b ! 1, the network lifetime determined by the optimal solution y* of problem (19) approximates the maximum network lifetime of the wireless sensor network. Proof. The proof of the lemma is given in Appendix B. h Thus, we call b the lifetime approximation exponent. The tradeo between network lifetime maximization and fair rate allocation of sensor nodes can be unied in the network utility maximization framework as X maximize cws U s xs 1 cV b s zs
s2 S

that the optimal solution x* and z* are unique. We can use the gradient projection method [28] to solve (20). Here we adopt a similar algorithm to the one used in [31]. At each iteration t, the data rate yr is updated by Dy r t j y r t ry P yt cU 0s xs t r

! 1 c X b z try r zi t : jy r t ws 0 cU s xs t i2r\S i 23 From Eq. (23), we observe that the update rule also suers from the same problem as previously viz. any route r cannot increase its rate once it reaches zero. Thus, we need to bound its data rate yr away from zero. With Assumption A1, which sets a small minimum rate requirement d for each route, this problem can be solved. Hence, at each iteration t, the data rate yr is updated by y r t 1 maxd; y r t Dy r t: 24

subject to

X
l2Lout i

P es l

r 2 R l y r

1 sl

X
l2Lin i

P e
r

r 2 R l y r

1 sl

Ei zi 8i 2 S; X y r xs ; 8s 2 S;
r 2 R s

yr P 0

8r 2 R: 20

Here, since both x and z are dummy variables, we can dene P the weighted objective function as P y c s2S ws U s xs 1 cV b s zs . Note that both U() and V() are twice dierentiable, and hence the objective function P(y) is dierentiable. The gradient with respect to data rate yr exists and is given by X b zi r y r zi ; 21 ry r P y cws U 0s xs 1 c
i2r\S

where s is the source sensor node of route r. A feasible solution y* is locally optimal if and only if for all r 2 R, y r P 0; r y r P y 6 0; y r r y r P y 0:

The detailed gradient algorithm is given in Algorithm 2. We observe that this algorithm is a variation of the scaled gradient projection algorithm. To demonstrate its convergence property, we rst introduce an additional assumption:Assumption A2: The data rate yr on each route r is upperbounded by a constant cr. This assumption is reasonable in wireless sensor networks since the bandwidth at the MAC layer is nite, and the data rate of routes cannot exceed the bandwidth. With Assumptions A1 and A2, we nd that the source rates x are also bounded, and the gradient $yP(y) is Lipschitz continuous, which guarantees the convergence of the gradient projection algorithm with constant step size. Dene R jRj, / maxs2S ws U 0s xs , and w maxr2R y r = U 0s xs . We have the following convergence result: Proposition 5. Let Y* be the set of optimal solutions of problem (20). With Assumptions A1 and A2, if constant step size j satises 0 < j < 2=R1=2 /w, then for any initial feasible solution y0, Algorithm 2 converges to an optimal solution y* 2 Y*. Proof. The proof of the proposition is given in Appendix C. h Algorithm 2. Fully distributed algorithm to solve (20)

22

Since both U() and V() are strictly concave, and both x and z are linear functions of y, the objective function P(y) is also concave. Hence a locally optimal solution is also a globally optimal solution. Generally, the optimal solution y* is not necessarily unique. However, the strict concavity of functions U() and the strict convexity of functions V() imply

34

J. Zhu et al. / Computer Networks 52 (2008) 2543

At each iteration t Step 1: Each sensor node i computes its price 1 c z b i ry r zi , and its revenue for all of its routes cws U 0s xs . Step 2: Compute the costs of all routes. for each route r 2 R do Routes a message from source to destination, aggregate the prices 1 czb i ry r zi charged by sensor nodes on it, and nally deliver this information back to the source node s. end for Step 3: Update the rate of all routes. for each sensor node s 2 S do for each route r 2 Rs do 1. Compute Dyr(t) with formula (23). 2. y(t + 1) = max(d, y(t) + Dyr(t)). end for end for In Algorithm 2, a sensor node i gives its route r an amount of revenue based on its marginal rate utility cws U 0s xs , and charges any route on it a price 1 czb i ry r zi , which is a product of the marginal power dissipation utility and normalized power dissipation for this route. A route r delivers a message to the sink node, and aggregates all the prices it receives from the relaying nodes. Then the sink node forwards this message back to the source node i. After receiving this message, sensor node i will add its price for route r to the aggregate price from relaying nodes, and then computes the new data rate yr based on the revenue and aggregate price of route r. When achieving the optimal solution, the revenue and price of each route with positive data rate should be equal. By iteratively updating the data rates of the routes with a constant step size, one optimal solution can be guaranteed, and this can be done asynchronously.

300 250 200 Ym 150 100 50

S11 S1

S8 S14 S13 S6 S19 S20 S15 S18 S9 S4

S17 S10

S16 S7

0 0

50

100

Xm

150

200

250

Fig. 1. Wireless sensor network topology.

sink node was randomly generated at the lower left corner of this square area. We assume that each sensor node has a maximum transmission range of 80 m, and the power dissipation for transmitting to and receiving from a neighboring node is determined by (3) and (5) in Section 3, respectively, where l = 50nJ/b, g = 0.0013pJ/b/m4, m = 4, and er = 50nJ/b [25]. The initial energy budget of each sensor node is 5 kJ. The utility function U() used here is logarithmic, and the weight ws of each sensor node s is set to 1. The initial data rates and minimum data rate requirement  are set to 1 bps and 0.01 bps, respectively. First we consider the scenario where the link error probabilities of all links are equal. We set these probabilities to 0.01. To gain a better understanding of the optimal performance of our rate allocation algorithms, we determine all the possible routes in the network for each sensor node. 6.1. Convergence results and tradeo by Algorithm 1 Fig. 2 depicts the convergence property of Algorithm 1 described in Section 4 with a step size j = 0.05. Here the tradeo factor c is set to 1.0 108. We only show the convergence results of sources rates; the convergence results of other variables are omitted due to space limitations. Here, we only show part of the sensor nodes as we nd that sensor nodes 9, 15 and 18 have the same allocated rates as sensor node 4, and the remaining sensor nodes (not shown here) have the same allocated rates as sensor node 1. We can see that source rate allocation of sensor

6. Performance evaluation In this section, we use some numerical results to evaluate our rate allocation algorithms. We consider the wireless sensor network illustrated in Fig. 1, where the topology is simple yet suciently large for us to investigate the performance of our algorithms. In this network, we have twenty sensor nodes indexed 120, and one sink node. The locations of sensor nodes were randomly generated over a 300 m 300 m square area, and the location of the

J. Zhu et al. / Computer Networks 52 (2008) 2543

35

4 3.5 Source Rate (kbps) 3 2.5 2 1.5 1 0.5 0 0 100 200 300 Iteration Number 400 500
Sensor 1 Sensor 2 Sensor 3 Sensor 4 Sensor 5 Sensor 12 Sensor 16

Fig. 2. The evolution of source rates of sensor nodes with step size j = 0.05 in Algorithm 1.

Fig. 3. The evolution of source rates of sensor nodes with step size j = 0.1 in Algorithm 1.

nodes approaches the optimal solution relatively fast, but oscillates near the optimal solution nally, which is consistent with what we found when developing this algorithm. The selection of the constant step size j is very important, since it controls the magnitude of the oscillations near the optimal solution, and the rate of convergence. Hence it can characterize the performance of our algorithm. Generally, the larger the step size j is, the faster we can approach the optimal solution, but the larger the oscillation is, and vice versa. This can be observed by comparing the convergence results of this algorithm with step size j set to 0.05 and 0.1, which are shown in Figs. 2 and 3 respectively. In Fig. 4, the dashed line denotes the optimal value of the objective function. From this gure, we see that the smaller the step size is, the closer the nal solution approaches the optimal solution. We can see the impact of tradeo factor c on the tradeo between network lifetime and fair rate allocation of sensor nodes with the partially distributed Algorithm 1 in Fig. 5, where c ranges from 1.0 109 to 1.0 107. The lifetime is shown in seconds, and the aggregate rate utility is the sum of the individual rate utility of all sensor nodes. To obtain the optimal solution, we adopt a diminishing step size rule in the algorithm instead of the constant step size rule. We use the simple step size rule j(t) = j(0)/t, and the convergence analysis can be found in [29]. The inherent tradeo between network lifetime maximization and fair rate allocation can be observed from this gure. Each point on the

13
12 Objective Function Value 11 10 9 8 7 6

x 10

optimal function value

Algorithm 1 with = 0.05 Algorithm 1 with = 0.1

Fig. 4. The evolution of objective value with step size j = 0.05 and j = 0.1 in Algorithm 1.

curve denotes a best rate allocation (evaluated by the aggregate rate utility) that can be obtained under the corresponding network lifetime. From this curve, we can observe that the larger the network lifetime is, the smaller the best rate allocation we can get. In our case, we use the logarithmic function as the utility function of a sensor nodes source rate. We observe that in general, when the source rate of each sensor node exceeds 1 bps, the network lifetime is asymptotically inversely proportional to a scale of the tradeo factor c. This observation can simplify the selection of c when we deploy wireless sensor networks, and want to guarantee a desired tradeo between network lifetime and fair rate allocation in the presence of other objectives.

36
200 180 Aggregate Rate Utility 160 140 120 100 80 60 0

J. Zhu et al. / Computer Networks 52 (2008) 2543

= 1.0 107 = 1.0 10

1.4

Algorithm 1 Algorithm 2 with = 10

Sensor 1

1.2 1 0.8 0.6 0.4 0.2

Sensor 2 Sensor 3 Sensor 4 Sensor 5 Sensor 12 Sensor 16

= 1.0 10

= 1.0 10

90

5 10 Network Lifetime (s)

15 x 107

0 0

200

400 Iteration Number

600

800

Fig. 5. The impact of tradeo factor c on the tradeo between network lifetime and fair rate allocation of sensor nodes.

Fig. 7. The evolution of source rates of sensor nodes with step size j = 0.2 in Algorithm 2.

6.2. Convergence results and approximation ratio of Algorithm 2 Figs. 6 and 7 show the convergence results of the fully distributed Algorithm 2 derived in Section 5 with step size j = 0.1 and j = 0.2, respectively. Here we set the tradeo factor c to 1.0 1080, and the lifetime approximation exponent b to 10. We see that if the step size j is selected according to the step size rule proposed in Proposition 5, then the larger the step size is, the faster the algorithm converges. We also observe that the fully distributed Algorithm 2 converges to an optimal solution of problem (20) in Section 5 in a few hundred iterations, and faster than the partially distributed Algo-

1.4
Sensor 1

1.2 Source Rate (kbps) 1 0.8 0.6 0.4 0.2 0 0 200 400 Iteration Number

Sensor 2 Sensor 3 Sensor 4 Sensor 5 Sensor 12 Sensor 16

600

800

Fig. 6. The evolution of source rates of sensor nodes with step size j = 0.1 in Algorithm 2.

rithm 1. Thus, Algorithm 2 has a better convergence property than Algorithm 1. However, Algorithm 2 may not achieve the optimal tradeo since it is only an approximation algorithm. Thus it is interesting to investigate the impact of the lifetime approximation exponent b on the performance of Algorithm 2. If the performance is satisfactory given a reasonably large b, then Algorithm 2 is preferred as it requires less communication overhead, and is easier to implement than Algorithm 1. We evaluate the performance by the approximation ratio of the optimal solution obtained by Algorithm 2 with respect to the optimal solution of problem (12). Comparing the two tradeo curves in Fig. 5, which are obtained by Algorithms 1 and 2, respectively, we observe that the same value of c can result in dierent optimal tradeos between network lifetime and fair rate allocation of sensor nodes. Thus the naive way of xing parameter c, and comparing the optimal solution obtained by Algorithm 2 under dierent value of b with the optimal solution obtained by Algorithm 1 does not work. Hence we use another approach. First we x the tradeo factor c, and obtain the optimal network lifetime and rate allocation of problem (12), then we set b for problem (20), and vary c such that the optimal aggregate rate utility is the same as that of problem (12), or the dierence is extremely small. After that, we normalize the optimal network lifetime of problem (20) with respect to the optimal network lifetime of problem (12), and regard this ratio as the approximation ratio. Here, we set c = 1.0 108 for problem (12), and the results are shown in Fig. 8.

J. Zhu et al. / Computer Networks 52 (2008) 2543

37

From this gure, we can see that when b = 2, we already guarantee a fairly good approximation ratio, i.e., 90.7%, and when b P 14, the approximation ratio exceeds 99%. Thus, the performance of Algorithm 2 is very good. Another intuitive example can be obtained from Fig. 5. Although the ranges of c are dierent, the two tradeo curves obtained by Algorithms 1 and 2 with b = 10 respectively are very close to each other. We compare the lifetime and source rates of sensor nodes obtained by Algorithms 1 and 2 under different value of b with the same aggregate rate utility in Table 1. From Table 1, we observe that when the tradeo factor c is set to 1.0 108 in Algorithm 1, the lifetime of sensor nodes 2, 5, 12 and 16 is very close to the network lifetime. This means that these sensor nodes are bottleneck nodes in the sensor network (the slight dierence between their individual lifetime is due to the precision of the implementation). With Algorithm 2, we nd that the larger the lifetime approximation exponent b is, the closer the network lifetime gets to the optimal network lifetime obtained by Algorithm 1, and the larger the number of bottleneck nodes in the network is. This implies when b grows large, Algorithm 2 can balance the energy consumption rate of sensor nodes more eciently. Also, from the table, we see that although the aggregate rate utilities are the same,

0.99 0.9 Approximation Ratio

0.8

0.7

0.6

0.5 0 2 4 6 8 10 12 14 16 18 20

Fig. 8. The impact of lifetime approximation exponent b on the performance of Algorithm 2.

the sources rates allocated to sensor nodes are not the same. We observe that the larger b is, the closer the rate allocation obtained by Algorithm 2 approaches to the rate allocation obtained by Algorithm 1. 6.3. The impact of routing algorithm In our formulation of the problem and in the previous examples, we assumed that a sensor node determines all its routes to the sink prior to starting

Table 1 Comparison of the individual lifetime and rate of sensor nodes obtained by Algorithms 1 and 2 with xed aggregate rate utility Sensor ID Algorithm 1 (c = 1.0 108) Lifetime 1.0 10 (s) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 0.6340 0.5021 4.3908 1.9810 0.5034 2.7067 0.6387 0.9393 1.0925 1.1645 1.1893 0.5033 2.0641 0.9595 2.4748 0.5047 0.5463 0.5386 1.7395 2.7520
7

Algorithm 2 (b = 1) Lifetime 1.0 10 (s) 0.8026 0.5592 3.7722 4.8375 0.5308 7.6066 1.1499 1.8214 1.5025 1.5345 2.1460 0.4032 3.6006 1.4780 6.1513 0.4516 0.6808 0.8301 3.7552 7.6826
7

Algorithm 2 (b = 10) Rate (kbps) 0.7313 1.4035 1.6937 1.1710 5.8671 0.5578 2.5703 0.6409 1.2970 1.0936 0.6636 2.7400 0.5742 0.6038 1.1927 3.1827 0.9184 1.6227 0.5675 0.5580 Lifetime 1.0 107 (s) 0.6433 0.5059 4.6587 2.1207 0.4994 2.7449 0.7280 0.9652 1.1696 0.7893 1.2083 0.5033 2.0932 0.9779 2.6492 0.5113 0.6561 0.6545 1.7641 2.7908 Rate (kbps) 0.7685 1.0189 1.3861 1.5495 3.1974 0.7552 1.6934 0.7555 1.5496 0.8175 0.7555 2.1252 0.7551 0.7553 1.5495 2.5817 0.8017 1.5946 0.7551 0.7551

Rate (kbps) 0.7705 0.9523 1.3115 1.6586 3.1013 0.7663 1.6599 0.7670 1.6599 0.7705 0.7670 2.0108 0.7652 0.7735 1.6586 2.5049 0.7705 1.6599 0.7652 0.7663

38

J. Zhu et al. / Computer Networks 52 (2008) 2543

data exchange. This in turn leads to the optimal multi-path routing (OMPR) since we can prove that our solutions are optimal. However, in real sensor networks, it may be impossible to explore all possible routes, and only a fairly small set of routes will be used by the sensor nodes. Thus, the impact of route choices on the performance of our algorithms is considered here. Obviously, convergence of our algorithms is still guaranteed since they do not depend on the route choices. Nevertheless, dierent route choices other than those assumed in the model may lead to sub-optimal results. To investigate this, we compare the results of the optimal algorithm with multi-path routing (including all possible routes) to the results obtained by two other routing algorithms, namely, single path minimum energy routing (MER) and 2-shortest path (link disjoint) routing (2SPR). Table 2 compares the performance of the three route choices with equal aggregate rate utility. From the table, we observe that although the three route choices obtain the same rate utility, the rate allocation of sensor nodes with minimum energy routing is quite dierent from that of the optimal multi-path routing and is unfair. In contrast, the rate allocations of sensor nodes with link-disjoint 2-shortest path routing and optimal multi-path routing are very close to each other. We observe that although the network lifetime with minimum energy routing
Table 2 Impact of dierent route choices with equal aggregate rate utility Sensor ID MER Lifetime 1.0 10 (s) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 0.7169 0.6624 9.8441 3.2726 0.4009 11.5525 1.6624 0.8493 1.0106 12.1122 13.1314 0.3902 2.7680 1.5667 5.0843 0.3931 0.5539 0.6523 8.2918 12.2309
7

is already close to the optimal network lifetime, i.e., around 80%, the lifetime of sensor nodes is not well balanced. The network lifetime with the link-disjoint 2-shortest path routing is even closer to the optimal network lifetime, and the power dissipations of sensor nodes are close to the optimal power dissipation with all the routes. From Table 2, we nd that the performance of minimum energy routing is not satisfactory as it can result in a much less fair rate allocation compared with the optimal rate allocation, and wastes more scarce energy of sensor nodes. The performance of the link-disjoint 2-shortest path routing is comparable with the optimal multi-path routing as it can achieve a rate allocation close to the optimal rate allocation, and at the same time, a network lifetime close to the optimal network lifetime. This suggests that for real implementation of our algorithms, it is sucient to only enumerate and explore a small number of disjoint routes from each sensor to the sink to obtain results that are nearly optimal. 6.4. Non-uniform link error probability scenario In the above, we assume the link error probabilities of all links are equal. This is not true in real networks. Usually, the link error probability of a particular link is related to the received power at the receiver, the level of interference in the sur-

2-SPR Rate (kbps) 0.6787 0.6787 0.6787 1.7933 10.7050 0.6787 5.7349 0.6787 1.7933 0.6787 0.6787 1.0405 0.6787 0.6787 1.7933 2.7843 0.6787 1.7933 0.6787 0.6787 Lifetime 1.0 10 (s) 0.6444 0.4769 5.8640 5.4133 0.4608 10.3838 1.0253 0.7634 1.6717 3.2968 11.8029 0.4585 2.4879 1.4082 8.4101 0.4632 0.4596 0.7644 7.4529 10.9935
7

OMPR Rate (kbps) 0.7551 1.0841 1.0648 1.0841 5.8547 0.7551 3.1381 0.7551 1.0841 0.8091 0.7551 1.5564 0.7551 0.7551 1.0841 1.6269 0.9570 3.1381 0.7551 0.7551 Lifetime 1.0 107 (s) 0.6340 0.5021 4.3908 1.9810 0.5034 2.7067 0.6387 0.9393 1.0925 1.1645 1.1893 0.5033 2.0641 0.9595 2.4748 0.5047 0.5463 0.5386 1.7395 2.7520 Rate (kbps) 0.7705 0.9523 1.3115 1.6586 3.1013 0.7663 1.6599 0.7670 1.6599 0.7705 0.7670 2.0108 0.7652 0.7735 1.6586 2.5049 0.7705 1.6599 0.7652 0.7663

J. Zhu et al. / Computer Networks 52 (2008) 2543

39

rounding environment, and so on. As these factors cannot be the same for all links, the error probabilities of links are dierent. Thus, we investigate here the tradeo between network lifetime and rate allocation of sensor nodes under the non-uniform link error probabilities scenario. First we consider an extreme case, where the error probabilities of all links except link (13, 1) are same, and set to 0.01. The error probability of link (13, 1) is set to 0.8. Table 3 shows the comparison of the lifetime and source rates of sensor nodes under dierent error distributions with tradeo factor c = 1.0 108. We can observe that the rate allocation obtained with non-uniform link error probabilities diers greatly from the one with no link errors. From the table, we observe that trac through the less reliable links is reduced by either redirection to other links, or by decrease in the source rates. Hence the lifetime of sensor node 12 does not change too much, although the aggregate rate utility decreases. From this table, we also observe that the rate allocated to sensor node 12 in the extreme case is reduced signicantly compared to the rate allocated to it in the case where no link errors occur. This implies that when we provide energy ecient and reliable communication in sensor networks by link layer retransmission schemes, we should consider
Table 3 Impact of error rate distribution on optimal result c = 1.0 108 Sensor ID No error Lifetime 1.0 107 (s) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 0.6379 0.5054 5.0520 1.9788 0.5068 2.7201 0.6518 0.9436 1.0913 1.1784 1.1946 0.5071 2.0743 0.9646 2.4719 0.5047 0.5519 0.5210 1.7481 2.7655 Rate (kbps) 0.7554 0.9787 1.2834 1.6773 3.1265 0.7674 1.6779 0.7795 1.6778 0.7554 0.7795 1.9677 0.7725 0.7639 1.6773 2.5332 0.7554 1.6781 0.7725 0.7674 With error Lifetime 1.0 107 (s) 0.6890 0.5033 6.2568 2.9587 0.5000 2.9206 0.7006 1.0200 1.6320 1.2676 1.2917 0.5242 2.2272 1.0396 3.6961 0.5028 0.5944 0.5621 1.8770 2.9695 Rate (kbps) 0.7044 0.8042 0.8736 1.1106 2.0861 0.7120 1.1101 0.7025 1.1101 0.7044 0.7025 0.9344 0.7071 0.7010 1.1106 1.6473 0.7044 1.1101 0.7071 0.7120

200
Uniform Link Error Probability

180 Aggregate Rate Utility 160 140 120 100 80

Nonuniform Link Error Probability

= 1.0 107

= 1.0 10

2 3 Network Lifetime (s)

5 x 107

Fig. 9. The impact of tradeo factor c on the tradeo between network lifetime and rate allocation of sensor nodes.

the link error probability in the energy consumption model. Otherwise, the obtained solution may be very far from representing the optimal solution. In real networks, the link error probabilities usually relate indirectly to the number of neighboring nodes. Hence we study the case where the error probability of a link is determined by 0.01 + 0.1 n where n is the number of neighboring nodes of the receiver. The tradeo curve is shown in Fig. 9. The tradeo curve with the assumption of the error probabilities of all links are 0.01 is also shown in the gure. Comparing these two curves, we see that although the network lifetime is still inversely proportional to the tradeo factor c, they achieve dierent tradeo between network lifetime and rate allocation of sensor nodes with the same value of c. This is another reason why we need to consider the impact of link error probabilities on the energy consumption model of sensor nodes. 7. Conclusions In this paper, we studied the tradeo between the network lifetime maximization and fair rate allocation problem with a reliability requirement in wireless sensor networks. We addressed the whole problem from the transport layer perspective, with multi-path routing. To guarantee reliable end-toend communication, because of its intrinsic energy eciency, a hop-by-hop retransmission scheme is adopted instead of the traditional end-to-end retransmission scheme used in reliable transport. We formulated the network lifetime maximization problem and the fair rate allocation problem as con-

40

J. Zhu et al. / Computer Networks 52 (2008) 2543

strained optimization problems, and invoked the NUM framework to guarantee fairness on the rate allocated sensor nodes. By introducing a tradeo factor c that characterizes the tradeo between the two conicting objectives, we combined these two problems into a single optimization problem, then derived a rate allocation algorithm to achieve the optimal tradeo. Since this algorithm is a partially distributed algorithm, it turned out to be less ecient in terms of message complexity. Thus, we further dened an additional utility function for each sensor node as a function of its normalized power dissipation, besides the utility with respect to its data rate. With this new utility, we could approximate the network lifetime maximization problem by invoking the NUM framework and studied the tradeo problem in a unied framework based on NUM. An ecient and fully distributed rate allocation algorithm is derived to approximate the optimal solution of the original problem. Numerical results show that both algorithms converge relatively fast to the optimal solution. We further observed that the fully distributed algorithm could approximate the optimal solution of the partially distributed algorithm very well, which makes it a good candidate for practical implementation. We also studied the impact of different route selection algorithms and link error distributions on the tradeo problem. We observed that the selection of routes and link error distribution have signicant inuence on the best tradeo achieved by the algorithms. From the results, we noticed that typically, exploring only a small set of candidate routes for each sensor node achieves performance levels that are close to those achieved by exploring all possible routes for each sensor node. This makes the distributed algorithm even more interesting for practical implementation. Appendix A. Proof of Proposition 1 (a) For each i 2 J(z*), dene X Qi y c w s U s x s 1 c z i :
s 2S

It is easy to verify that Qy jd mini2J z Qi y jd, and thus we have Qy jd Qy min fjry Qi y dg:
i2J z

A : 2

Note that we drop the third term on the right-hand side of (A.1) as it goes to zero much faster than the second term when j is small enough. Based on (A.2), it is simple to infer that clause (a) of the proposition holds. (b) We prove this by contradiction. Suppose (14) does not hold at an optimal solution y*, then there exists a p 2 J(z*) and a route r on it such that X cr y r w s U s x s 1 cry r zp > 0:
s 2S

If we increaseP y r by a relatively small positive value j, then both s2S ws U xs and zp increase, and the change of objective function Q(y*) is ! X DQy  j cry r ws U s xs 1 cry r zp > 0:
s2 S

Thus, the objective value is improved, which contradicts the assumption that y* is optimal. Hence clause (b) holds. Appendix B. Proof of Lemma 4 Dene matrices A1 and A2 in space RjSjjRj where 2 elements a1 ir and air are given, respectively, by ( es l ; l Lout i \ r ; 1sl Ei a1 ir 0; Lout i \ r /; ( r e ; l Lin i \ r; 1sl Ei a2 ir 0; Lin i \ r /: We then have z = (A1 + A2)y, and the objective functions of problems (11) and (19) can be written b1 as kA1 A2 yk1 and kA1 A2 ykb1 =b 1, respectively, where k k1 and k kp denote the maximum norm and the p-norm respectively. Suppose y* and y b are the optimal solution of problem (11) and (19), respectively. Note that y b is also an optimal solution of objective function k(A1 + A2)yk(b+1). We have the following inequalities:
1 2 kA1 A2 y b k1 6 kA A yb kb1

Let j be a small positive constant, and change the solution by jd such that J(z*) = J(z* + jd), then we have ! X 2 Qi y jd Qi y jry Qi y d o kd r :
r 2R

6 kA1 A2 y kb1 6 j Sj
1=b1

A:1

kA1 A2 y k1 :

J. Zhu et al. / Computer Networks 52 (2008) 2543

41

Denote the network lifetime in problem (11) and (19) by Tnet and T b net , respectively. We observe 1 that Tnet = 1/k(A1 + A2)y*k1, and T b net 1=kA 2 b A yb k1 . Thus, T net 6 T net . From the above inequalities, we have 1 j Sj
1=b1

References
[1] I.F. Akyildiz, W. Su, Y. Sankarasubramaniam, E. Cayirci, Wireless sensor networks: a survey, Computer Networks 38 (4) (2002) 393422. [2] Y. Xue, Y. Cui, K. Nahrstedt, A utility-based distributed maximum lifetime routing algorithm for wireless networks, in: Proceedings of the Second International Conference on Quality of Service in Heterogeneous Wired/Wireless Networks (QSHINE05), 2005, p. 18. [3] S. Tilak, N.B. Abu-Ghazaleh, W. Heinzelman, A taxonomy of wireless micro-sensor network models, ACM SIGMOBILE Mobile Computing and Communication Review 6 (2) (2002) 2836. [4] Z.H. He, S.K. Mitra, From rate-distortion analysis to resource-distortion analysis, IEEE Circuits and Systems Magazine 5 (3) (2005) 618. [5] V. Srinivasan, C.-F. Chiasserini, P.S. Nuggehalli, R.R. Rao, Optimal rate allocation for energy-ecient multipath routing in wireless ad hoc networks, IEEE Transactions on Wireless Communications 3 (3) (2004) 891899. [6] O.B. Akan, I.F. Akyildiz, Event-to-sink reliable transport in wireless sensor networks, IEEE/ACM Transactions on Networking 13 (5) (2005) 10031016. [7] F.P. Kelly, A.K. Maulloo, D.K.H. Tan, Rate control for communication networks: shadow prices, proportional fairness and stability, Journal of the Operational Research Society 49 (3) (1998) 237252. [8] J. Mo, J. Walrand, Fair end-to-end window-based congestion control, IEEE/ACM Transactions on Networking 8 (5) (2000) 556567. [9] M. Bhardwaj, A.P. Chandrakasan, Bounding the lifetime of sensor networks via optimal role assignments, in: Proceedings of IEEE Infocom, 2002, pp. 15871596. [10] Z. Hu, B. Li, On the fundamental capacity and lifetime limits of energy-constrained wireless sensor networks, in: The 10th IEEE Real-Time and Embedded Technology and Applications Symposium (RTAS 2004), 2004, pp. 29. [11] J.H. Chang, L. Tassiulas, Maximum lifetime routing in wireless sensor networks, IEEE/ACM Transactions on Networking 12 (4) (2004) 609619. [12] K.D.K. Kalpakis, P. Namjoshi, Ecient algorithms for maximum lifetime data gathering and aggregation in wireless sensor networks, Computer Networks 42 (6) (2003) 697716. [13] A. Sankar, Z. Liu, Maximum lifetime routing in wireless ad hoc networks, in: Proceedings of the IEEE Infocom, 2004, pp. 10891097. [14] V. Srinivasan, P. Nuggehalli, C.F. Chiasserini, R. Rao, Cooperation in wireless ad hoc networks, in: Proceedings of the IEEE Infocom, 2003, pp. 808817. [15] K. Kar, M. Kodialam, T.V. Lakshman, L. Tassiulas, Routing for network capacity maximization in energy-constrained ad-hoc networks, in: Proceedings of the IEEE Infocom, 2003, pp. 673681. [16] Y.T. Hou, Y. Shi, H.D. Sherali, Rate allocation in wireless sensor networks with network lifetime requirement, in: Proceedings of the ACM MobiHoc, 2004, pp. 6777. [17] F. Stann, J. Heidemann, RMST: Reliable data transport in sensor networks, in: Proceedings of the IEEE SNPA, 2003. [18] D. Tian, N.D. Georganas, Energy ecient routing with guaranteed delivery in wireless sensor networks, in: Proceedings of the IEEE WCNC, 2003.

T net 6 T b net 6 T net :

Thus, limb!1 T b net T net , and the lemma holds. Appendix C. Proof of Proposition 5 Since the data rates y on the routes are bounded, the source rates x are also bounded. Therefore, / 0 and w are nite as the gradient U (xs) is bounded for each sensor node s. We observe that function P(y) is dierentiable, hence the following inequalities regarding the gradient ry r P y hold: X b zi r y r zi ry r P y cws U 0s xs 1 c
i2r\S

6 cws U 0s xs 6 c/: As function U() is strictly concave, and function V() is strictly convex, function P(y) is strictly concave. By denition of concave functions, for any points y1 and y2, we have P(y1) P(y2) 6 $yP(y)T(y1 y2). Thus kP y1 P y2 k 6 kry P y y1 y2 k 6 kry P yk ky1 y2 k 6 R1=2 c/ky1 y2 k: From the above inequalities, we see that function P(y) is Lipschitz. Therefore, the limiting point of the sequences of points {y(t)} generated by the gradient algorithm with constant step size j is optimal [28, Proposition 1.2.3]. Next we derive the stepsize rule. Given the update direction d(t) at iteration t, we have    y r t  w  6 jry P ytj: jd r tj  r P y t r cU 0 x t y r  c
s s T

Thus, we have jry P yt dtj kdtk

w=ckry P ytk
2

2 2

w=c kry P ytk

c : w

According to [28, Proposition 1.2.3], we nd that if 0 < j < 2=R1=2 /w, the gradient algorithm converges to an optimal solution.

42

J. Zhu et al. / Computer Networks 52 (2008) 2543 John Ka-Lok Hung received his B.Eng. degree and M.Phil. degree in Computer Science and Engineering in 2004 and 2006, respectively, from the Hong Kong University of Science and Technology (HKUST). He is currently pursuing a Ph.D. degree at the HKUST. His research interests include modeling and performance evaluation of multi-hop wireless networks, cross-layer optimization, and protocol design for wireless networks.

[19] B. Deb, S. Bhatnagar, B. Nath, Information assurance in sensor networks, in: Proceedings of the ACM WSNA, 2003. [20] B. Deb, S. Bhatnagar, B. Nath, ReInForM: Reliable information forwarding using multiple paths in sensor networks, in: Proceedings of the IEEE LCN, 2003. [21] A. Willig, H. Karl, Data transport reliability in wireless sensor networks a survey of issues and solutions, Praxis der Informationsverarbeitung und Kommunikation 28 (2) (2005) 8692. [22] H. Nama, M. Chiang, N. Mandayam, Utility-lifetime tradeo in self-regulating wireless sensor networks: a cross-layer design approach, in: Proceedings of the IEEE ICC, 2006. [23] J. Zhu, S. Chen, B. Bensaou, K.-L. Hung, Tradeo between lifetime and rate allocation in wireless sensor networks: a cross layer approach, in: Proceedings of the IEEE INFOCOM, 2007. [24] V. Raghunathan, C. Schurgers, S. Park, M.B. Srivastava, Energy-aware wireless microsensor networks, IEEE Signal Processing Magazine 19 (2) (2002) 4050. [25] W. Heinzelman, A. Chandrakasan, H. Balakrishnan, An application-specic protocol architecture for wireless microsensor networks, IEEE Transactions on Wireless Communications 1 (4) (2002) 660670. [26] K. Miettinen, NonLinear Multiobjective Optimization, Kluwer Academic Publishers, 1999. [27] S. Boyd, L. Vandenberghe, Convex Optimization, Cambridge University Press, 2004. [28] D.P. Bertsekas, Nonlinear Programming, Athena Scientic, 1999. [29] N.Z. Shor, Minimization Methods for Non-dierentiable Functions, Springer-Verlag, 1985. [30] R. Madan, Z.Q. Luo, S. Lall, A distributed algorithm with linear convergence for maximum lifetime routing in wireless sensor networks, in: Proceedings of the Allerton Conference on Communication, Control and Computing, 2005. [31] F. Kelly, T. Voice, Stability of end-to-end algorithms for joint routing and rate control, ACM SIGCOMM Computer Communication Review 35 (2) (2005) 512.

Junhua Zhu is now a Ph.D. student in the Department of Computer Science and Engineering, the Hong Kong University of Science and Technology. He received his B.Sc. degree in Computer Science and Technology from Fudan University, PR China, in 2004. His current research interests are in the areas of communication quality and reliability in wireless sensor networks, and cooperation issues in wireless mesh and sensor networks.

Brahim Bensaou received an Engineering Degree in Computer Science (with distinction) from the University of Science and Technology Houari Boumediene of Algiers Algeria in 1982, and a DEA degree from University Paris XI in Computer Science in 1988. He earned his Doctorate degree in Computer Science from the University Paris VI in 1993. From 1990 to 1994, he was a research assistant at France Telecom Research labs near Paris, where he was involved in the early designs and study of ATM technology. In mid 1995 he joined the Hong Kong University of Science and Technology as a Research Associate where he spent nearly 2 years working on various problems in congestion and trac control. In 1997 he joined the Centre for Wireless Communications, a national R& D center in Singapore (now known as the Institute of Infocomm Research, I2R A-Star) as a Member of Technical Sta, where he worked as a System Architect on the design of Quality of Service (QoS) enabled MAC protocols and scheduling algorithms in a wireless ATM network prototype. In 1998 he was promoted to a Senior Member of Technical Sta, and was instrumental in forming a small R&D group in the area of wireless networking at the CWC. He led the group, for a year and a half, and then moved to Academia at HKUST in fall 2000 where he is now a faculty member in the Department of Computer Science and Engineering. His general areas of research are in QoS-enabled wired/wireless networks, including ad hoc networks, sensor networks and wireless LAN, where he has published more than 80 research papers in prominent conferences and journals, received numerous research grants, graduated nearly 20 postgraduate students out of which four Ph.D.s and invented three US patents of which one is licensed. He is an Associate Editor of The IEEE Communications Letters and is a senior member of the IEEE, and a member of the ACM.

J. Zhu et al. / Computer Networks 52 (2008) 2543 t-Abdesselam obtained his Farid Na engineer degree in computer sciences from University of Sciences and Technologies of Algiers (USTHB) Algeria, in 1993 and a master degree in computer Descartes sciences from University Rene (Paris 5) France, in 1994. After 2 years spent in the industry working as a software engineer, he joined the University of Versailles (UVSQ) France, and got his Ph.D. degree in computer sciences in 2000. In 1998, he worked as an associate researcher at University of Western Ontario, London Ontario Canada. From 09/1999 to 08/2000 he was an assistant professor at University of Sciences

43

and Technologies of Lille France. From 09/2000 to 08/2003 he was an associate professor at INSA of Lyon and a research nes Alpes. Since 09/2003 he is an member of the INRIA Rho associate professor at University of Sciences and Technologies of Lille and research member of the INRIA Futurs. His research interests lie in the eld of quality of service and security in IPbased networks, mobile ad hoc, sensor, and mesh networks, and peer to peer networks. He has been on the technical program committee of dierent IEEE conferences, including Globecom, LCN, and ICC, and chaired some of their sessions. He has also served as a technical program chair of the IEEE Workshop on Wireless Local Networks 2007 (WLN07) and the International Workshop on Peer to Peer Networking 2007 (PPN07).