Vous êtes sur la page 1sur 10

WIRELESS SENSOR NETWORKS

THE PROBLEM OF MEDIUM ACCESS CONTROL IN WIRELESS SENSOR NETWORKS


YALIN EVREN SAGDUYU AND ANTHONY EPHREMIDES, UNIVERSITY OF MARYLAND

ABSTRACT
A31 A1

R1

A32

The primary performance objectives of wireless sensor networks are energy conservation, throughput improvement, scalability, and self-configuration, whereas fairness and temporal delay are often the secondary issues. However, current wireless MAC protocols fall short from matching the characteristics of sensor networks.

In this article we revisit the problem of scheduled access through a detailed foray into the questions of energy consumption and throughput for MAC protocols in wireless sensor networks. We consider a static network model that rules out simultaneous transmission and reception by any sensor node and consequently requires partitioning of nodes into disjoint sets of transmitters and receivers at any time instant. Under the assumption of circular transmission (reception) ranges with sharp boundaries, a greedy receiver activation heuristic is developed relying on the network connectivity map to determine distinct receiver groups to be activated within disjoint time intervals. To conserve limited energy resources in sensor networks, the time allocation to each receiver group is based on the residual battery energy available at the respective transmitters. Upon activating each receiver group separately, the additional time-division mechanism of Group TDMA is imposed to schedule transmissions interfering at the non-intended destinations within separate fractions of time in order to preserve the reliable feedback information. The two-layered time-division structure of receiver activation and Group TDMA algorithms offers distributed and polynomial-time solutions (as required by autonomous sensor networks) to the problems of link scheduling as well as energy and throughput-efficient resource allocation in wireless access. The associated synchronization and overhead issues are not considered in this article.

INTRODUCTION
Wireless sensor networks consist of battery-operated sensing devices with computing, data processing, and communicating components. Such a network includes a large number of distributed sensor nodes that organize themselves into multihop wireless networks and collaborate on common tasks such as location sensing, event detection, and local actuator control. The primary performance objectives of wireless sensor networks are energy conservation, throughput improvement, scalability, and self-configuration, whereas fairness and temporal delay are often secondary issues. Since sensor nodes share a common wireless medium, an efficient medium

access control (MAC) operation is required. However, the current wireless MAC protocols such as IEEE 802.11 and Bluetooth fall short of matching the characteristics of sensor networks and cannot effectively support their applications. In this article we outline a throughput- and energy-efficient MAC approach that allows distributed implementation and supports multihop communication as required by autonomous and large-scale wireless sensor networks with high throughput needs and energy constraints. The extent of studies on multiple access has been traditionally limited to simple networks with multiple transmitters and a single destination. This model is clearly not sufficient to represent self-organizing wireless sensor networks with multiple dynamically changing transmitter-receiver pairs. As an extension of MAC operation to multidestination networks, Nguyen et al. [1, 2] looked at the problem of contention-based access in wireless networks with two fixed receivers and used conflict resolution algorithms to explore the bounds on the maximum stable throughput. The Group time-division multiple access (TDMA) algorithm was proposed in [1] as a time-division mechanism in a two-destination network in order to separate in time interfering groups of nodes with packets addressed to different destinations. The fundamental idea of scheduling transmissions is not new. However, its use in this context is. Each group is assigned separate fractions of time depending on traffic needs. The Group TDMA method was analyzed in terms of throughput properties in [2], and the optimal time allocation was determined as function of the offered loads independent of the underlying multiple access protocol within each group of users. This analysis can be extended to multidestination networks with arbitrary topology. In [1, 2], a fixed assignment of transmitterreceiver pairs is assumed in contrast to the dynamic and autonomous nature of sensor networks, where all nodes are both able as well as obligated to transmit and receive packets either as parts of source-destination pairs or for relaying purposes (as required by multihop operation in large-scale sensor networks). If we further assume that only a single transceiver per node is available, we need to rule out simultaneous packet transmission and reception by any node in the network. Then it is necessary to develop a

44

1536-1284/04/$20.00 2004 IEEE

IEEE Wireless Communications December 2004

mechanism that will activate nodes as either transmitters or receivers for disjoint time intervals. This requirement is unavoidable. The problems of deriving optimal channel access schedules for multihop networks and network partitioning into activation sets are both NP-complete [3, 4], and require heuristic suboptimal solutions for practical use. In this article we introduce a greedy receiver activation method based on partial knowledge of the network connectivity map to partition nodes into disjoint transmitter-receiver sets. Rather than ensuring conflict-free schedules as in standard link scheduling, we allow multiple transmission assignments to each receiver and rely on an arbitrary single-receiver MAC protocol to resolve the unavoidable packet conflicts. This protocol could be contention-based or conflict-free depending on the application and performance measures, such as throughput, energy efficiency, or complexity. To obtain reliable feedback information from each receiver, Group TDMA eliminates the secondary conflicts in terms of packet collisions due to transmissions at nonintended receivers. The predetermined receiver groups are activated within disjoint time intervals in a time-division mechanism. In wireless access, whether in cellular, ad hoc, or sensor networks, energy efficiency is of paramount importance. Inspired by recent work on energy-efficient multiple access in wireless networks [5], we can use battery energies and node lifetimes as decision criteria in temporal allocation for distinct receiver groups in order to make best use of finite and nonrenewable energy resources. The intuitive idea is to extend node lifetimes by allocating more time to transmissions by those nodes that have higher residual energy. In summary, we outline a resource allocation and link scheduling scheme based on two-layer time-division operation. The outer layer (1) allocates disjoint fractions of time (depending on residual energy) to activate distinct sets of receivers (predetermined on the basis of network topology). The inner layer (2) creates time orthogonality (based on throughput properties) between interfering transmitter groups for each receiver group.

nas and generate packets at a common rate. We assume that nodes address their packets to any receiver in their transmission ranges with equal probability. We consider a slotted system where all transmissions are synchronized into unit time slots. Any packet transmission is successful only if no other packet is simultaneously transmitted to the same receiver in the given time slot. A collision occurs if multiple transmitter nodes attempt to transmit (i.e., interfere) simultaneously. An idle slot is observed if there is no packet transmission attempt in the particular time slot. Each transmitter receives immediate and correct information on the channel output (i.e., whether a success, a collision, or an idle slot was observed in the preceding time slot). A separate channel based on scheduled access is dedicated to feedback control packets. These are the classical assumptions of the collision channel model.

TOPOLOGY-BASED RECEIVER ACTIVATION IN MULTI-DESTINATION NETWORKS


We propose a topology-based greedy heuristic to sequentially determine distinct receiver groups to be activated in a time-division mechanism. We now describe the logic of the mechanism for selecting receiver nodes. As the first step of receiver activation, an arbitrary node is chosen to initiate the first receiver group. The decision is either completely random or follows a particular prioritybased rule. Then the activated receiver node designates all nodes within a fixed chosen range as transmitter nodes. This can be done in a distributed manner by exchanging control information (about the current transmitter-receiver assignments) between the neighboring nodes. Next, a node outside the receiving range of the first activated receiver node is chosen as the second receiver. It also designates all nodes in its range (which may overlap with that of the first receiver) as transmitter nodes to itself. We continue with sequential assignments of transmitter-receiver pairs until all nodes are included in either a receiver or transmitter activation group at least once. We repeat the same procedure several times by selecting next a previously designated transmitter node as a receiver and running the same algorithm. Thus, we create a sequence of different receiver activation groups, until each node is activated both as a receiver and a transmitter at least once in a full cycle of activation periods. Actually, including each node in (at least) one receiver group is a sufficient condition to terminate the process of forming new distinct receiver groups, since all nodes will be consequently designated also as transmitters at least once over the full cycle of receiver activation periods.

Any packet transmission is successful only if no other packet is simultaneously transmitted to the same receiver in the given time slot. A collision occurs if multiple transmitter nodes attempt to transmit simultaneously.

THE SYSTEM MODEL


We consider a static model of wireless sensor networks in which simultaneous transmission and reception by any node is not allowed. However, each node needs to act (interchangeably) as transmitter or receiver. We assume circular transmission (reception) ranges with sharp boundaries such that no successful transmission or interference can be observed beyond that range. This is a simplifying unrealistic assumption; however, we allow these circular regions to overlap significantly, and thus the essence of interference is captured. We assume that each node lies within the transmission (reception) range of at least one other node. We consider arbitrarily large (or renewable) energy resources and fixed traffic load distributions among transmitters. In this article we consider the problem of multiple access for both unlimited transmitter population and a finite number of nodes. Transmitters are equipped with omnidirectional anten-

ILLUSTRATIVE EXAMPLE OF RECEIVER ACTIVATION


We use the simple network shown in Fig. 1 to illustrate the greedy receiver activation heuristic. We pick node 1 as the first activated receiver. Nodes 2, 3, 4, and 5 are designated as transmitters, since they are within the receiving range of node 1. If node 6 is the second activated receiver, we classify nodes 7 and 8 as transmitters. Similarly, if node 10 is the last activated receiver, nodes 9, 11, and 12 are classified as transmitters.

IEEE Wireless Communications December 2004

45

7 3 1 4 2 12 5 11 6 8 10 9

receiver is activated in each successive receiver group. For the case of n nodes in the region under consideration, the number of distinct receiver groups N (note this is a different quantity than Ni) has the following bounds: I. Tandem networks: n N n, L min g(n), + 1 r II. Planar networks: n Nn 3 A min h(n), 2 r

n Figure 1. The simple multihop network model for illustration of the receiver
activation method. We exclude nodes 1, 6, and 10 from the future list of receiver candidates, and repeat the same procedure, until all nodes are activated at least once as receivers. The node sets {1, 6, 10}, {2, 4, 7, 12}, {5, 3, 8, 12}, and {9, 11, 6, 1} form valid receiver groups that can be activated in sequence. The question is what time fraction of activation to assign to each of these groups. This issue is addressed later.

STARTING POINT: RANDOM ACCESS FOR COLLISION CHANNELS


The problem of multiple access for a single activated receiver has been extensively studied and well-known random access (e.g., stabilized slotted Aloha, First-Come-First-Serve [FCFS] algorithm) and scheduled access (e.g., time-, frequency-, or code-division multiple access) solutions have been proposed. When we consider a large number of transmitters, some form of random access (rather than scheduled access) is unavoidable. The splitting algorithms for collision resolution in random access provide higher values of stable throughput than stabilized slotted Aloha. Splitting a group of packets that have collided in a slot can be implemented based on various criteria, such as coin toss, node or packet ID, arrival time, or residual energy [5]. In this article we consider the classical collision channel model and use the FCFS algorithm to resolve packet collisions, although almost any arbitrary MAC protocol can be assumed. The FCFS algorithm initiates a new collision resolution period (CRP) whenever a packet collision occurs (i.e., multiple packets are simultaneously transmitted in a given time slot). All packets that arrive within a specified time allocation interval are transmitted in the first period of this CRP. If there is another collision, the time allocation window is further shortened and the same procedure is repeated, until all packets involved in the original collision are successfully received. Synchronization among transmitter-receiver pairs can be achieved as the receiver node sends some form of synchronization information initiating a new communication with a control message. Transmitter nodes synchronize their clocks with the receiver on receiving this information. The exact method of synchronization and further implementation issues are out of scope of this article.

TOPOLOGICAL CONSTRAINTS ON THE SIZE AND NUMBER OF RECEIVER GROUPS


The number of activated receivers in each receiver group is constrained by the total number of nodes n as well as by the topology properties including the network size and the radius r of the common reception (and transmission) range. We define N i as the number of elements in receiver group Gi, for 1 i N. The cardinality of any receiver group strongly depends on the effects of overlapping between the reception ranges of the activated receiver nodes. The upper bound on Ni is imposed by the total number of nodes n and the constraint that the separation between two activated receivers cannot be smaller than the radius r of reception range. The lower bound on N i can be achieved if the reception regions of activated receivers are nonoverlapping but still cover the remaining nodes (already designated as transmitters). For illustrative purposes, we consider a tandem network of length L and a planar network of area A. For the case of n nodes in the region under consideration, the number of activated receivers N i in any receiver group G i has the following bounds: I. Tandem networks: 1 Ni min {g(n), L/r +1} II.Planar networks: 1 Ni min {h(n), 3A/r2} where the quantities g(1) = 1, g(n) = 2(n 1)/3 for n > 1 and h(1) = 1, h(n) = 12(n 1)/13 for n >1 are derived from basic geometry. The proposed receiver activation approach produces a minimum number of distinct receiver groups if the maximum number of receiver nodes (with minimum overlapping between their reception ranges) is included in each receiver group and each node only appears in a single receiver group. On the other hand, the maximum number of receiver groups is achieved if only one new

GROUP TDMA ALGORITHM FOR TWO-DESTINATION NETWORKS


In this section we look at the multiple access problem in a simple network model with two activated receivers (i.e., two potential destina-

46

IEEE Wireless Communications December 2004

tions) and multiple transmitters, as shown in Fig. 2. We assume that transmitters are within the reception range of at least one receiver. Transmission ranges are circular with sharp boundaries, and beyond that range no transmission or interference is possible. However, these circles may overlap significantly. Packets can be addressed to either of the two receivers with equal probability if the nodes lie in the intersection of these circles. Otherwise, the packets are destined for the receiver they can reach. Nodes have immediate access to ternary feedback from their intended destinations: whether a collision, a success, or an idle was observed during the preceding slot. We assume unlimited populations of unbuffered users in each region. We assume that these users follow an arbitrary MAC protocol with maximum stable throughput Smax. For example, this could be slotted Aloha with Smax = 1/e = 0.3679 or FCFS with Smax = 0.4878 (packets per time slot).

A31 A1 A2

R1

A32

R2

n Figure 2. The simple two-destination network model for illustration of the


Group TDMA algorithm. tion, we define f1, f2, f31 and f32 as the fractions of the total traffic load generated by transmitter groups A1, A2, A31, and A32, respectively. If we employ Group TDMA with temporal assignments of x1, x2, and x3 to transmitter groups {A1, A2}, A31, and A32, the total stable throughput must satisfy f1 x1 Smax, f2 x1 Smax, f31 x2 Smax, f32 x3 Smax (1)

FEEDBACK RELIABILITY PROBLEM IN MULTI-DESTINATION NETWORKS


We denote receiver nodes R1 and R2, as shown in Fig. 2. For each time instant (or time slot), we identify four distinct groups of transmitters. We define A1 and A2 as disjoint transmitter groups in the reception range of each receiver node R1 and R 2 , respectively. We define A 3 as the group of transmitters that lie in the intersection of the two circles and hence have both receivers in their transmission ranges. We assume that nodes from groups A1 and A2 randomly generate packets destined only to R 1 and R 2 , respectively, whereas nodes of group A 3 transmit to either R 1 or R 2 with equal probability. At each time the group A3 consists of groups A 31 and A 32, where nodes in A3i, i = 1, 2, are transmitting to receiver Ri. If the MAC protocol used is stabilized Aloha, there is no problem of propagation of feedback error. If, however, we use FCFS with the socalled first improvement [6], there is potential instability created by misinterpretation of the feedback signal from the channel (since collisions at Ri can be caused by packets destined for either receiver); this is equivalent to errors in the feedback signals. However, if plain FCFS is used, there is no problem of such stability. The cross-collisions between packets with different destinations may reduce the maximum stable throughput value but will not cause instability at low traffic rates.

where 0 x 1 , x 2 , x 3 1, and 3 x i = 1. Our i=1 objective is to maximize over x1, x2, and x3 subject to the stability conditions, which are given by Eq. 1 in terms of linear constraints. The first two constraints in Eq. 1 can be combined to max(f1, f2) x1Smax. Then the stable throughput is maximized if we choose x1, x2, x3 to satisfy each of the constraints max(f 1, f 2) x 1S max, f 31 x 2S max, and f32 x3Smax with equality. The throughputoptimal temporal allocation is given by max( f1, f2 ) * x1 = , max( f1, f2 ) + f31 + f32
* x2 = * x3 =

f31 , max( f1, f2 ) + f31 + f32 f32 max( f1, f2 ) + f31 + f32

(2)

DESCRIPTION OF THE GROUP TDMA ALGORITHM


A solution to the feedback reliability problem was proposed in [1, 2] promoting the Group TDMA algorithm as a time-division mechanism to distinguish four groups of transmitters, A1, A2, A31, and A32, and schedule transmissions of {A1, A 2 }, A 31 , and A 32 over three nonoverlapping time intervals x 1 , x 2 , and x 3 , respectively. The slots allocated to various groups do not need to be implemented as contiguous blocks of slots and can have an arbitrary order as long as the overall frame length is sufficiently large so that any given slot apportionment (i.e., fractions x1, x2, and x3) can easily be realized. For the case of unlimited transmitter popula-

This is derived in [2] and can be obtained by solving the simple linear program. The solution suggests that separately activated transmitter groups with larger traffic loads should be allocated for longer time intervals. We can interleave the slots dedicated to different groups as long as the resulting time allocation satisfies Eq. 2. We assume sufficiently large frame lengths so that we can exactly implement the optimal time fractions.

THROUGHPUT EFFICIENCY OF THE GROUP TDMA ALGORITHM


We compare Group TDMA to the options of: 1 Activating all transmitter-receiver pairs simultaneously (i.e., no time division) 2 Activating receivers one at a time (i.e., each group is activated separately)

IEEE Wireless Communications December 2004

47

A1 A1 A31 R2 R1 A4 A32 r

A2

r: Transmission radius

: Guard zone parameter

n Figure 3. The simple two-destination network operating under the protocol

model with more realistic interference effects (solid and dashed circles correspond to the transmission and interference ranges, respectively.)

Then the maximum stable throughput under method 1 must satisfy

* = 1

Smax , max( f1, f2 ) + f31 + f32

and under method 2 it must satisfy

* = 2

Smax = Smax f1 + f2 + f31 + f32

where S max is the maximum stable throughput achievable for the single destination case without interference from the adjoining group and S is the maximum stable throughput for the max single destination with neighboring node inter ference. Clearly, S max S max . By contrast the maximum stable throughput of the Group TDMA satisfies

* =

Smax . max( f1, f2 ) + f31 + f32

Since max(f 1 , f 2 ) f 1 + f 2 and S max S max , we see that Group TDMA achieves higher stable throughput than method 1 or 2.

GROUP TDMA ALGORITHM UNDER A MORE REALISTIC CHANNEL MODEL


So far, we have assumed that the transmission (or equivalently interference) and reception ranges are both circular areas with common sharp boundaries, and beyond that range no transmission or interference can be observed. We can adapt a more realistic criterion for successful packet reception, the protocol model (as introduced in [7]) that extends the interference effects beyond transmission ranges. According to the protocol model, all nodes have a common range r for transmissions, and a transmission is successful if and only if the distance between the intended transmitter-receiver pair is smaller than or equal to a transmission-reception radius r, and the dis-

tance of every other concurrent transmitter to the particular receiver is greater than or equal to the interference radius (1 + ) r. The quantity 0 accounts for a guard zone that prevents a neighboring node from transmitting over the same single channel at the same time. Figure 3 illustrates the two-destination network model, in which the transmission and interference ranges become distinguishable due to the additional interference effects. We partition the network into five subregions with distinct transmission and interference properties. Regions 1 and 2 contain nodes that have only receivers R1 and R2 as their destinations, respectively, and cannot cause interference at the other receiver. Nodes in the reception ranges of both receivers are included in region 3. On the other hand, region 4 consists of nodes that are in the reception range of R1 but can also interfere at R2. Similarly, region 5 consists of nodes that are in the reception range of R2 but can also interfere at R1. We denote fi the fraction of traffic load generated by transmitter nodes in region i, where 1 i 5. The set of the nodes in region i is denoted by Ai. We partition A3 into two subgroups, where A3,1 and A3,2 denote nodes that attempt to reach receiver R1 and R2, respectively, with the traffic loads f3,1 and f3,2. According to the Group TDMA formulation, we divide the time interval into three disjoint time fractions, x 1 x 2 and x 3 The , , . packet transmissions from A1 and A2 are allocated within x1 fraction of time. The packet trans missions from A3,1 and A4 are allocated within x2 fraction of time. The rest of the time, the time fraction of x3, is dedicated to transmissions from A3,2 and A5. The optimal temporal allocation and maximum stable throughput are still given by Eqs. 2 and 3, where we need to replace x* by xi* i for i = 1, 2, 3, and we let f1 = f1 f2 = f2 f31 = f31 , , + f4 and f32 = f32 + f5. , We can easily evaluate the effects of on * under the assumptions that the traffic load fi is directly proportional to the area of region i and nodes address their packets with equal probability to any of the receivers in their transmission ranges. Figure 4 depicts the maximum stable throughput per destination, */Smax, as a function of the traffic load fraction f 3 for different values of . To simplify the analysis, we continue in the rest of the article with the assumption of = 0. The important thing to note in this section is that more realistic models (i.e., with > 0) can be tracked in the same fashion as we analyze Group TDMA.

GROUP TDMA IN MULTI-DESTINATION NETWORKS


We illustrate Group TDMA operation using the simple network shown in Fig. 1. If nodes 1, 6, and 10 are activated in the first receiver group, we include nodes 2 and 5 in A 1 , node 7 in A 2 , and nodes 9, 11, 12 in A3. Nodes 3 and 4 belong to A1,1 or A1,2 depending on whether their packets are addressed to node 1 or 6, respectively. Node 8 belongs to A2,2 or A2,3 if its packets are addressed to node 6 or 10, respectively. The activation period is divided into fractions of time to be assigned separately to each transmitter group. Nodes from A1,

48

IEEE Wireless Communications December 2004

A 2 , A 3 are allocated within x 1 fraction of time. Nodes from A1,1, A2,3 are allocated within x2 fraction of time. The residual fraction of time x 3 is assigned to transmissions from A1,2 and A2,2.

1 =0 = 1/4 = 1/2 =1 2

THROUGHPUT-OPTIMAL TIME ALLOCATION BY GROUP TDMA


We assume systems with either unlimited energy resources or an unlimited number of nodes to eliminate changes in the traffic load characteristics of any transmitter group. We consider the receiver activation period in which a particular receiver group G i is activated. The transmitter nodes in that group are divided into ci number of disjoint transmitter groups Gi,k, 1 k ci. Transmitter group Gi,k is activated within xi,k fraction of time in the given receiver activation period such ci that k=1 x i,k = 1. We define R i,k as the set of receivers that can be reached by transmitter group G i,k and define the jth receiver in group (j) (j) R i,k R i,k . We define A i,k as the subset of nodes that belong to the kth transmitter group Gi,k and (j) have packets destined to the receiver Ri,k . The fraction of the traffic load generated by node (j) (j) group Ai,k is given by fi,k . We assume that each transmitter node generates one-hop packet transmissions with equal rate (either as a source or for relaying purposes). Provided that receiver group Gi is activated, the quantity i stands for the total rate of packet arrival to the rest of the nodes that are classified as transmitters in this period. All nodes in transmitter group Gi,k, k {, , ci}, must jointly satisfy the stability condition of i f(j) xi,k Smax, j: R(j) Ri,k i,k i,k (4) The optimal temporal allocation and maximum achievable throughput can be obtained according to a linear programming solution (similar to the one previously obtained for the two-destination case), and are given by xi*, k = max j: R( j ) R
i ,k i ,k

0.9

0.8 1* /Smax

0.7

0.6

0.5

0.4 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Fraction of traffic f3

n Figure 4. Effects of the receiver range overlapping (i.e., the value of f3 ) and

interference level (i.e., the value of in the protocol model) on the maximum value of the normalized stable throughput (i.e., */Smax).

from different groups can cause secondary interference effects. Therefore, transmitter groups Gi,1, Gi,2, and Gi,2 (i.e., the packet transmissions originating from the three distinct transmitter groups) need to be activated separately within disjoint fractions of time x1,i, xi,2, and xi,3, respectively. The special nature of the tandem topology reduces the number of distinct transmitter groups to three for each activated receiver group.

fi(kj ) ,

TOPOLOGY-INDEPENDENT BOUNDS ON THE MAXIMUM STABLE THROUGHPUT


Consider a tandem network deployed over length L and a two-dimensional planar network of area A both with common transmission radius r. We denote * the maximum stable throughput i achievable by an unlimited population of nodes operating according to the Group TDMA method during the ith receiver activation period. For any tandem and planar topology, we have Smax L S L * max and i 4r 2r Smax A S A * max i (3 3 + )r 2 r 2

ci ( j) j k =1 max j : Ri(,k) Ri ,k fi, k

, k {1,ci} and (5)

* = i

Smax

ki=1 max j: Ri(,kj) Ri,k

fi(kj ) ,

Note that the quantity * does not represent i the end-to-end throughput but denotes the maximum number of successful one-hop transmissions per time slot.

(6)

ILLUSTRATIVE EXAMPLE OF TANDEM NETWORKS


As an extension to networks with an unlimited number of destinations, we consider the tandem network model with at most two-neighbor connectivity where nodes are placed in tandem on a single line. Provided that the receiver group Gi is activated, we enumerate the activated receivers from left to right. Transmitters that are in the reception range of one receiver only belong to the transmitter group Gi,1. The rest of the transmitters (i.e., transmitters in the reception ranges of two receivers) are divided into two subgroups Gi,2 and Gi,3 such that Gi,2 consists of nodes transmitting to the odd-numbered receivers and Gi,3 consists of nodes transmitting to the even-numbered receivers. Simultaneous transmissions of nodes

where S max is the maximum stable throughput achievable by an arbitrary MAC protocol for the single destination case.

GROUP TDMA ALGORITHM FOR THE FINITE POPULATION OF TRANSMITTERS


Instead of revisiting the stability problem of multiple access systems with finite numbers of transmitters [8], we assume that the maximum achievable (stable) throughput of MAC protocol Smax(Ti) is known as a function of the number of transmitters T i . We use the following simple example to illustrate the temporal allocation

IEEE Wireless Communications December 2004

49

We propose to use node lifetimes and energy consumption rates as measures for the time allocation to the different receiver groups. This is a sensible criterion for sensor networks.

E0(G2) E1(G4) E2(G1)

E0(G1) E0(G2) E0(G3) E0(G4)

E1(G1) E1(G2) E1(G3) E1(G4)

E2(G1) E2(G2) E2(G3) E2(G4)

n Figure 5. Illustration of temporal allocation among four receiver groups over successive activation periods.
K K K

solutions of the Group TDMA algorithm in the period and Em(Gi) the change in the amount of case of finite number of transmitters. We assume energy Em(Gi) during the mth period. that the ith receiver group G i is activated with The intuitive energy-efficient solution for receiver group activation is that group RG m two receivers and three transmitters. The recepshould be the one (from all possible receiver tion region of each receiver contains two transE mitters,Ewhereas the(G ) E (G ) overlapping area(G ) two ) E (G ) E (GGi, 1 i EN) that maximizes (Gm(Gi). In E4 of E4(G2 4 groups 4) 3(G1) E3(G2) E3 3 3 4 1 3 4 5(G1) E5(G2) E5(G3) E5 4) other words, we activate (at any time instant) reception ranges includes only one transmitter. only the receiver group for which the respective Transmitters are divided into four groups each (1) (2) transmitters have the highest amount of residual of cardinality one. Nodes in Ai,1 and Ai,1 are cumulative energy. activated within a fraction of time equal to xi,1, (1) (2) We denote LT the function to be maximized, whereas transmissions of nodes in Ai,2 and Ai,3 namely the residual system lifetime, which is are assigned within xi,2 and xi,3 fractions of time, defined as LT = min1iN LTi, where LTi denotes respectively. If we assume that the overall traffic load is homogeneously distributed among transthe lifetime of energy supplies dedicated to mitters (i.e., each transmitter has the same packpacket transmissions to receiver group Gi. et arrival rate), the load fractions are given by The approach to the given optimization prob(1) (2) (1) (2) fi,1 = fi,1 = 1/3, fi,2 = fi,3 = 1/6. Under the stalem is based on load balancing, that is, equalizing bility condition given byEq. 4 with S max = the cumulative residual energies of different node Smax(1), the optimal temporal allocation is x* = groups. The idea is to keep the quantity Em(Gi) i,1 * * 1/2, x i,2 = x i,3 = 1/4. The maximum achievable of all Gi close to each other over successive actithroughput can be expressed as * = 3/2Smax(1), vation periods m so that no node group (transmiti where Smax(1) = 1 is the obvious solution due to ting to a particular receiver group) runs out of energy earlier than other node groups. As a result, the absence of primary or secondary interference the minimum of node lifetimes, LT, is maximized. effects for this simple example. Hence, 3/4 packets per slot can be transmitted to each receiver The underlying theoretical solution for the on the average in this example. given receiver activation policy is that the mth activated receiver group RGm maximizes Em(Gi), as lim tm 0 for all values of m. A new receiver ENERGY-EFFICIENT TEMPORAL ALLOCATION activation period m + 1 is initiated only if the residual energy of transmitters for RG(m) falls FOR RECEIVER ACTIVATION below the residual energy of transmitters for another receiver group. The optimal solution We now depart from the discussion of Group suggests switching between receiver groups with TDMA for a given receiver group and address infinitesimal activation durations. A suboptimal the issue of scheduling the activation of the difbut practical solution is to activate first the receivferent receiver groups. If these receiver activaer group with the highest total energy of corretion groups have been predetermined based on sponding transmitters and to replace RGm with the network topology according to the heuristic outlined earlier, it remains to determine the actianother receiver group for receiver activation vation order and duration of each receiver group period m + 1 if E m+1 (RG m ) falls below minGiRGmEm+1(Gi) c. We introduce the conGi. We propose to use node lifetimes and energy consumption rates as measures for time allocastant c to prevent rapid changes in the activation tion to the different receiver groups. This is a process. An intuitive solution for selecting the sensible criterion for sensor networks. We define length of activation period with RG m = G i is RGm as the receiver group activated in the mth c/(Gi), where (Gi) denotes the rate of change activation period with an allocated time fraction (measured in unit energy per unit time) in the of tm. We assume that the energy of each node is cumulative residual energies of the nodes whenever receiver group Gi is activated. A sample soluequally dedicated to transmissions for each receiver in its transmission range. We denote tion is illustrated in Fig. 5 for four receiver groups. Next, we explore the optimal time allocation Em(Gi) the total energy available for transmiswith c = 0. According to the energy-efficient sions to G i before the mth receiver activation

50

IEEE Wireless Communications December 2004

(y): energy consumption rate

receiver activation, we select Em(Gi) for all Gis with the objective of making the node lifetimes approach each other. In other words, the length of temporal allocation to G i, namely i, should be selected inversely proportional to (Gi). The value of i is simply given by E0 (Gi ) (Gi ) , i = N E0 (Gi ) i =1 (G ) i 1 i N

1.4 Stabilized slotted aloha FCFS without improvement FCFS with improvement FCFS without improvement

1.2

1 Stabilized slotted aloha

(7)

0.8

0.6

The energy consumption rates (Gi), i {1, , N} strongly depend on the underlying MAC protocol and the transmitter-receiver activation by the Group TDMA algorithm. We can express (Gi) for any i {1, , N} as follows:

0.4

FCFS with improvement

0.2

(Gi ) =

k =1

xi , k

ci

j j : Ri(,k) Ri ,k

(i fi(kj ) / xi, k ) ,

(8)

0 0 0.05 0.1 0.15 0.2 0.25 0.3 y: arrival rate 0.35 0.4 0.45 0.5

where (y) is the energy consumption rate of an arbitrary single-receiver MAC protocol (e.g., FCFS) operating with rate y packets per time slot. The time fractions xi,k, load fractions fi,k(j), number of transmitter groups c i and stable throughput i can be obtained from the previously outlined analysis of Group TDMA. The energy consumption rates (y) are depicted in Fig. 6 as function of the achievable stable throughput y under MAC protocols such as the stabilized slotted Aloha and FCFS algorithms with and without improvement. Although in this section we omit the details of the analysis and the notation is somewhat cumbersome, we intend to show that the energybased criterion results in a concrete solution to the optimal receiver group activation schedule.

n Figure 6. The energy consumption rate (y) as function of the achievable stable
throughput y (with unit energy for packet transmissions). links without violating any of conditions IIV for the same fraction of time. This can be formulated as a link coloring problem [9]. The problems of determining the edge chromatic number of graphs (i.e., the fewest number of colors necessary to color each graph edge so that no two graph edges incident on any graph vertex have the same color) and optimal link scheduling are both NP-complete [3, 4, 10]. Instead of solving the standard scheduling problem, we rely on a receiver activation heuristic to determine disjoint subsets of transmitters and receivers at each time instant (so that condition I is satisfied and possible violations of other conditions are reduced, but not eliminated, for all links) and on the Group TDMA method to create time orthogonality between links violating conditions II and IV.

JOINT RECEIVER ACTIVATION AND GROUP TDMA AS SOLUTIONS TO MODIFIED LINK SCHEDULING PROBLEM
STANDARD AND MODIFIED LINK SCHEDULING PROBLEMS
The standard form of general link scheduling [3, 4, 9] involves the assignment of channels (i.e. time slots, frequencies or codes) to connecting links between nodes so that all links assigned to the same channel may transmit in a conflict-free fashion. The network topology is described by a directed graph where directional links between nodes are only possible if nodes are within each others transmission-reception ranges. For conflict-free packet transmission, the following conditions should be satisfied: I Nodes cannot simultaneously transmit and receive packets. II Nodes cannot transmit packets to multiple destinations in the same time slot. III Primary packet conflicts multiple number of simultaneous transmissions to the same receiver are not allowed. IV Secondary packet conflicts interference effects at nonintended receivers are not tolerated. Standard link scheduling allocates conflict-free

DISTRIBUTED IMPLEMENTATION FOR GROUP TDMA ALGORITHM


We can set up the transmitter group classification as a link coloring problem. We assume that transmitters can discover receivers up to a twohop distance. Two receivers are called neighbors if there is at least one transmitter in the intersection of their reception ranges. Interfering transmitter groups are assigned to distinct fractions of time (i.e., different colors are assigned to links from different transmitter groups). Transmitters with only one receiver in their transmission ranges acquire membership in group A1, and all links from A1 are given color C1. Next, an arbitrary receiver R1 is selected such that any transmitter that has multiple receivers in its transmission range including R1 as the intended destination initiates a transmitter group A2 (i.e., links from A 2 to R 1 are given a new color, C 2). Next, we consider all neighbors of R 1 for link coloring purposes. If R2 is a neighboring receiver of R1, we assign different colors to all links from transmitters that lie in the intersection of R1 and R 2 to the particular receiver R 2 . We continue with coloring links to receivers one by one.

IEEE Wireless Communications December 2004

51

0.5 Maximum throughput per destination (packets/slot) 0.45 Simultaneous operation Group TDMA

main contribution in this article and the principal point we want to get across.

NUMERICAL RESULTS
For numerical evaluations, we consider both static tandem and planar network models with 1000 unbuffered nodes as approximate models for infinitely dense wireless sensor networks. We consider systems with first unlimited energy supply and then with hard finite energy constraints. For the latter case, we assume that each node has an amount of initial battery energy Emax = 106 (unit energy). Each packet transmission consumes r 2 units of battery energy. We assume that nodes generate packet transmissions with the same rate according to a common Poisson process and employ the FCFS collision resolution algorithm to resolve primary packet conflicts. In the single destination case, the maximum stable throughput achievable by the unlimited node population employing the FCFS algorithm (with the first improvement) is Smax = 0.4878 (packets per unit time or time slot), which represents only a lower bound on the maximum stable throughput achievable by a finite number of transmitters. The value of the common transmission (reception) radius characterizes the distribution of the activated transmitter-receiver pairs on the network as well as specifies the overlapping between the reception regions, on which the operation of receiver activation and Group TDMA strongly depends. To illustrate the topology effects, we introduce the quantities 2r/L and r 2/A, which denote the ratios of the transmission range to the network length and network area in tandem and planar networks, respectively. We first apply the topology-based receiver activation heuristic (without energy-efficient solutions) to unlimited energy systems and compare the Group TDMA algorithm with simultaneous operation of the activated receivers. For both cases, equal fractions of time are allocated to each receiver group. The network approaches a single-destination system for large values of transmission ranges, whereas the number of one-destination systems increases with smaller transmission ranges. Figure 7 depicts the maximum achievable throughput per destination (over a single hop) as functions of the quantities 2r/L and r 2 /A for tandem and planar networks. Simulation results indicate the superior performance of the Group TDMA algorithm over simultaneous operation of the transmitter-receiver pairs (for the entire range of transmission radius r). Lower values of the maximum stable throughput are achieved in planar networks than in tandem networks. This is expected because of the increased overlapping effects between the reception ranges in planar topologies. We also consider systems with hard energy constraints and evaluate the performance improvement by incorporating energy-efficient solutions into the topology-based receiver activation. We run both receiver activation heuristics over the layer of the Group TDMA algorithm operating with the maximum stable throughput. The energy-efficient receiver activation has the objective of maximizing the system lifetime,

0.4

0.35 Tandem networks

0.3 0.25 0.2

Planar networks 0.15 0.1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 2r/L or r2/A = ratio of transmission range to network length or area

n Figure 7. Comparison of Group TDMA and simultaneous operation of transmitter-receiver pairs for different reception ranges in tandem networks. Transmitters that originate links with new colors initiate new transmitter groups. For receiver activation group Gi, we denote ri the maximum number of intersections of reception ranges (i.e., the maximum number of neighbors) for each receiver and we denote ei the modified edge chromatic number, which is the minimum number of colors necessary to color graph edges so that no two graph edges violating condition IV have the same color (i.e., links to neighbor receivers are assigned different colors). Note that ri + 1 ei, which follows from plain geometry. Since the separation between receiver nodes is greater than the reception radius r (as a consequence of receiver activation), there exists a fixed upper bound on the number of intersections of reception ranges for each receiver. At most 13 different colors are needed for planar networks (with ri = 12), and at most three different colors are needed for tandem networks (with ri = 2). As a result, transmitters with any intended destination R choose one of the finite number of available group memberships different than those previously acquired by other transmitter groups with intended destinations that are neighbors to R. Packets are addressed randomly to any of the receivers in the transmission range so that condition II is also satisfied. If the receiver activation has already partitioned nodes to subsets of transmitters and receivers (so condition I is satisfied for all links), the remaining problem of creating time orthogonality among transmitter groups (so conditions II and IV are satisfied for all possible links) can be solved in polynomial time by the distributed Group TDMA method. So we see that the entire problem of scheduling transmitters and receivers in a sensor network, a problem that needs to be solved one way or another, reduces to a combination of standard graph coloring techniques and Group TDMA with arbitrary MAC protocol. This is the

52

IEEE Wireless Communications December 2004

which is defined as the length of time interval from the start of network operation until the first time the energy supplies dedicated to any of the activated receiver groups are completely depleted. The corresponding time allocation follows the solution outlined earlier. On the other hand, the topology-based receiver activation (without energy-efficient solutions) allocates equal fractions of time to each receiver group. Figure 8 depicts the system lifetimes for both receiver activation heuristics. Simulation results verify that the solutions based only on the connectivity map are outperformed by the energyaware receiver activation heuristic for all transmission ranges. The gap between the two heuristics increases for intermediate values of r, where there are several potentially interfering multidestination systems. The performance of both methods becomes identical as r increases, so in the end we have a single activated receiver in each receiver activation group.

3.5

x 109 Energy-efficient receiver action Purely topology-based receiver action

3 System lifetime (number of slots)

2.5 Tandem networks 2

1.5 Planar networks 1

0.5 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 2r/L or r2/A = ratio of transmission range to network length or area

CONCLUSIONS
In this article we rediscover the value of scheduled access in wireless sensor networks from the perspectives of throughput and energy efficiency. We propose a two-layered time-division mechanism based on receiver activation and Group TDMA as a form of link scheduling and resource allocation with suboptimal but polynomial time solutions. We develop a topology-based greedy heuristic to determine distinct receiver groups to be activated within disjoint fractions of time, and determine temporal allocations based on cumulative battery energies left at transmitter groups to extend the node lifetimes. We use the Group TDMA method to formulate a linear programming solution to the problem of throughputoptimal temporal allocation for transmissions to activated receivers, and derive bounds on the maximum stable throughput for tandem and planar networks. We also evaluate via numerical examples the performance improvement by energy-efficient receiver activation and throughput-efficient Group TDMA. Far from constituting a complete solution to the MAC issue in sensor networks, this work identifies a fruitful approach to handle the coordination of transmissions and receptions in such a network. It is only a first step in the process of examining the operation of sensor networks, but it focuses on the problem at a fundamental level, rather than at a level of immediate deployment.

n Figure 8. Comparison of energy-efficient and topology-based receiver activations for different reception ranges in tandem and planar networks.
[6] D. Bertsekas and R. Gallager, Data Networks, 2nd Ed., Prentice Hall, 1992. [7] P. Gupta and P. R. Kumar, The Capacity of Wireless Networks, IEEE Trans. Info. Theory, vol. 46, no. 2, Mar. 2000, pp. 388404. [8] W. Szpankowski, A Multiqueue Problem: Bounds and Approximations, Adv. Appl. Probab., vol. 26, 1994, pp. 498515. [9] S. Ramanathan and E.L. Lloyd Scheduling Algorithms for Multihop Radio Networks, IEEE/ACM Trans. Net., vol. 1, no. 2, Apr. 1993, pp. 16677. [10] I. Holyer, The NP-Completeness of Edge Colorings, SIAM J. Comp., vol. 10, 1981, pp. 71820.

BIOGRAPHIES
ANTHONY EPHREMIDES (tony@eng.umd.edu) received his B.S. degree from the National Technical University of Athens (1967), and M.S. (1969) and Ph.D. (1971) degrees from Princeton University, all in electrical engineering. He has been at the University of Maryland since 1971, and currently holds a joint appointment as professor in the Electrical Engineering Department and the Institute of Systems Research (ISR). He is co-founder of the NASA Center for Commercial Development of Space on Hybrid and Satellite Communications Networks established in 1991 at Maryland as an offshoot of the ISR. He was a visiting professor in 1978 at the National Technical University of Athens, Greece, and in 1979 at the Electical Engineering and Computer Science Department of the University of California at Berkeley and INRIA, France. During 19851986 he was on leave at Massachusetts Institute of Technology and the Swiss Federal Institute of Technology, Zurich. He has also been director of the Fairchild Scholars and Doctoral Fellows Program, an academic and research partnership program in satellite communications between Fairchild Industries and the University of Maryland. He has been President of the Information Theory Society of the IEEE (1987), and served on the Board of the IEEE (1989 and 1990). His interests are in the areas of communication theory, communication systems and networks, queuing systems, signal processing, and satellite communications. YALIN EVREN SAGDUYU (sagduyuy@eng.umd.edu) received his B.S. degree from Bogazici University, Turkey, and M.S. degree from the University of Maryland at College Park in 2000 and 2002, respectively, all in electrical engineering. He is currently working toward his Ph.D. degree at the University of Maryland, where he has been a graduate research assistant with ISR since 2000. His research interests include wireless communication, ad hoc and sensor network design, stochastic games, and optimization.

REFERENCES
[1] G. D. Nguyen, J. E. Wieselthier, and A. Ephremides, Multiple-Access for Multiple Destinations in Ad-hoc Networks, Proc. WiOpt 03, Sophia-Antipolis, France, Mar. 2003. [2] G. D. Nguyen, J. E. Wieselthier, and A. Ephremides, Collision-resolution Algorithms for Multiple Destinations in Wireless Networks, Proc. Conf. Info. Sci. and Sys., Baltimore, MD, Mar. 2003. [3] E. Arikan Some Complexity Results about Packet Radio Networks, IEEE Trans. Info. Theory, vol. IT-30, July 1984, pp. 68185. [4] A. Ephremides and T. Truong, Scheduling Broadcasts in Multihop Radio Networks, IEEE Trans. Commun., vol. 38, no. 4, Apr. 1990, pp. 45660. [5] Y.E. Sagduyu and A. Ephremides, Energy-Efficient Collision Resolution in Wireless Ad Hoc Networks, Proc. IEEE INFOCOM, San Francisco, CA, Apr. 2003.

IEEE Wireless Communications December 2004

53

Vous aimerez peut-être aussi