TDMA Scheduling Algorithms for Sensor Networks

Sinem Coleri Ergen and Pravin Varaiya Department of Electrical Engineering and Computer Sciences University of California, Berkeley, CA 94720 Email: {csinem, varaiya}@eecs.berkeley.edu July 2, 2005

Abstract
In developing algorithms for scheduling TDMA transmissions in multi-hop networks, the problem is usually to determine the smallest length conict-free assignment of slots where each link or node is activated at least once. This is based on the assumption that there are many independent point-to-point ows in the network. In sensor networks however often data are transferred from the sensor nodes to a few central data collectors. The scheduling problem is therefore to determine the smallest length conict-free assignment of slots during which the packets generated at each node reach their destination. We show that this problem is NP-complete. We rst propose two centralized heuristic algorithms for solving the problem: One is based on direct scheduling of the nodes, node based scheduling, whereas the other is based on scheduling the levels in the routing tree before scheduling the nodes, level based scheduling. The performance of these algorithms depends on the distribution of the nodes across the levels. We then propose a distributed algorithm based on the distributed coloring of the nodes, that increases the delay by a factor of 10 70 over centralized algorithms for 1000 nodes. We also obtain upper bound for these schedules as a function of the total number of packets generated in the network.

work lifetime. Proposed MAC protocols for sensor networks provide either contention based access or time division multiple access (TDMA). The former, e.g. IEEE 802.11 [10], consume more energy than TDMA protocols because they waste energy in collisions and idle listening. Moreover, they do not give delay guarantees. TDMA protocols are more power efcient since nodes in the network can enter inactive states until their allocated time slots. They also eliminate collisions and bound the delay. For example, the TDMA protocol for a trafc monitoring network described in [6] has a lifetime of 1,200 days compared with ten days using the IEEE 802.11 protocol. The main task in designing a TDMA schedule is to allocate time slots depending on the topology and the node packet generation rates. A proper schedule not only avoids collisions by silencing the interferers of every receiver node in each time slot but also minimizes the number of time slots hence the latency: The larger latency may require a higher data rate (and hence higher energy consumption) to satisfy a deadline [3]. We therefore try to nd a TDMA schedule that minimizes the number of time slots. TDMA algorithms consider either one-hop or multi-hop scheduling. The former are for networks in which the nodes are one hop away from the base station [3, 7], and allocate time slots in the reverse channel depending on allocation request and deadline of the nodes. Because the base station is the common receiver of the transmissions, only one node can transmit in a slot. In some sensor networks however direct transmission from all sensor nodes to the base station may not be feasible nor power efcient [5]. Multi-hop TDMA scheduling is more challenging than onehop scheduling because spatial reuse of a time slot may be possible: More than one node can transmit at the same time slot if their receivers are at non-conicting parts of the network. There are two types of conicts, namely, primary conict and 1

1 Introduction
Wireless sensor networks have been proposed for a wide range of monitoring applications such as trafc and seismic monitoring, and re detection [8]. Such networks consist of a group of nodes, with sensing, signal processing and wireless communication capabilities and limited battery energy. The nodes must quickly report the results to a data collection node or access point. Since the nodes are battery-powered, the medium access control (MAC) protocol is critical in determining net-

secondary conict. A primary conict occurs when a node transmits and receives at the same time slot or receives more than one transmission destined to it at the same time slot. A secondary conict occurs when a node, an intended receiver of a particular transmission, is also within the transmission range of another transmission intended for other nodes. In the context of TDMA, the problem is to determine the smallest length conict-free assignment of slots where each link or node is activated at least once [12]. Previous work on scheduling algorithms focus on either decreasing the length of schedules [12, 9, 2] or distributed implementation [13, 4, 19, 20, 15].

are analyzed in Section 7. Simulations are given in section 8. Section 9 collects some conclusions.

Network and transmission model

We consider a network comprising a single access point (AP) and several sensor nodes that periodically generate data, possibly at different rates, for transfer to the AP. Links are assumed to be bidirectional. This is required for proper functioning of network protocols such as distributed Bellman-Ford algoPrevious scheduling algorithms activating each link or node at rithms [16]. Bidirectionality is achieved if all sensor nodes least once during a TDMA frame is based on the assumption transmit at the same power. Differences in actual transmission that there are many independent point-to-point ows in the net- power due to the hardware differences can be compensated by work. In sensor networks however often data are transferred setting up links based on received signal strength as explained from the sensor nodes to a few central data collectors. In trafc in [6]. monitoring [18], for example, the nodes sense the passage of vehicles at several freeway locations or at an intersection, and The network is represented by a graph G = (V, E). V is transmit the data to the access point on the side of the freeway the set of nodes, including the access point AP as node 1. or intersection. The packets are transferred to the access point N = |V | is the number of nodes in G. The (undirected) edges over the routing tree in multiple hops. The problem therefore E V V are the (transmission) links to be scheduled. The is to determine the smallest length conict-free assignment of graph forms a tree. All trafc is destined for AP , so every data slots during which the packets generated at each node reach packet at a node is forwarded to the nodes parent in the tree rooted at the AP. the access point over the routing tree. To the best of our knowledge, this is the rst work that exploits application-specic characteristics of sensor networks to improve the delay. This problem has not been discussed before in the literature. Since each packet is relayed on the routing path from the originating sensor node to the access point, the problem requires considering precedence relations: If the packet follows the routing path (k, k 1, ..., 2, 1), node j should not be scheduled before node i for that packet if j i. Precedence constrained graphs have been studied in the context of allocating the tasks in the precedence constrained task graphs to the processors in the processor network so that the schedule length is minimized [14]. In this problem however the tasks are already assigned to the processors since a packet at node j has to be processed by node j. The difculty on the other hand comes from the requirement of eliminating primary and secondary conicts in the processor network. A node may interfere with another node, so these nodes should not transmit simultaneously. The interference graph C = (V, I) is assumed known. I V V is the set of edges such that (u, v) I if either u or v can hear each other or one of them can interfere with a signal intended for the other (even if they cannot hear each other). So, if u is transmitting, v should not be scheduled to receive from another node at the same time.

The conict graph corresponding to G = (V, E) and C = (V, I) is called GC = (V, EC). In GC, each node i V corresponds to the link (i, pi ) E where pi is the parent of node i in the routing tree G rooted at AP. EC comprises the edges between node pairs in G that should not transmit at the same time. It is generated by taking into account the primary and secondary conicts described in Section 1. EC contains two kinds of edges. First, if (i, j) E, (i, j) EC, because a The rest of the paper is organized as follows. The transmis- parent node and a child node cannot transmit at the same time. sion and network model is described in Section 2. Section Second, if (i, j) I or (i, j) E and cj is a child of j in G, 3 describes the scheduling problem and proves that it is NP- (i, cj ) EC: Because i and j interfere, if i is transmitting, complete. We then propose two heuristic centralized algo- the child cj of j cannot transmit at the same time because j rithms for solving the problem: The one in Section 4 is based would hear from both i and cj . on direct scheduling of the nodes whereas the other is based A scheduling frame is the time duration that starts when each on scheduling the levels in the routing tree before scheduling node has generated an integer number of packets and ends the nodes as explained in Section 5. Section 6 describes a towhen all these packets have reached AP . It is divided into ken based distributed scheduling algorithm. The algorithms time slots. A slot is long enough to transmit one data packet 2

plus a guard interval to compensate for synchronization errors. A schedule assigns one or more time slots to each edge in G or, equivalently, to each node in GC. A node u may receive a packet from its child v during a time slot assigned to (v, u) E or to node v V since its parent u is already known. We use the following notation. The distance d(u, v) between nodes u and v is the number of edges in the path between them in G; and a node u is at level k if it is at distance k from AP .

GP=(VP,EP) v

AP

GC=(V,EC)

2 v w 4 1 w 2 w 3 w 4

1 v 3 v 1 v G=(V,E,I) 3 v 2 v 4

AP

3 The scheduling problem

v v 1

2 v v 3 4

Figure 1: Transformation from GP = (V P, EP ) to GC = (V, EC) and then to a tree network G = (V, E) (solid lines Theorem 1 The scheduling problem is NP-complete. belong to E) with interference graph C = (V, I) (dashed lines Proof We reduce the NP-complete problem of nding the belong to I). chromatic number of a graph to the scheduling problem.1 Let GP = (V P, EP ) with V P = {v1 , , vN } be an instance of them is interfered by a transmission of the other. Thus GC of a graph whose chromatic number we want to nd. We rst is indeed the conict graph corresponding to the tree graph G construct a conict graph GC = (V, EC). First, GC includes and interference graph C. all the nodes and edges of GP . Next, for each node vi , add another node wi . Then add edges (wi , wj ), (vi , wj ) EC for Consider the minimum schedule length for GC such that each all i, j. Lastly add another node AP and edges (AP, wi ) for node vi , wi , 1 i N , has one packet destined for AP . all i. See gure 1. A packet in wi takes the path (wi , AP ) and a packet in vi takes the path (vi , wi , AP ). Because each wi conicts with the The conict graph GC is such that if wi is active, none of the nodes w , j = i and all nodes v , it takes N slots to transmit j i nodes in V \ {wi } can be active at the same time. Also, if vi the packets generated at level one to AP , independently of the is active, none of the nodes wj or the conicting nodes from rest of the network. Also, when the N packets from level two V P , determined by the edges EP , can be active. arrive at level one, it takes another N slots to forward them to We now construct a tree G = (V, E) and an interference graph AP . C = (V, I) whose conict graph is GC = (V, EC). The Thus to minimize the time to transmit all packets to AP , we edges of the tree are E = {(AP, wi ), (wi , vi ) | 1 i N }. must minimize the time to transmit the packets from level two Because AP is a parent of wi , (wi , AP ) EC for all i; more- to level one. But the conict graph at level two is determined over (wi , wj ) EC for all i, j, because they have the same by the original graph GP , so the minimum scheduling time is parent, AP . And (vi , wi ) EC because wi is the parent of exactly 2N +c, where c is the chromatic number of the original vi . graph GP . 2 Let I consist of edges (vi , AP ) for all i, and (vi , wj ), (vj , wi ), whenever (vi , vj ) EC. Since (vi , AP ) I and (wj , AP ) E, (vi , wj ) EC for all i, j. Lastly, if (vi , wj ) I and (vj , wi ) I, i = j, (vi , vj ) EC because the parent of one
1 The chromatic number of a graph G is the smallest number k such that G is k-colorable. G is k-colorable if its vertices can be colored using k different colors in such a way that adjacent vertices have different colors.

The scheduling problem is difcult because many subsets of non-conicting nodes are candidates for each time slot, and the subset selected for transmission in one slot affects the number of transmissions in the next time slot, as some schedulable nodes may not have any packets to transmit because of the subset selected in the previous slot.

Lemma 1 Assume that node i V has generated gi packets to transmit. The minimum schedule length is at least iV gi . Proof AP can receive at most one packet in each slot, so at least iV gi slots are needed for all packets to reach AP . This gives the lower bound. 2

4 Node Based Scheduling Algorithm

Input: Vc = {2, 3, ..., N }, conict graph GCc = (Vc , ECc ). Output: One color assigned to each node {(2, c2 ), (3, c3 ), ..., (N, cN )} in which ci {1, 2, , ..., M } and M is the number of colors. begin Order the nodes as (n1 , n2 , ..., nN 1 ) for l = 1 to N 1 i=1 while ( j assigned to color i st. (j, nl ) ECc ) i=i+1 assign color i to nl end Figure 2: Assigning one color to each node in the network.

nodes corresponding to the current time slot from the network coloring, additional nodes assigned to other colors are added Any algorithm can be used to color the conict graph GCc as long as the resulting set is non-conicting. The running such that nodes i and j are assigned different colors if (i, j) time of the algorithm is then O(ldmax |V |), where dmax is the ECc . Computing the chromatic number of a graph is NP- maximum degree of a node in GC and l is the total number of complete. Incremental methods appear to be the heuristic slots in the schedule. choice of vertex coloring [13]: Vertices are colored sequentially with the colors chosen in response to colors already as- Two examples are given in Figures 6 and 7. G = (V, E) signed in the vertexs neighborhood. These methods vary in and C = (V, I) are shown on the left with the resulting how the next vertex is selected and how it is assigned a color. GC = (V, EC). The nodes are ordered based on their degrees in GC for the coloring, which are (s3, s2, s4, s5, s6, s7, s1) Figure 2 shows such a heuristic coloring algorithm. At the and (s2,s3,s5,s6,s1,s4) in Figures 6 and 7 respectively. The beginning of the algorithm, the nodes are ordered according resulting schedules are shown in part-(a) of the gures. to some criterion, e.g. non-increasing order of degree since high degree vertices have more color constraints and so are more likely to require an additional color if inserted late. This 5 Level Based Scheduling Algorithm algorithm assigns a slot to node i in O(i) steps, so the running 2 time of this algorithm is O(|V | ). The level based scheduling algorithm has three parts. In the rst part, we obtain a linear network GL = (V L, EL) with 4.2 Scheduling the Network interference graph CL = (V L, IL) resulting in conict graph GCL = (V L, ECL) corresponding to the original network. A superslot in node based scheduling algorithm is a collection In the second part, we color this linear network. In the third of consecutive time slots such that each node with at least one part, we schedule the links in the original network, (u, v) E, packet at the beginning of the superslot transmits at least one based on the coloring of the linear network. packet during the superslot. Because two nodes assigned the same color can transmit at the same time, the number of slots in a superslot is at most equal to the total number of colors 5.1 The linear network used for coloring the network. If the original tree network has depth N , the linear network The algorithm is given in Figure 3. After determining the GL = (V L, EL) has nodes V L = {v1 , , vN } with node 4

4.1

Coloring the network

Input: Graph G = (V, E) with conict graph GC = (V, EC), color assignment of the nodes Vc using M colors. Output: Transmission schedule for nodes of G. begin while (at least one packet has not reached AP ) for s = 1 to M sets = set of nodes corresponding to color s with at least one packet T = sets if T = setos = set of nodes not corresponding to color s with at least one packet for each node k setos if (k, j) EC j T / T = T {k} assign current slot to set T update the place of the packets end Figure 3: Node Based Scheduling Algorithm.

Input: (V, E, I, EC). Output: (V L, EL, IL, ECL). begin add node v1 to V L l=2 while l levelOf T ree add node vl to V L add edge (vl1 , vl ) to EL If (u, v) I(EC) with u at level l and v at level j satisfying j < l add edge (vj , vl ) to IL(ECL) l++ end Figure 4: Algorithm to nd linear network corresponding to original network. level i and the other from level j, can transmit at the same time.

A superslot in level based scheduling algorithm is a collection of consecutive time slots such that each level of the tree with at least one packet at the beginning of the superslot forwards vl corresponding to all nodes at level l in the original network at least one packet to the lower level during the superslot. Beand edges (vi , vi+1 ) EL for 1 i < N . The interference cause two nodes at different levels assigned the same color can graph CL = (V L, IL) includes edge (vj , vl ) if there is an in- transmit at the same time, the number of slots in a superslot is terference edge between a node at level j and any node at level at most equal to the total number of colors used for coloring l in the original network for j, l 1. The resulting conict the linear network. graph GCL = (V L, ECL) thus includes edge (vj , vl ) if the The algorithm is given in Figure 5. After determining the levtransmissions of a node at level j and a node at level l con- els corresponding to the current time slot from the linear netict in the original network. The algorithm in gure 4 nds work coloring, a nonconicting set of nodes at these levels that EL,CL and ECL. Its running time is O(|V |2 ). have packets to transmit are selected for transmission. Additional nodes from other levels are then added as long as the resulting set is non-conicting. The running time of the algo5.2 Coloring the linear network rithm is O(ldmax |V |), where dmax is the maximum degree of a node in GC and l is the total number of slots in the schedule. Any coloring algorithm can be used to color the conict graph of the linear network GCL = (V L, ECL). The algorithm Two examples are given in Figures 6 and 7. G = (V, E) given in Figure 2 can be used for this purpose with Vc = V L and C = (V, I) are shown on the left with the resulting and GCc = GCL as input, and one color assigned to each GCL = (V L, ECL). The levels are in increasing order for coloring. The ordering does not affect the number of colors node in V L and the number of colors, M , as output. used in linear network nor the schedule length for these examples. The resulting schedules are shown in part-(b) of the gures. 5.3 Scheduling the original network Figure 6 illustrates a topology where level based scheduling If nodes vi , vj in the linear network are assigned the same performs better than node based scheduling whereas Figure color, they do not interfere. By construction of the linear net- 7 illustrates a network where node based scheduling outperwork any two nodes in the original network, one chosen from forms level based scheduling. Figure 6 demonstrates the ad-

vantage of level based scheduling in balancing the movement of packets across the network in a network of higher density of the packets at high levels. In topologies of equal density of the packets across the network or higher packet density at low levels, giving equal chance to the nodes balances the movement of packets as shown in Figure 7.
G=(V E) , AP GC=(V EC ) , AP GCL =(VL, EC L ) DFStree AP v1 v2

s1 s2 s3 s5 s6 s7

s1 s2 s3 s5 s6 s7

s1 s2 s3 s5 s6 s7

v3 v4

s4

s4

s4

Input: Graph G = (V, E) with conict graph GC = (V, EC), color assignment of the corresponding linear network GCL using M colors. Output: Transmission schedule for nodes of G. begin while (at least one packet has not reached AP ) for s = 1 to M sets = set of levels corresponding to color s T = for j = 1 to |sets | T = T {a nonconicting set of nodes from level sets (j) with at least one packet} if T = setos = set of levels not corresponding to color s for each node k belonging to a level in setos if (k, j) EC j T / T = T {k} assign current slot to set T update the place of the packets end Figure 5: Level Based Scheduling Algorithm.

a) superslot #:1 slot #: 1 2 s2 3 4 5 6 s7 2 7 s3 8 s2 9 s1 3 10 s3 11 s2 12 s1 4 13 s3 14 s2 15 s1 5 16 s3 17 s2 18 s1 6 19 s2 20 s1

schedule: s3

s4,s1 s5,s1 s6

b) superslot #:1 slot #: 1 2 3 s3 2 4 5 6 s3 3 7 8 9 s3 4 10 11 12 s3 5 13 s1 14 s2 15 s3 6 16 s1 17 s2 7 18 s1

schedule: s1,s4 s2 c) superslot #:1 slot #: 1 2

s1,s5 s2

s1,s6 s2

s1,s7 s2

2 3 s3 4 5 6 s7 7 8 s2 9 s3 10 11 12 -

3 13 s1 14 s2 15 s3 16 17 18 -

4 19 s1 20 s2 21 s3 22 23 24 -

schedule: s1,s4 s2

s1,s5 s6

superslot #:5 slot #: 25 26 27 s2 s3 28 29 30 -

6 31 s1 32 s2 33 34 35 36 -

7 37 s1 38 39 40 41 42 -

schedule: s1

Figure 6: An example network where level based scheduling performs better than node based scheduling. a) Schedule for node based scheduling algorithm. b) Schedule for level based scheduling algorithm. c)Schedule for distributed scheduling algorithm.

Distributed Implementation

The node based and level based scheduling algorithms described above require complete topology information, and there are two options for implementation. The rst option is to send the topology information to a central controller, which then performs the slot assignment and sends it back to the nodes in the network. The second option is that each node learns the entire network topology and executes the algorithm independently to produce identical schedules. Both options may require a lot of communication among the nodes, and may become inappropriate for large networks. Distributed algorithms, in which the schedules of the nodes are generated based on the local topology information of the nodes, are preferred in large networks due to their scalability. However, it is very hard to obtain a distributed version of node

G=(V E, I) , AP

GC=(V EC ) , AP

GCL =(VL, EC L )

DFStree AP

s1 s4

s2 s5

s3 s6

s1 s4

s2 s5

s3 s6

v1 v2

s1 s4

s2 s5

s3 s6

edge interference edge a) superslot #:1 slot #: 1 2 3 4 2 5 6 s3

each node picks as many of the remaining colors as it can for transmission. At both stages, the nodes send this information to their one-hop and two-hop neighbors in Gu so that all their interferers in GC learn about the assignment. The DFS traversal starts with a TOKEN message generated at the AP. Upon receipt of the token, the node performs the color assignment and then sends this information to its one-hop and two-hop neighbors in Gu . It then sends the token to each of its neighbors in G who have not received the token yet. Once it nds that all its neighbors have received the token, it sends the token back to its parent, which is the node from which it receives the token for the rst time. At the end of the traversal, the token carries the information of the number of colors used in the network back to the AP. This distributed algorithm turns out to be the distributed version of the color assignment algorithm shown in Figure 8. Once the colors are assigned to each node, the nodes only transmit in the assigned timeslots assigned to these colors if they have a packet to transmit. The total number of token transmissions is O(|E|) at each stage and the total number of transmissions for distributing the color assignments is O(dmax |V |), in which dmax is the maximum degree of the nodes in Gu . Two examples are given in Figures 6 and 7. The coloring of the nodes are shown on the top right of the gures. The colors of the nodes are assigned at the rst stage whereas the colors of the small circles next to the nodes are assigned at the second stage. The DFS traversal order are (s1 , s2 , s3 , s4 , s5 , s6 , s7 ) and (s1 , s4 , s2 , s5 , s3 , s6 ) in Figures 6 and 7 respectively. The resulting schedules are shown in part-(c) of the gures.

schedule: s2,s4 s3

s1,s5 s6,s1 s2

b) superslot #:1 slot #: 1 2 2 3 3 4 s3 4 5 s1 5 6 s2 6 7 s3

schedule: s1,s5 s4,s6 s2 c) superslot #:1 slot #: 1 2 3

2 4 5 6 s2 s3 -

schedule: s1,s5 s2,s4 s3

s6,s1 -

Figure 7: An example network where node based scheduling performs better than level based scheduling. a) Schedule for node based scheduling algorithm. b) Schedule for level based scheduling algorithm. c)Schedule for distributed scheduling algorithm. based and level based scheduling algorithms. The main reason is that these algorithms check whether the nodes that are potential transmitters in the current slot have any packets and skip that slot otherwise. Another reason is that the algorithms schedule the nodes in other colors if some of the nodes that are potential candidates for the current slot do not have any packets. These however may not be performed in their distributed version since the nodes cannot know how many packets their interferers have due to the lack of knowledge of global topology information. To get an idea of the performance of a distributed algorithm, we propose a simple algorithm based on the distributed coloring of the network similar to the one described in [13]. Notice that the conicting nodes, the nodes that have edge between them in GC, are either one hop or two hops away from each other in the graph Gu = (V, E I). Assume the nodes i and j can transmit to each other if (i, j) E I, e.g. this can be performed by increasing the transmission range of the nodes [6]. We also assume that all the transmissions during the generation of the schedules are successful, which can be guaranteed by an acknowledgement. The nodes rst learn about all of their one hop and two hop neighbors in Gu and their parents so that they can determine their interferers in GC.

Analysis of the Algorithms

We consider four cases: Case 1: The tree graph G = (V, E) is linear, that is each node u V has at most one child. The interference graph C = (V, I) is such that I = .

Case 2: The tree graph G = (V, E) is general. The interference graph C = (V, I) satises the ancestor property, that is, there do not exist u, v, b such that (u, v) I and |d(u, b) d(v, b)| > 1. This represents the case where shortest Similar to the algorithm described in [13], the color assignpath routing is used with the cost of each path being equal to ment is performed in two stages. During the rst stage of the the number of nodes on that path and only nodes that can hear algorithm, each node picks one slot for transmission in the oreach other can interfere, which is the assumption of previously der of the traversal of the depth rst search (DFS) [11] of the proposed TDMA scheduling algorithms. graph G. In the second stage, the DFS is repeated and now

Input: V = {1, 2, ..., N }, G = (V, E), Gu = (V, E I), conict graph GCc = (Vc , ECc ). Output: Color assignment of the nodes in Vc such that each color corresponds to a maximal nonconicting set in GCc . begin Order the nodes as (n1 , n2 , ..., nN ) in DFS traversal of G for l = 1 to N if nl = 1 i=1 while ( j assigned to color i st. (j, nl ) ECc ) i=i+1 assign color i to nl M is maximum assigned color for l = 1 to N if nl = 1 for i = 1 to M T is the set of nodes assigned to color i if (no node j T st. (j, nl ) ECc ) add color i to the color set of nl end Figure 8: Distributed Scheduling Algorithm. Case 3: The tree graph G = (V, E) is general and the interference graph C = (V, I) is such that the maximum difference between the levels of two interfering nodes is K.

At the beginning of the frame, each node has exactly one packet. In the rst superslot, one packet is transmitted from any level to the next lower level. Because each node is a parent of exactly one node except for the node at the highest level |V | 1, it also receives one packet during the superslot. Thus, at the end of the rst superslot, each node at level less than |V | 1 has exactly one packet to transmit, the node at level |V | 1 has no packet, and each node has transmitted exactly one packet during the superslot. This means that at the end of the rst superslot, each packet has moved by one hop and one packet has reached the nal destination AP . In the same way, at the beginning of the second superslot, each node at level less than |V | 1 has one packet to transmit, and at the end of the second superslot, each packet has moved by one more hop, there are no more packets at levels greater than or equal to |V | 2 and the node AP has received exactly one packet. Continuing in this manner, at the end of (|V | 1) superslots, all packets will have reached the nal destination AP . The maximum number of time slots in each frame is at most the product of the maximum number of slots in each superslot and the maximum number of superslots necessary for all packets to reach the destination AP , namely 3(|V | 1). Case 2. Because the interference graph of the tree network satises the ancestor property, the corresponding linear tree interference graph CL satises IL = . It can therefore be colored optimally with 3 colors.

First assume that we select exactly one node to transmit from each level (of the original tree graph G = (V, E)) corresponding to the color of the slot. At the beginning of the frame, each node has one packet. In the rst superslot, one packet is transmitted from each level to the next lower level. Except at the Case 4: The tree graph G = (V, E) and the interference graph highest level, each level receives one packet. Therefore, one C = (V, I) are both general. packet has moved one hop closer to the AP at each level, one Theorem 2 Assume that each node has one packet to transmit. packet from level one has reached AP , and nodes at the level For level based scheduling algorithm, in cases 1 and 2 the of the depth of the tree may have no more packets. maximum length of the frame is 3|V | 3 time slots; in case At the end of the second superslot, the number of packets 3 it is (K + 2)(|V | 1); and in case 4 it is (|V | 1), in transmitted from one level to one lower level is again one exwhich is the number of colors used in the linear network cept, possibly, for level depth. Each level less than depth 1 corresponding to G and C. has one packet to transmit, while nodes at levels depth or Proof Case 1. If the tree graph G is linear and the interference graph C satises I = , the corresponding linear tree interference graph CL also satises IL = . It is easy to see that this linear tree can be colored optimally with three colors when the number of levels is more than two. The colors are assigned in a round robin fashion starting with the node at level 1. depth 1 may have exhausted all packets. Continuing in this manner, by the end of i-th superslot, there are no more packets above some threshold level, and there is at least one packet at levels lower than this threshold. Since each level below the threshold is guaranteed to have a packet, and all levels with at least one packet can transmit once in each superslot, one packet reaches AP in each superslot. Therefore, the number of superslots required for all packets to reach AP 8

scheduling, schedules only one node containing at least one packet from each level of the routing tree G = (V, E) rooted at the AP and does not schedule any node if that level does not The scheduling algorithm allows a subset of non-conicting contain any packet. By the same reasoning as in the proof of nodes (instead of a single node) at each level to transmit so the Theorem 2, the number of superslots required for all packets resulting frame length will also be at most 3(|V | 1). to reach AP is the number of packets in the network, which is iV gi . iV gi is also the maximum number of superslots Case 3. The worst case is when there is an interfering edge between a node at level j and every node at level i with |i required for all packets to reach AP in node based scheduling. j| K. The corresponding linear graph can be colored by The maximum number of slots in each superslot is the number K + 2 colors in that case. Assign color 1 to v1 . The color of colors. The result follows. of the nodes {v2 , , vK+2 } cannot be 1. Assign the smallest 2 color, 2, to node v2 . The color of {v3 , , vK+3 } cannot be 2. Assign the smallest color, 3, to v3 . Continuing in this way, Remark The chromatic number of a graph GC is less than or vK+2 is assigned color K + 2. Node vK+3 is assigned color equal to 1 + degmax [17], where degmax is the maximum de1, since its color is restricted not to be 2, , K + 2. Thus, gree of the nodes in GC. The maximum length of the schedule the algorithm colors this network with K + 2 colors in a round in node based and level based scheduling algorithms is thererobin fashion with color 1 assigned to v1 . The interference fore (1 + degmax )iV gi , in which degmax is the maximum graph of any other network is a subgraph of this worst case. degree of the nodes in GC and GCL respectively. Since the minimum schedule length is iV gi as shown in Lemma 1, The same reasoning as in Case 2 now indicates that at least the worst case ratio of the length of the frame to the optimal one packet reaches AP in each superslot so the number of length is 1 + deg max . superslots needed is at most |V | 1. Hence the frame length is at most (K + 2)(|V | 1) time slots. Lemma 2 Assume that node i V has generated gi packets to transmit. For distributed scheduling algorithm, the maximum Case 4. The number of superslots required for all packets to length of the frame is iV gi , in which is the number of reach AP is the number of packets in the network, which is colors used in the conict graph GC. |V | 1. The maximum number of slots in each superslot is the number of colors, . The upper bound on the frame length Proof The proof is the same as that of Theorem 4. 2 is then (|V | 1). 2 Theorem 3 Assume that node i V has generated gi packets to transmit. For level based scheduling algorithm, in cases 1 8 Simulation and 2 the maximum length of the frame is 3iV gi time slots; in case 3 it is (K + 2)iV gi ; and in case 4 it is iV gi , The goal of the simulations is to compare the delay perforin which is the number of colors used in the linear network mance of the centralized node based and level based schedulcorresponding to G and C. ing algorithms, and the distributed algorithm. Proof The proof is similar to that of Theorem 2. The number of superslots required for all packets to reach AP is the number of packets in the network, which is iV gi . The maximum number of slots in each superslot is the number of colors. In the simulations, 1000 nodes are randomly distributed in a circular area of radius 100 units. The density of the nodes is 1 inside the radius 100 and 2 between the radius 100 and 2 2 100 units. The transmission range, denoted rs , is chosen to be 2 slightly larger than the threshold necessary for network connectivity [1]. Theorem 4 Assume that node i V has generated gi packets to transmit. For node based scheduling algorithm, the maxi- The results discussed below are averages of the performance mum length of the frame is iV gi , in which is the number of ten different random congurations. Shortest path routing is used to construct the routing tree rooted at the AP, which of colors used in the conict graph GC. corresponds to G = (V, E) in Section 2. The interferers of Proof The proof is similar to that of Theorem 2. In node based the nodes that form C = (V, I) are the nodes that are inside a scheduling, during each superslot, each node is given at least larger range rm , rm rs , of each node other than its parent one chance to transmit. and children in the routing tree G. In the coloring part of node and level based scheduling algorithms, the nodes are ordered Let us assume that another algorithm, namely node-level based

is |V | 1. Since there are three slots in each superslot, the maximum frame length is again 3(|V | 1).

in non-increasing order of degree since high degree vertices have more color constraints and so are more likely to require an additional color if inserted late. Figure 9 shows the delay of node based and level based algorithms as a function of 1 ratio, for the case rm = 2. As ex2 rs pected from the examples in Figures 6 and 7, the level based algorithm performs better for low 1 ratios whereas node based 2 1 schedule performs better for high 2 ratios. Figure 10 shows that this is true up to a certain value of rm ratio. Node based rs scheduling algorithm performs better at high rm ratios. rs
2500

70 range whereas the ratio of the number of colors used in the distributed algorithm to that of centralized algorithm is in the 1 1.3 range. This suggests the basic disadvantage of distributed algorithms to be the scheduling of the nodes that do not have any packet in a specic slot and the elimination of other nodes as a result. The delay ratio therefore increases as rm rs ratio increases due to the increase in the number of colors used in the original network.
10
2

node based level based

2400

2300

performance ratio

10

2200

delay (slot)

2100

delay ratio,1/2=1/9 delay ratio,1/2=9 color ratio,1/2=1/9 color ratio, / =9

2000

1900

1800 10 1700
0

1.5

2.5

3.5

4.5

rm/rs

1600

1500

10

1/2

Figure 11: Performance ratio of the distributed algorithm to the centralized node based scheduling algorithm in delay and number of colors used in GC.

Figure 9: Comparison of the delay of node based and level based scheduling algorithms for different 1 ratios and rm = 2 rs 2.
4500

Conclusion

4000

node based,1/2=1/9 level based,1/2=1/9 node based, / =9 1 2 level based, / =9

3500

3000

2500

2000

1500

The common scheduling problem in multi-hop networks employing a TDMA MAC protocol is to determine the smallest length conict-free assignment of slots where each link or node is activated at least once. This is based on the assumption that there are many independent point-to-point ows in the network. In sensor networks where data are often transferred from the sensor nodes to a few central data collectors, the problem is to determine the smallest length conict-free assignment of slots during which the packets generated at each node reach their destination. This optimization problem is shown to be NP-complete.
2.5 3 3.5 4 4.5 5

delay (slot)

1000

1.5

We propose two centralized heuristic algorithms for solving the problem: node based scheduling and level based schedulFigure 10: Comparison of the delay of node based and level ing. In node based scheduling, the schedule is obtained based on the coloring of the original network. The nodes of the color based scheduling algorithms for different rm ratios. rs corresponding to each slot with at least one packet are choFigure 11 shows the performance of distributed algorithm in sen rst and additional nodes are added afterwards. In level terms of the ratio of its delay to that of the centralized node based scheduling on the other hand the original network is based scheduling algorithm. The delay ratio is in the 10 rst transformed to a linear network where each node correrm/rs

10

sponds to a level in the original network. The schedule of the original network is then obtained based on the coloring of the linear network. This scheduling algorithm schedules a non-conicting set of nodes corresponding to each level of the color for the current slot and then schedules additional nodes if possible. The movement of the packets across the network is balanced much better in level based scheduling for topologies of higher density of the packets further away from the common sink whereas giving equal chance to the nodes in node based scheduling performs better in topologies of equal density of the packets across the network or higher packet density at low levels.

[5] S. C. Ergen and P. Varaiya. On multi-hop routing for energy efciency. IEEE Communications Letters, to appear in 2005. [6] S. C. Ergen and P. Varaiya. Pedamacs: Power efcient and delay aware medium access protocol for sensor networks. IEEE Transactions on Mobile Computing, to appear in 2005. [7] H. Fattah and C. Leung. Energy-efcient packet transmission over a wireless link. IEEE Wireless Communications, 9(5):7683, 2002.

[8] C. Sharp J. Polastre, R. Szewczyk and D. Culler. The We also propose a simple token based distributed algorithm mote revolution: Low power wireless sensor network deto understand the performance of distributed algorithms comvices. In Hot Chips 16: A Symposium on High Perforpared to centralized ones. The distributed algorithm is based mance Chips, August 2004. on a two-stage coloring algorithm at the end of which nodes assigned the same color form a maximal nonconicting set. [9] C. Y. Ngo and V. O. K. Li. Centralized broadcast schedulWe observe that the delay in distributed algorithm increases ing in packet radio networks via genetic-x algorithms. by a factor of 10 70 over centralized algorithms for 1000 IEEE Transactions on Communications, 51(9):1439 nodes although the number of colors used in coloring the net1441, 2003. work is almost the same. This suggests the basic disadvantage of distributed algorithms to be the scheduling of the nodes that [10] LAN-MAN Standards Committee of the IEEE Computer Society. Wireless LAN medium access condo not have any packet and the elimination of the scheduling trol(MAC) and physical layer(PHY) specication. IEEE, of other nodes as a result. This is hard to avoid in a distribNew York,NY, 1997. uted fashion since the global topology information is required to know whether the interfering nodes have any packets. Dis- [11] T. L. Magnanti R. K. Ahuja and J. B. Orlin. Network tributed scheduling algorithms that improve upon this token Flows. Prentice Hall, Inc, New York, NY, 1993. based algorithm in the context of sensor networks is an interesting research direction. [12] S. Ramanathan and E. L. Lloyd. Scheduling algorithms for multihop radio networks. IEEE/ACM Transactions on Networking, 1(2):166177, 1993.

References
[1] S. B. Wicker B. Krishnamachari and B. Bejar. Phase transition phenomena in wireless ad-hoc networks. In Symposium on Ad-Hoc Wireless Networks, IEEE GLOBECOM 2001, November 2001.

[13] R. Ramaswami and K. K. Parhi. Distributed scheduling of broadcasts in a radio network. In INFOCOM 1989, pages 497504, April 1989. [14] P. Kumar S. Bansal and K. Singh. An improved duplication strategy for scheduling precedence constrained graphs in multiprocessor systems. IEEE Transactions on Parallel and Distributed Systems, 14(6):533544, 2003.

[2] G. Chakraborty. Genetic algorithm to solve optimum tdma transmission schedule in broadcast packet radio networks. IEEE Transactions on Communications, [15] M. Dawande S. Gandham and R. Prakash. Link scheduling in sensor networks: Distributed edge coloring revis52(5):765777, 2004. ited. In IEEE INFOCOM 2005, March 2005. [3] B. Prabhakar E. Uysal-Biyikoglu and A. El Gamal. Energy-efcient packet transmission over a wireless link. [16] R. S. Sreenivas S. Narayanaswamy, V. Kawadia and P. R. Kumar. Power control in ad-hoc networks: Theory, arIEEE/ACM Transactions on Networking, 10(12):487 chitecture, algorithm and implementation of the compow 499, 2002. protocol. In Proceedings of European Wireless 2002, [4] A. Ephremides and T. V. Truong. Scheduling broadcasts February 2002. in multihop radio networks. IEEE Transactions on Communications, 38(4):456460, 1990. 11

[17] D.P. Jacobs S. T. Hedetniemi and P. K. Srimani. Fault tolerant distributed coloring algorithms that stabilize in linear time. In International Parallel and Distributed Processing Symposium (IPDPS) 2002, April 2002. [18] B. Dundar S. Ganesh C.W. Tan S. Y. Cheung, S. Coleri and P. Varaiya. Trafc measurement and vehicle classication with a single magnetic sensor. 84th Annual Meeting, Transportation Research Board, January 2005, chosen to be published in Journal of Transportation Research Board, to appear in 2005. [19] B. Tavli and W. B. Heinzelman. Mh-trace: Multihop time reservation using aadaptive control for energy efciency. IEEE Journal on Selected Areas in Communications, 22(5):942953, 2004. [20] Z. Wu and D. Raychaudhuri. D-lsma: Distributed link scheduling multiple access protocol for qos in ad-hoc networks. In IEEE GLOBECOM 2004, pages 1670 1675, November 2004.

12