PROCEEDINGS OF THE IEEE, VOL. 91, NO. 8, AUGUST 2003

---

TAEK JIN KWON, MEMBER, IEEE, MARIO GERLA, FELLOW, IEEE, VIJAY K. VARMA, SENIOR MEMBER, IEEE, MELBOURNE BARTON, SENIOR MEMBER, IEEE, AND T. RUSSELL HSING, FELLOW, IEEE

Contributed Paper

    High capacity real-time data communications in sensor networks usually require multihop routing and ad hoc routing protocols. Unfortunately, ad hoc routing protocols usually do not scale well and cannot handle dense situations efficiently. These two issues—scalability and density—are the major limitations when we apply ad hoc routing schemes to sensor networks. Passive clustering (PC) classifies ad hoc/sensor nodes into critical and noncritical nodes without any extra transmission. By 2-b piggybacking and monitoring user traffic (e.g., data polling requests from a sink), PC deploys the clustering structure “for free.� Moreover, PC makes even the first flooding as efficient as all subsequent floodings (i.e., no initialization overhead). PC introduces many benefits, including efficient flooding and density adaptation. As a result, PC reduces control overhead of ad hoc routing protocols significantly and, as a consequence, enables ad hoc routing in large, dense sensor networks. The resulting structure can be utilized in cluster-based ad hoc network/sensor networking as well as for active node selection.
  

    Keywords—Ad hoc/sensor networks, clustering, density adaptation, on-demand efficient flooding.

    Manuscript received November 16, 2002; revised March 26, 2003.
    T. J. Kwon, V. K. Varma, and M. Barton are with Telcordia Technologies, Red Bank, NJ 07701 USA (e-mail: tkwon@research.telcordia.com; vvarma@research.telcordia.com; mbarton@research.telcordia.com).
    M. Gerla is with the Department of Computer Sciences, University of California in Los Angeles, Los Angeles, CA 90095 USA (e-mail: gerla@cs.ucla.edu).
    T. R. Hsing is with Telcordia Technologies, Morristown, NJ 07960 USA.
Digital Object Identifier: 10.1109/JPROC.2003.814920

I.  INTRODUCTION
II.  FLOODING IN ON-DEMAND ROUTING
III.  EFFICIENT FLOODING
IV.  PASSIVE CLUSTERING
    A.  Protocol Overview
    B.  Operational Description
V.  SELECTIVE GATEWAY PASSIVE CLUSTERING
    A.  Gateway Selection Heuristic
    B.  Flooding Improvement On the Fly
    C.  Properties of the Passive Cluster Solution
VI.  SIMULATION STUDY
    A.  Flooding Delivery Experiments
    B.  Comparison with Active Clustering
    C.  No Mobility—Dynamic Traffic
    D.  Adaptive Density
VII.  CONCLUSION
REFERENCES

I.  INTRODUCTION

    The differences between sensor networks and ad hoc networks are summarized in [1]. The major limitations that prohibit ad hoc network routing protocols from being applied in sensor networks are the limited scalability and the inability to adapt to high-density sensor distributions. In fact, the number of nodes in sensor networks can be several orders of magnitude higher than that in ad hoc networks. Sensor nodes are densely deployed as well. These factors contribute to a drastic increase in the control overhead (e.g., route discovery, topology updates, neighbor finding, etc.) of ad hoc network routing protocols.
    There are two dominant strategies in ad hoc network routing: proactive routing and on-demand routing. Proactive routing mechanisms like optimized link state routing (OLSR) [2] and topology dissemination based on reverse-path forwarding (TBRPF) [3] discover and maintain routing tables without considering the usage of routing information. Their basic mechanism is similar to the routing mechanisms in wired network—they are link-state routing schemes [4] with ad hoc network enhancements to cope with faster topology changes and wireless environment characteristics. With periodic exchanges of hello messages (neighbor learning) and topology (link) advertisements (by flooding), the complete routing information is maintained. The routing activities continue even when the network carried no user data traffic. The control overhead of proactive routing is dependent on the number of nodes in the system. For this reason, there is an upper bound in the number of nodes unless a well-defined hierarchy is introduced. On-demand routing schemes like ad hoc network on-demand distance vector (AODV) [5] and dynamic source routing (DSR) [6] are routing strategies unique to ad hoc networks. Instead of calculating routes in the background all the time, routes are found and maintained only when they are needed. As a result, there is a route discovery latency delay in this category if the requested route is not readily available. The control overhead of on-demand routing increases with the number of active communication pairs.
    In sensor networks with a limited number of active communication pairs (e.g., intrusion detection), an on-demand routing strategy offers several advantages over proactive routing. First, it does not maintain unnecessary routing entries. Proactive routing maintains as many routing entries as the number of nodes in the network, which can be tens of thousands even though only a few active routes are needed. This introduces significant storage as well as processing overhead on each of the sensor nodes. The control overhead of on-demand routing in this case is much less than proactive routing. Moreover, lower control overhead extends the life of sensors and reduces interference of the shared media. Finally, it is also helpful for stealth operations—there is no transmission unless data transmission is required, whereas there are continuous hello/topology broadcastings in a proactive routing mechanism. For all these reasons, proactive routing is not appropriate for large ad hoc network/sensor networks.
    This paper presents a novel efficient flooding method suitable for on-demand routing protocols. In Section II, we briefly go over the functionality of flooding in on-demand routing protocols. Section III surveys efficient flooding mechanism in general. We introduce passive clustering (PC) in Section IV and discuss selective gateway PC in Section V. Simulation studies and the relative advantages and limitations of the proposed mechanism are discussed in Section VI. Section VII provides the conclusions of this work.

II.  FLOODING IN ON-DEMAND ROUTING

    Flooding is a packet delivery process that delivers a packet to all the connected nodes in the network. A packet flood usually requires at least one relay (broadcast) by each node. On-demand routing and multicasting protocols utilize flooding as a route discovery mechanism. Since in on-demand routing there is no prior underlying routing or topology information that can guide a packet to its destination, a path search query must be flooded to the entire network, or at least through a certain section (scope) of it. Once the path search query packet reaches a destination by flooding, the destination can report a path to the source as a reverse path (dotted arrows in Fig. 1) through which the search packet came. In case of multiple search packets, the destination selects the best path and reports a complete route to the source. In case there are asymmetric links, the destination must report the route to the source with another flooding. As an example, AODV uses scoped flooding to find a route.

Fig. 1. Scoped flooding in AODV—“S� transmits a route-request with . Only the striped nodes (direct neighbors of a source S) relay route search flooding packet once. The TTL value of the packet is decremented by one as the intermediate nodes rebroadcast the packet. Flooding packets with TTL higher than one are relayed (rebroadcast.) The destination replies and traces the route back to the source. If the destination were located further than two hops, then the source had to reinitiate the route search with a bigger TTL value.

    In AODV, by tagging “time to live (TTL)� on each route-request flooding packet, a source gradually enlarges flood search diameters until it finds a destination or TTL reaches the threshold. Once the diameter hits the threshold, AODV uses complete flooding (no scope limitation). i.e., it covers the entire network by setting the TTL value of the route request to the maximum network diameter. Fig. 1 illustrates scoped flooding. On the other hand, another popular on-demand routing, DSR depends on complete flooding (flooding relay by all of connected nodes) if a source cannot find a path to destination in a single hop. If the communication patterns are “local� (within certain threshold), incremental scoped flooding is effective. On the other hand, if destinations are typically located many hops away, it is actually wasteful to run incremental scoped flooding.

Fig. 2. MPR in OLSR—the cross-marked node is an MPR for S. Whenever S initiates link-advertising flooding, only MPR nodes relay the flooding packet. To maintain MPR, every node participates in neighbor list advertisements (hello messaging) and MPR maintenance processes. S and D can be MPR nodes for the cross-marked node. Each node in OLSR maintains its own set of MPR nodes.

    Each node in AODV identifies a flooding packet with a combination of the source ID and the sequence number carried by the flooding packet. It relays a flooding packet once—duplicates (received from other neighbors) are discarded immediately. In DSR, each flooding packet carries the list of intermediate relay node IDs in order to avoid loops and endless duplications. Each node in DSR drops the flooding packet if it saw it before (a match of its ID list.) Flooding generates replicated packets, namely, at least one replica for each node. Thus, flooding overhead is significant in a large and dense sensor network.

III.  EFFICIENT FLOODING

    Flooding in ad hoc networks is used to find a feasible route (as illustrated in previous on-demand routing examples) or to advertise routing information (as in some of the link state routing protocols). If the network is dense, it is not necessary for every node to relay the flood search packet. In fact, it may suffice to use only a subset of nodes as relays. Reducing unnecessary flooding relay may save control overhead greatly when the connectivity is high (say, hundreds of neighbors). Efficient flooding (relaying flood packets by a subset of nodes) is rather a common practice in proactive routing protocols where neighbor information is gathered and propagated as part of routing. In OLSR, for example, by utilizing neighbor information (e.g., the two-hop neighbor information is collected by exchanging neighbor lists), a node selects only a small number of direct neighbors who will relay the packet to the complete set of two-hop neighbor. (e.g., crossed node in Fig. 2.) However, this procedure requires not only continuous hello messaging but also extra signaling to designate the neighbors to relay flooding packets (in Fig. 2, S nominates the crossed-node as its relaying node). Note that every node in the system has to perform the same tasks (neighbor learning and nominations). Multipoint relay (MPR) [2] implements the idea in OLSR. Hello overhead increases quickly as the number of neighbor grows—the lengths of the hello packets get longer and the possibility of collision becomes higher. For this reason, none of the on-demand routing protocols promotes hello messaging or efficient flooding (à la MPR), as it would introduce additional stress into network.
    There are many ways to reduce the number of forwarding participants. All of the approaches concern selecting the dominant set, i.e., a minimal subset of forwarding nodes which is sufficient to deliver the flooding packet to every other node in the system. There are two basic approaches for selecting the dominant set: without and with a clustering structure.
    The first approach (no clustering) includes the building of a source tree with the maximal number of leaf nodes [3], [7], [8] or the building of a well-covered mesh [9][10][11]. By excluding leaf nodes from forwarding participation, the method can improve flooding efficiency. To build such a source tree, two-hop connectivity information is necessary. To collect the required information, at least two complete floodings from a source are necessary. The first flooding (which can be replaced with well-coordinated hello messages) is to learn the one-hop neighbors. The second flooding is to report the direct (one-hop) neighbor lists. By collecting the complete neighbor lists of all of its neighbors, a node can construct the two-hop connectivity, i.e., the list of nodes that are two hops away. From this list, each node selects the minimum set of one-hop neighbors which cover all the downstream two-hop neighbors. This problem can be reduced to the well-known “set-cover� problem (NP-complete). Starting from a source and applying this procedure recursively, one generates the nonleaf nodes of a minimal flooding tree. Span [9] and geographic adaptive fidelity (GAF) [11] build their dominant set as a well-covered mesh. Span selects nodes that are potentially on critical paths as coordinators, i.e., members of a dominant set. GAF partitions the region with a grid such that any nodes in neighboring cells can communicate each other; one node per cell is selected to form the dominant set. The complexity of the selection algorithm in this category is dependent on the number of neighbors (except for GAF, which requires global positioning system information instead). In other words, complete neighbor list knowledge is always the assumption. Note that the neighbor-learning procedure is not trivial in ad hoc networks, and it involves substantial overhead with high node density and mobility.
    The second approach is based on a clustering structure. Ephremides et al. first introduced the concept of clustering as a linked cluster algorithm (LCA) in the DARPA packet-radio network [12]. This was adopted and refined as the lowest ID (LID) and highest degree clustering algorithm by Gerla et al. [13]. Clustering was applied in many areas, including improving channel efficiency [14], routing performance [15], and power control [16]. Clustering in ad hoc networks was later revisited by Basagni [17], who proposed distributed and mobility-adaptive clustering (DMAC) to avoid the stationary limitation while the clustering process takes place.
    The basic concepts of efficient flooding with a cluster structure are described in [18] by Mase et al. They formulated efficient flooding for clustered ad hoc networks. The exclusion of ordinary nodes from flooding relay was formally presented as a gateway forwarding (GWF). Selected gateway forwarding (SGF) is a greedy solution of the set-cover problem (described later) to improve the flooding efficiency of GWF further. However, the extra cost (active signaling for clustering and gateways selections, i.e., exchanging extra packets) is difficult to validate. Simulation results in Section VI-B show how significant the cost is.
    To illustrate efficient flooding with clustering, let us consider the node example in Fig. 3 (we have not shown ordinary nodes in the figure.) Let be a transmission range, and the size of the roaming space be where k is an even number (Fig. 3 depicts the case of ). There are nodes in the square, but in the figure we only show the nodes at coordinates where either or is an integer smaller than . This “selection� of nodes is known as two-hop clustering, i.e., any two nodes in a cluster are separated by at most two hops. The nodes at the center of the circles are “clusterheads,� and the light-shaded nodes in between are “gateways.� Clearly, such nodes represent a connected set. They are in fact the dominant set required to forward the flood packets. Without the cluster overlay shown in Fig. 3, each flood packet is relayed exactly times if single relay is assumed where each node must rebroadcast the packet once. On the other hand, broadcasts suffice if only clusterheads and gateways forward the packet. Note that in the cluster restricted forwarding, all nodes still receive the flood packet. The flooding reduction is, thus, .

Fig. 3. Selective gateway flooding scenario.

    In the case of and , the number of broadcasts required in the cluster is 21 instead of 99. In other words, 78.8% of transmissions can be saved. This is not even a very dense network (each node has about 12 neighbors). As we increase the number of nodes in the system (and, therefore, the density), the clustering structure and, thus, the broadcast remains the same. As a result, the savings increase sharply with the node density.

IV.  PASSIVE CLUSTERING

    In this section, we introduce a new cluster formation protocol called “passive� clustering. The main novelty (with respect to conventional “active clustering,� e.g., LID, highest degree, or DMAC [17]) is that PC does not require explicit signaling. Another advantage is that PC efficiently controls the number of flooding packets on the fly, without requiring any prior structure. This reduces overhead and alleviates other limitations. Here, we present the concept of PC and illustrate its operation by example. The proof of correct operation and the detailed description can be found in [19].
    
A.  Protocol Overview

    PC is a cluster formation protocol that does not use dedicated protocol-specific control packets or signals. Conventional clustering algorithms require all of the participating network nodes to advertise cluster-dependent information repeatedly. Moreover, most of the existing clustering schemes require the execution of a separate clustering phase prior to any network layer activity (e.g., routing). With PC, we avoid these limitations. By monitoring user data packets that piggyback 2 b of cluster status information, we can build impromptu clusters for mobile wireless networks. Thus, the cluster infrastructure can be constructed as a byproduct of user traffic, without any dependency on the routing protocol.
    In PC, each node collects neighbor information from the MAC sender address carried by the incoming packets, and can construct clusters without collecting the complete neighbor list. This is an innovative approach to clustering which virtually eliminates major source of cluster overhead—the time latency for initial clustering construction as well as the communication overhead for neighbor information exchanges. Instead of using protocol-specific signals or packets, cluster status information (2 b for four states: initial, clusterhead, gateway, and ordinary node states) of a sender is stamped in a reserved field in the packet header. Sender ID (another key piece of information for clustering) is carried by all the existing MAC protocols, and it can be retrieved from the MAC header without any extra burden. Since the MAC packets are transmitted in broadcast (instead of unicast) mode in flooding, every node receives and reads the packets (in a promiscuous way) and, thus, participates in PC.
    Surprisingly, simulation results show that PC can form better clusters (well-connected clusters) than conventional clustering schemes based on weight (i.e., ID, degree, etc.) information [20]. The simulation results are backed up by analytical proofs [19]. This is because PC (as used in the support of ad hoc network routing schemes) uses network traffic that emanates from sources (i.e., the source in search of a path). If a cluster structure is constructed by flooding from a single source, the resulting structure is completely immune from logical isolation and lack of connectivity. Clustering stability and fast convergence time are other important properties required of clustering algorithms. To improve clustering stability, to speed up convergence, and most importantly, to avoid the “stationarity� requirement during the neighbor-learning and clustering phase, we developed a new clusterhead election rule which does not require any weight information. We call this rule “first declaration wins.�

Fig. 4. Clustering structure constructed by PC (with the same flooding scenario in Fig. 1). Solid circles represent declared clusters. Every node in the diagram is a critical node (solid—clusterheads; striped—gateways; gray—clusterhead-ready nodes.) There is no ordinary node in the example. The number of flooding relay was the same as in Fig. 1. But the resulting structure and routing overhead are similar, no matter how many neighbors are in the transmission range of the source “S,� whereas the control overhead is tightly coupled with the number of neighbors in the previously described cases (both AODV and OLSR.) The marked nodes have not sent any packet in the scenario. If any gray node transmits any packet, it becomes a clusterhead and declares a cluster. Dotted circles represent candidates of clusters.

    With the first declaration wins rule, a node which first claims to be the clusterhead remains the clusterhead and “rules� the rest of nodes in its clustered area (radio coverage). There is no waiting period (to make sure all the neighbors have been checked) unlike in all the weight-driven clustering mechanisms.
    
B.  Operational Description

    When a node is ready to become a clusterhead and has packets to send, it declares that it is a clusterhead by stamping its clustering state claim in the packets. As shown in Fig. 4, all the neighbors of the source “S� are ready to be a clusterhead when they receive the route request. Since PC does not support explicit control packets or signals of its own, a clusterhead-ready node must postpone its claim until it has outgoing “application� packet-level traffic, for example, flood search packet traffic. Among six of the illustrated cluster-ready nodes, only three could claim the role (solid nodes). After a successful transmission from an aspiring clusterhead, every node within radio coverage learns the presence of the clusterhead by monitoring the “cluster� state of the received packets. At this point, the neighbors of the clusterhead record the clusterhead information (clusterhead ID and the most recent transaction time-timestamp) and change their clustering states as discussed later.
    The readiness of being a clusterhead is determined by network activities as well as by the node's clustering state. After a period of inactivity (i.e., no incoming or outgoing traffic for longer than the cluster timeout period), all the nodes revert to the INITIAL state. Only nodes in INITIAL state can be clusterhead candidates—in other words, two-hop is the minimum distance between any two clusterheads, since all neighbors of a declared clusterhead exit the INITIAL state. After a clusterhead successfully asserts its state, it functions as a clusterhead. Clusterheads collect neighbor information by monitoring the network traffic. They are responsible for relaying intracluster packets.
    A node that hears more than one clusterhead becomes a GATEWAY. It reverts to ORDINARY node if it does not hear from more than one clusterhead for a given period. In the next section, we will describe a slightly modified procedure (selective gateway) in which a fraction of elected gateways gives up the role and reverts to ORDINARY nodes upon hearing a certain number of other gateways. A node that is neither a clusterhead nor a gateway is an ordinary node. The ordinary node does not forward flooding packets. It is precisely this forward-suppression mechanism that reduces flood overhead. Gateway nodes and clusterheads, on the other hand, forward the flood packets. Because of the passive nature of the collection mechanism, neighbor information is kept in “soft state� (i.e., it times out) and is possibly incomplete. Note here again that complete neighbor information is no longer necessary to form the structure. By using timestamps for neighbor information, we preserve the freshness of the information. Ordinary nodes and gateways keep a list of their clusterhead(s) in soft states. The timeout period has to be carefully chosen based on node mobility and communication pattern. Nonclusterhead nodes can collect neighbor clusterhead(s) information in a passive way. If a received packet is from a clusterhead (after checking the status information in the packet), the nonclusterhead node compares the sender ID of the packet with its clusterhead list and adds/refreshes accordingly.

V.  SELECTIVE GATEWAY PASSIVE CLUSTERING

    In typical ad hoc network/sensor network layouts, one discovers that the number of gateways is quite significant and in fact may exceed the number of ordinary nodes. Typically, not all of the gateways have to relay the flooding packets to assure reliable diffusion. It is mandatory to reduce the number of gateways in order to achieve efficient flood search packet suppression. Careful gateway selection is the proposed solution.
    To select the strictly minimal set of gateways, each clusterhead needs to collect the clusterhead list for each of its gateway, and then choose the minimum number of gateways that can cover all the neighboring clusterheads. This is a set-cover problem as stated in the flooding gateway selection (FGS) protocol. FGS was proposed by Mase et al. [18] to improve the flooding efficiency in a dense situation. If it were not for gateway selection, the gain is limited (because almost all of nodes are either clusterheads or gateways) as shown in their simulation results. Mase et al. [18] also reported chances of structural isolations in clustered networks based on the conventional weight-based clustering (LID or degree based) as well. The structural disconnections may end up with unnecessary detoured routes or failures in route search. To avoid clustering overhead considerations, they had to assume an additional channel for clustering-related and gateway selection packet exchanges. In ad hoc network/sensor networks, the gain from the efficient flooding with clustering is easily overridden by its overhead. As in FGS, the set-cover approach would introduce extra communication (it is even more expensive than clustering) and computation overhead, since the procedure requires clusterhead list reports from gateways. Even worse, it may lead to traffic bottleneck situations. In order to avoid the communication, computation complexity and extra channel (bandwidth) requirements, we introduce a heuristic solution to this problem as follows.
    
A.  Gateway Selection Heuristic

    Instead of selecting a single gateway between adjacent clusterheads (two hops away), we developed a heuristic algorithm that enables a limited number of gateways, and at the same time, preserves adequate connectivity within the resulting cluster structure. The selection algorithm provides many advantages, including on-the-fly flooding improvement, redundant connectivity, and higher overall flooding efficiency. The heuristics also promotes “distributed gateways� implementations [12]—i.e., two ordinary nodes work together to relay intercluster traffic. The basic mechanism of this heuristics is to control the number of gateways in an area without breaking the passive nature of PC.
    Every nonclusterhead node monitors and keeps track of the number of clusterheads (NC) and the number of gateways (NG) that can be overheard. Whenever a nonclusterhead node hears a packet from a clusterhead or a gateway, the node becomes a gateway if , where is a coefficient properly chosen based on the desired degree of gateway redundancy and is a gateway redundancy factor (, ). Otherwise, the nonclusterhead node becomes an ordinary node. The larger the number of clusterheads that a node can hear, the greater the chances of becoming a gateway. By manipulating , , we can control the number of gateways in the system—it is also helpful to avoid the bottleneck situations. The larger the number of gateways, the lower the gain in forwarding overhead reduction. On the other hand, if there are too few gateways, connectivity may be impaired leading to poor network performance. In this paper, and are global system parameters and are both set to one. The values of and should be chosen based on considerations including channel quality, noise level, as well as traffic pattern. For this reason, and can be local parameters, i.e., they can be locally adjusted to provide better adaptability and flexibility. In dense networks where packet collisions abound, higher values of those parameters lead to more gateways and better network performance by distributing network traffic over more gateways. Conversely, in low-density conditions, we propose to keep the parameters low to discourage multiple gateway creation. By introducing these heuristics, PC strikes a good balance between clusterheads and gateways and retains only a handful of forwarding nodes for flood search no matter how high the node density is. The gateway selection procedure is fully distributed, and requires only local information that can be overheard. No active packet exchanges, e.g., clusterhead-list exchange, is required.
    
B.  Flooding Improvement On the Fly

    Let us consider the example of single-source (a sink) flooding from a cold start. Every node is in the Initial state, and a source broadcasts a RouteRequest packet (as in AODV, for example). The immediate neighbors of the source receive the packet, and change their state to Clusterhead-Ready. When one of the neighbors is ready to forward the packet, it changes its state to Clusterhead, and broadcasts the RouteRequest packet with the Clusterhead state assertion. This time, all the nodes, including the source that receive the relayed RouteRequest packet from that newly proclaimed clusterhead, are eligible to become gateways, since they have heard from one clusterhead, and from no gateways (for simplicity, in this case we assume and .) Now, one of the candidate gateways except the source relays the flood search packet. This relay does not switch any gateways back to ordinary nodes because they still have the number of clusterheads which is equal to or smaller than the number of gateways (0 or 1). Let us say that a second gateway (within range of the first declared gateway) relays the flooding packet. Thereafter, none of the nodes in the intersection area of those two gateways can become a gateway. They become ordinary nodes after they receive the second flooding packet: the count for clusterhead equals one but they have heard from two gateways already. One may notice that there is a chance of “connectivity� loss with these heuristics. However, extensive simulation experiments have shown that the risk of flood delivery failure to certain areas of the network is negligible even with moderate node density. The simulation study in Section VI-A backs up the claim. As a result, the nodes are partitioned into two classes. The first is the class of critical nodes (initial, clusterhead, and gateway) which participate in flooding and routing operation; the second is the class of ordinary nodes which do not relay flooding. Nonsource nodes in the second class can go to sleep to improve energy efficiency if the network topology remains the same. As we have just described, every node in the intersection between two declared gateways immediately becomes an ordinary node, thus improving flooding efficiency on the fly. In conventional, active clustering approaches, such improvement was possible only after most of the clusters are constructed.

Fig. 5. A snapshot of the selective gateway PC.

    Fig. 5 shows a working example of selective gateway PC. This snapshot was taken from an actual simulation experiment with randomly placed 100 nodes in a 600 600 roaming space with 150 m of transmission range and ten constant bit rate (CBR) communication pairs which introduce 1 packet/s each. More details are found in the following section. All the nodes in the system turned out to be well connected by the cluster structure. There were 33 flooding participants out of 100 nodes (black in color).
    
C.  Properties of the Passive Cluster Solution

    It is appropriate at this point to compare and contrast PC with traditional LID active clustering. We have already discussed the impact of the background updating procedure and the neighbor list broadcast requirements on the control traffic overhead caused by active clustering. Here we focus on the structure of the solutions. Typically, one finds that the two solutions are comparable (in terms of number and layout of clusters). But there are fundamental differences, as follows.
    

1. Active clustering is carried out independently, in the background and in parallel across all nodes in the network, while PC is “on-demand� and is initiated by a single “source,� namely, the first source that needs to send data. Thus, active clustering tends to grow increasing aggregations of clusters, which eventually might turn into disconnected islands (the latter requires the “distributed gateway� feature—i.e., gateway to gateway links—to reestablish connectivity). PC does not suffer from this problem (although it can also be extended to support distributed gateways).
   
2. PC features the “selective gateway� provision. Popular active clustering schemes do not include such feature.
   
3. The LID feature tends to make the clustering more sensitive to mobility (frequent changes)—the clusterhead can be more easily challenged by newcomers with lower ID.
   

    Another important issue is the suitability of PC for low-energy operations, as in battlefield scenarios or sensor network applications. Repeated selection of the same subset of clusters and gateways can be detrimental to low power operation in that it creates uneven energy consumption. In this respect, PC is beneficial. In fact, it favors even distribution, since at each new cluster formation round (caused by the arrival of a new user data session, say), new clusters and gateways are selected as the source changes and/or, even in the case of same source, the random timers cause different clusterheads and gateways to assert their role first. In the case of “permanent� traffic pattern where the cluster structure tends to persist, a possible remedy is to associate the clusterhead and gateway status with a minimum energy level requirement. When energy drops below this threshold, the role is given up, triggering a new election.

VI.  SIMULATION STUDY

    The simulation models used for the performance evaluation were implemented in the GloMoSim library [21]. The GloMoSim library is a scalable simulation environment for wireless network systems using the parallel discrete-event simulation language called PARSEC [22]. The distributed coordination function (DCF) of IEEE 802.11 [23] is used as the MAC layer in our experiments. The radio model uses characteristics similar to a commercial radio interface (e.g., Lucent's WaveLAN). The radio propagation range for each node is 150 m, and channel capacity is 2 Mb/s. The roaming space is 600 600 m square (except for the experiment described in Section VI-D.) The random waypoint model [24] was used for node mobility.
    We use AODV [5] as a routing protocol in evaluation because AODV is one of the most popular on-demand protocols. The only modification we made to AODV was to limit the flood search forwarding function to “nonordinary nodes.�
    
A.  Flooding Delivery Experiments

    To test the path-finding performance (reachable test), we ran the following simulations. We first deployed 100 nodes in the simulation space at random. Without node mobility, we chose 2400 random source and destination pairs and ran the selective gateway flood search from cold start one by one. Only one data packet is sent from each source to each destination. There is only one source and destination in a given period (which is larger than the cluster timeout of 2 s to ensure that no residual clustering structure remains after the single transmission. The source finds the destination with the “scoped� selective gateway flood search. A short data packet will be pumped through a path if the path is successfully found. One hundred percent packet delivery was observed with the experiment.
    We also ran a batch of experiments with mobile nodes. We randomly selected communication pairs with varying node speeds of 0, 2, 4, 6, and 8 m/s. The source sends out a packet every 15 s to the destination for 100 min (400 tries). The slow packet rate (1/15 packet/s) is to ensure that the cluster structure built by the previous packet delivery dissolves after cluster timeout (2 s). With the speeds of 0, 2, and 4 m/s, we observed 100% packet delivery. In the case of 6 m/s and 8 m/s, we observed 99.25% and 98.25% of packet delivery ratio, respectively. Such packet drop was investigated and traced to route breakage—after a source finds a path, the source cannot deliver the packet because the path was broken in the interim due to the motion.
    
B.  Comparison with Active Clustering

    

Fig. 6. Throughput (2 m/s).

We first compared the efficient flooding mechanism introduced by active and passive clustering. The active clustering (LID, highest degree, etc.) pays the extra control overhead for the hello messages. It should be noted that LID can be built in the process of hello message exchanges. We implemented active clustering as closely as possible to the original description of the LID algorithm [14]; however, nonrealistic assumption cannot be satisfied—e.g., the stationary assumption [17] that nodes are static (no mobility) when neighbor information is collected.
    To obtain the correct neighbor information, the following corrective hello messaging was adopted.

1. Every node broadcast hello messages once per second. The hello message carries the sender's neighbor list, sender ID, and cluster related information.
   
2. Once a node receives a hello message that does not list its ID in the neighbor list, it notifies the sender of discrepancy by sending another hello message.
   
3. If there is no additional hello message during the epoch, a node can assume that it has collected the complete neighbor list, and the clustering can be concurrently performed.
   

    

Fig. 7. Throughput (6 m/s).

In addition to the original active clustering method, we also evaluate a hybrid clustering scheme that utilizes the overhearing neighbor discovery as in PC. We have, thus, four protocols to compare:

1. AODV that does not incur a single hello overhead, but does not attempt to improve its flooding mechanism;
   
2. AODV-PC that is equipped with efficient flooding introduced by PC;
   
3. AODV-PCH that uses overhearing neighbor discovery as well as additional hello messaging to obtain uncollected neighbor lists; and
   
4. AODV-AC (active clustering, LID) that uses hello messaging as previously described.
   

CBR source and destination pairs are randomly chosen, and 10 to 90 pairs are introduced to increase the offered load. Data packets are 512 B long (1 packet/s). Each simulation is executed for 5 min of simulation time.
    As can be observed in Figs. 6 and 7, active clustering as well as hybrid clustering promotes network saturation. In higher mobility situation where neighbor changes are more frequent, active clustering and hybrid clustering experience saturation at a lower offered load. AODV also saturated in the higher mobility case.

Fig. 8. Delivery ratio.

Fig. 9. Delay.

    
C.  No Mobility—Dynamic Traffic

    

Fig. 10. Throughput.

Fig. 11. Delay.

In this experiment, we froze the node positions and injected short sessions with bursty traffic to model sensor applications. The packet rate was 0.4 packet/s (3 packets per session). A given number of new source and destination pairs were selected to participate in such bursty communication every 3 s with randomized starting times. This simulation tests the path-finding capability of original AODV and AODV with selective gateway PC in various network load conditions. Because there is no mobility, there is no packet delivery loss due to a path break. This scenario is very similar to that of a sensor network where all the nodes are fixed and communication patterns are short and bursty. Fig. 8 shows packet delivery ratio as a function of number of communication pairs. AODV with selective gateway PC outperforms conventional AODV in the whole range of the simulation window. The delay (see Fig. 9) also reports the faster network saturation in case of AODV. These do demonstrate the effectiveness of the PC in reducing flood redundancy. The selective gateway PC finds paths well, and at the same time, it reduces interference of multiple flooding searches by limiting flood packet relays.
    
D.  Adaptive Density

    To demonstrate the density adaptability of the proposed scheme, we increased the number of nodes in the system without changing the size of the roaming space 1500 m by 500 m. Traffic sources are CBR. The source-destination pairs (20 pairs) are totally randomized. Data packets are all 512 B long (2 packets/s). Each simulation is executed for 10 min of simulation time. We report the average throughput and the average delay of the network. As we can observe in Figs. 10 and 11, PC preserves the throughput and controls the average end-to-end delay of AODV. On the other hand, we could not even finish simulations with 700 or more nodes in case of AODV without PC enhancement due to the excessive control overhead. This experiment confirms that the cost of flooding increases as the density increases. It also exhibits the importance of using some form of forwarding node selection to avoid density “explosion.�

VII.  CONCLUSION

    We have presented an overview of various flooding mechanisms and described their advantages and limitations. It is shown that on-demand clustering algorithm can provide efficient flooding in ad hoc network/sensor networks. PC can be truly called “on-demand� clustering; if there is no network usage, there is no clustering. As the user starts transmitting data, this protocol deploys clusters as quickly as the destination-finding process finds the routes. We proposed superimposing an on-demand cluster structure which can be quickly deployed in the “unstructured� network, and let only critical nodes participate in the flooding process. This will save significant flooding overhead (route search cost) in dense ad hoc network/sensor networks in a large scale. However, the gain of the proposed method is limited in small and sparsely deployed ad hoc networks. Another limitation is the longer hop counts (less than 1 hop on average) because the reduced flooding does not utilize every node as a relay.
    Efficient flooding with PC is the only mechanism for on-demand routing without extra bandwidth requirements. Due to its passive nature, PC does not require any control packets specific to this protocol. Thus, it can reduce the cost of flood search significantly without introducing any extra bandwidth requirements. As an added advantage, there is no preparation time or overhead for selecting dominant sets. Therefore, the number of flooding relays can be significantly reduced even during the first flooding. This is the unique feature and the principal advantage of the proposed mechanism. It is especially useful for ad hoc network/sensor networks with dynamic topology. Its efficiency increases linearly with the number of neighbors, i.e., with node density. Besides assisting with flood reduction, the clustering structure offers several other side benefits. In particular, it can be beneficial to routing scalability, reliability, and QoS support.
    PC is a self-sufficient clustering scheme. The protocol collects all the necessary information by overhearing and does not require costly information like global topology knowledge from the lower layer. The resulting cluster structure is better than other existing clustering algorithms in terms of stability, mobility robustness, and connectivity. It can be utilized in cluster-based ad hoc network/sensor networking [15] as well as sleep node selection. Unlike other clustering mechanisms, PC can build the cluster structure with partial neighbor information which, in most cases, is the only possible information available in ad hoc network/sensor networks. In many areas, including military applications (e.g., SensIT), this feature has merit, since it permits building the clusters without exposing network topology details to eventual eavesdroppers.

REFERENCES

Taek Jin Kwon (Member, IEEE) received the B.S. and M.S. degrees in computer engineering (with honors) from Seoul National University, Seoul, Korea, in 1989 and 1991, respectively, and the Ph.D. degree in computer science from University of California, Los Angeles, in 2000.     He is currently a Research Scientist with Telcordia Technologies, Red Bank, NJ. His research interests include scalability issues in ad hoc network/sensor networks and IP mobility management, network security, ad hoc network routing protocols, and home networking systems.

Mario Gerla (Fellow, IEEE) received the Laurea degree in electrical engineering from the Politecnico di Milano, Milano, Italy, in 1966, and the M.S. and Ph.D. degrees in engineering from the University of California, Los Angeles (UCLA), in 1970 and 1973, respectively.     He joined the Faculty of the UCLA Computer Science Department in 1977. His research interests cover the performance evaluation, design, and control of distributed computer communication systems and high-speed computer networks (B-ISDN and optical networks).

Vijay K. Varma (Senior Member, IEEE) received the M.Tech. degree in electrical engineering from the Indian Institute of Technology, Kanpur, in 1978 and the Ph.D. degree in electrical engineering from Southern Methodist University, Dallas, TX, in 1986.     He is a Senior Scientist in the Wireless Systems Department of Telcordia Technologies, Red Bank, NJ. He has 18 years experience in wireless communications as a researcher and systems engineer. Since joining Bellcore (now Telcordia Technologies) in 1985, he has been involved in various systems issues, including speech coding, signaling and call control protocols, wireless data, and mobility management protocols, for wireless and personal communications. He has published several papers in the area of wireless communications and has given many tutorials, organized workshops, and sessions at various IEEE conferences. His current research interests include 3G wireless network architectures, wireless-to-IP network interworking, WLAN/3G wireless network integration, and ad hoc networking.     Dr. Varma is a Member of Tau Beta Pi. He received the Frederick E. Terman award from the School of Engineering and Applied Science, Southern Methodist University.

Melbourne Barton (Senior Member, IEEE) received the B.Sc. degree in electrical engineering with first-class honors from the University of the West Indies, St. Augustine, Trinidad, in 1978, the M.Sc. degree in telecommunications systems from the University of Essex, Colchester, U.K., in 1981, and the Ph.D. in electrical engineering from the University of Rhode Island, Kingston, in 1988.     From 1983 to 1990, he was a Lecturer in the Department of Electrical Engineering at the University of the West Indies. In 1990, he joined the Applied Research Area of Telcordia Technologies as a Research Scientist, and is currently a Director in the Wireless Systems and Networks department. He has worked on ADSL technologies, network integration issues relating to wireless ATM, wireless NGN, wireless network configuration management, wireless network security, and 3G and WLAN integration.     Dr. Barton is a former British Commonwealth and Fulbright-LASPAU scholar.

T. Russell Hsing (Fellow, IEEE) received the B.S. degree in electrophysics from the National Chiao-Tung University, Hsin-Chu, Taiwan, R.O.C., in 1970, and the M.S. and Ph.D degrees in electrical engineering from the University of Rhode Island, Kingston, in 1974 and 1977, respectively. From 1995 through 1997, he finished business courses from the Stanford Business Graduate School, Palo Alto, CA; the Massachusetts Institute of Technology Sloan School, Cambridge; the University of Texas, Austin; and the University of Illinois, Urbana-Champaign.     He has been affiliated with research laboratories at Burroughs, Detroit, MI: Xerox, Webster, NY; GTE Labs, Waltham, MA; Telco Systems Fiber Optics Corporation, Norwood, MA; and TASC, Reading, MA. He joined Telcordia Technologies (formerly Bellcore), Morristown, NJ, in 1986 and is now an Executive Director. He manages overall wireless research programs and business strategy development in Telcordia Technologies, Applied Research.     Dr. Hsing is a Fellow of the SPIE—The International Society of Optical Engineering.