HCS: hierarchical cluster-based forwarding scheme for mobile social ...

1 downloads 241 Views 2MB Size Report
Dec 19, 2014 - Abstract Clustering has been shown to be a highly effective way to reduce network traffic in mobile ad hoc networks. Many clustering schemes ...
Wireless Netw (2015) 21:1699–1711 DOI 10.1007/s11276-014-0876-x

HCS: hierarchical cluster-based forwarding scheme for mobile social networks Sun-Kyum Kim • Ji-Hyeun Yoon • Junyeop Lee Sung-Bong Yang



Published online: 19 December 2014  Springer Science+Business Media New York 2014

Abstract Clustering has been shown to be a highly effective way to reduce network traffic in mobile ad hoc networks. Many clustering schemes have been proposed. However, none of these schemes can be directly applied to a mobile social network because they are designed for well-connected networks and require timely information sharing among the nodes. In this paper, we propose the hierarchical clustering-based forwarding scheme (HCS), which implements hierarchical clustering on social information. Each node constructs hierarchical clusters based on common neighbor similarity at the end of the warm-up period. The nodes then forward a message to other nodes based on the clustering information and similarity scores. HCS exploits the shortcuts on the path toward the destination node with the help of social similarity and node movement patterns. Experiments were performed on an NS-2 network simulator. The results show that HCS reduces network traffic compared to non-clustering schemes, such as Epidemic, SimBet, PRoPHET, and common neighbor similarity schemes, while maintaining acceptable transmission delay compared to the Wait scheme.

S.-K. Kim  J.-H. Yoon  J. Lee  S.-B. Yang (&) Department of Computer Science, Yonsei University, Seoul, Korea e-mail: [email protected] S.-K. Kim e-mail: [email protected] J.-H. Yoon e-mail: [email protected] J. Lee e-mail: [email protected]

Keywords Hierarchical clustering  Forwarding  Routing  Mobile social network (MSN)  NS-2

1 Introduction With advances in wireless communication technologies and the popularity of mobile devices, mobile social networks (MSNs)—also known as opportunistic networks [1] or pocket switched networks [2]—have rapidly become an increasingly popular field of networking research. MSNs are applicable to mobile wireless communications such as Sami Network Connectivity Project [3], Zebranet [4], Shared Wireless Info-Station [5], Vehicular Delay Tolerant Networks [6], and so on. MSNs have evolved from mobile ad hoc networks (MANETs) [7] and delay-tolerant networks (DTNs) (also known as disruption tolerant networks) [8] with social characteristics [9]. MSNs carry a more general concept with human-carried devices and include DTNs. Also, MSNs do not assume any compatibility with the Internet architecture, nor any a priori knowledge about the network topology, the areas of disconnections, or future link availability [10]. MSNs use contact opportunities and rely on devices carried by humans to relay messages for others [9]. However, like DTNs, MSNs suffer from intermittent connectivity and long-lasting disconnections due to low node density, short transmission ranges, and free node mobility. In addition, there may be no complete paths between the source and destination nodes [1]. Thus, the forwarding or routing schemes for MANETs are not applicable, and development of such schemes in MSNs have become a challenging problem. In MSNs, nodes communicate with the multi-hop and relay nodes that forward message addresses to other nodes in MSNs. In this case, however, forwarding is not ‘‘on the

123

1700

fly’’ because the relay nodes store messages when no forwarding opportunity exists, such as when there are no nodes within the transmission range. Moreover, they exploit any contact opportunity with other nodes in order to forward the message [10]. This forwarding mechanism is called a store, carry, and forward scheme and is performed hop by hop. In MSNs, node mobility creates opportunities for communication, whereas mobility in MANETs is viewed as a disruption of connections among nodes [10]. Therefore, the key issue in message forwarding is selection of proper nodes for message (or a copy of the message for a multi-copy scheme) handed [11]. Clustering has been shown to be a highly effective way to reduce network traffic in MANETs, and many clustering schemes have been proposed. However, none of them can be directly applied to MSNs because they are designed for well-connected networks and require timely information sharing among the nodes [12]. Among the proposed clustering schemes in MSNs, a distributed clustering scheme based on an exponentially weighted moving average [12] has been proposed. However, this scheme requires gateway nodes, each of which plays a role as a bridge between two nodes in different clusters. If a gateway node operates unpredictably, the scheme suffers from low performance. The DTN hierarchical routing (DHR) scheme [13] has been proposed to improve routing scalability. However, DHR is based on a deterministic mobility model in which all nodes move according to strict, repetitive patterns. Hence, the method is very difficult to implement in MSNs. Because mobile nodes have limited resources, such as bandwidth, power consumption, channel utilization, network size, and so on, the nodes in MSNs experience some communication difficulties. Therefore, as network traffic increases, network problems, such as bottlenecks, slow communication, and noise problems, are unavoidable. In addition, applications in MSNs should be relatively delaytolerant. However, it is still of interest to minimize the delay whenever possible [24]. To resolve this problem, we propose the hierarchical cluster-based forwarding scheme (HCS) to reduce network traffic while maintaining acceptable transmission delay. HCS constructs hierarchical clusters [14] using social information with the home-cell community-based mobility model (HCMM) [15]. Hierarchical clustering is a highly effective way for clustering nodes in MSNs because each node plays a similar role in flat clustering. This method is simple and effective in small networks but not applicable to large-sized MSNs [14]. HCS exploits agglomerate hierarchical clustering in the same way as in DHR. However, HCS adopts the common neighbor similarity to construct hierarchical clusters instead of the contact probability used in DHR. In HCS, each node utilizes both hierarchical clustering and similarity scores. The clustering information

123

Wireless Netw (2015) 21:1699–1711

helps deliver the message with a smaller number of hops owing to social similarity and node movement patterns. The experimental results show that HCS reduces network traffic compared to non-clustering schemes, such as Epidemic [18], SimBet [24], PRoPHET [29], and common neighbor similarity [32] schemes, while maintaining decent transmission delay compared to the Wait scheme. Mobile nodes in MSNs generally more frequently visit certain places, like home communities, while visiting other locations only occasionally [16]. Because the home community-based mobility model (HCMM) reflects such a characteristic, it is well-suited for use in MSNs. The main contributions of this paper can be summarized as follows. 1.

2.

3.

We propose a scheme using agglomerative (bottomup) hierarchical clustering with common neighbor similarities. Cluster level-based forwarding is introduced. To more efficiently deliver a message to the destination node, similarity-based forwarding compensates for level-based forwarding. Using this technique, a node sends a message to other nodes with higher similarity scores with respect to the destination. We conduct extensive simulations for experiments with the network simulator NS-2 and compare the results with non-clustering schemes, such as Epidemic, Wait, SimBet, PRoPHET, and common neighbor similarity schemes.

The rest of this paper is organized as follows. In Sect. 2, we discuss related work. After introducing the simplified MSN model in Sect. 3, we describe the proposed scheme in Sect. 4. The simulation environment and results are presented in Sect. 5. Finally, the conclusion and future work are outlined in Sect. 6.

2 Related work Existing routing protocols for MANETs [7], such as Dynamic Source Routing, Ad Hoc On-Demand Distance Vector, Split Multipath Routing, Shortest Multipath Source, and AntHocNet, have been introduced. A doublelayered peer-to-peer system using clustering was also proposed for improved routing performance [17]. None of the current schemes are applicable to MSNs regardless of improved performance because of the requirement for no complete path between the source and destination in MSNs. The opportunistic routing schemes can be classified in two categories: zero knowledge schemes and non-zero knowledge schemes. Zero knowledge schemes use no social information, while non-zero knowledge schemes

Wireless Netw (2015) 21:1699–1711

take advantage of the information about node behaviors or social relationships in order to make decisions for forwarding messages. Non-zero knowledge schemes are the ones utilized in MSNs. Zero knowledge schemes include Epidemic [18], Sprayand-Wait [19], Controlled routing protocol for DTN based on hierarchy forwarding and cluster control (CRHC) [20], Backpressure with adaptive redundancy (BWAR) [21], Backpressure-based routing scheme [22], Homing spread [16], and Hotspot-based forwarding scheme (HFS) [23]. In Epidemic when each node meets other nodes, it distributes the message to each of them, creating the replicas of the message. In Spray-and-Wait a node ‘‘sprays a number of copies into’’ some nodes in the network and then ‘‘waits’’ until one of these nodes meet the destination. First, a fixed number k of messages is transferred by spraying a half of kcopies. When each node has the last message after spraying messages, it waits until the moment when it meets the destination node during the wait phase. In CRHC, a node spreads a half of the messages to other nodes until meeting the destination when the destination is in the same cluster. On the other hand, a node sends a message to the clusterhead in the different cluster first, when the destination node is in a different cluster. However, it is not applicable to our environment because the network should select cluster heads and stable nodes in advance, where we can easily predict their moving patterns. BWAR takes advantages of replication to reduce delay under low load conditions. It creates copies of packets in a new duplicate buffer upon an encounter, when the transmitter’s queue occupancy is low. In Backpressure-based routing scheme a node can make source rate, packet routing, and forwarding decisions without the notion of end-to-end routes, using information about queue backlogs and link states. Both BWAR and Backpressure-based routing schemes are applicable to the environment considering queue, link states and load conditions. Homing Spread utilizes community structure to identify suitable relay nodes. These approaches spread messages to the detected communities via relay nodes; however, they incur extra delivery overhead for mobile nodes. HFS floods messages only in hotpots, where nodes often interact; however, the size of the hotspots is limited. Various non-zero knowledge schemes have been proposed for MSNs [11, 24–31]. The non-zero knowledge schemes can be further classified into three schemes: centrality/similarity-based, social context-based, and probability-based. The centrality/similarity-based schemes include SimBet [24], Bubble Rap [25], and SANE [26]. SimBet makes use of the exchange of both betweenness centrality metrics and the locally determined social ‘‘similarity’’ to the destination node. When a node encounters other nodes, it transfers a message to the node with the higher utility values of betweenness centrality and

1701

similarity until reaching the destination node. Bubble Rap takes advantage of both global and local centrality. The bubble-up operations transmit a message to the destination node or its community. However, when the destination belongs only to a community whose members all have low global centrality values, such a strategy may fail. In this case, a relay node in the same local community as the destination node cannot be identified. SANE utilizes user interests and similarity. Social context-based schemes include Label [27] and HiBop [28]. In Label, each node is assumed to hold the label information of other nodes in its social community, similar to name tags used in a conference. Based on the labels, the routing scheme selects nodes for directly forwarding messages to the destination or for acting as the next-hop node that shares the same label as that of the destination. HiBop requires personal information, such as residence, work, hobbies and fun, as well as system information. Finally, probability-based schemes include PRoPHET [29], PeopleRank [30], and MobySpace [31]. PRoPHET first estimates a probabilistic metric called the delivery predictability, P(a, b), which indicates how likely it is that node b will receive a message from node a during a warmup period. Two nodes exchange the summary vector of the information on the messages and the delivery predictability vector. The information in the summary vector is used to determine which messages should be sent for requesting information from other nodes. PeopleRank uses the PageRank algorithm of Google as a guide for forwarding decisions. Whenever two neighbor nodes in the social graph meet, they exchange their current PeopleRank values and their numbers of social graph neighbors. MobySpace takes advantage of the knowledge concerning node mobility; however, it requires global information for routing. Non-zero knowledge schemes are very effective in forwarding messages. However, most non-zero knowledge schemes require global information for forwarding decisions. These schemes therefore exploit real datasets for their simulations. These real datasets can be processed in advance because they contain information on mobility, contact trace, and social interaction graphs [23].

3 Simplified MSN model This system obeys the rules of typical message forwarding, whereby each node forwards a message to the destination node. We assume the network is represented by the graph G = hV, Ei, where the vertex set V consists of all nodes, and edge set E consists of the social links between nodes. Each node in MSNs has a unique identifier and is denoted

123

1702

Wireless Netw (2015) 21:1699–1711

by Ni, i = 1, 2, …, m, where m is the number of nodes in the network. Each node Ni keeps track of a set Ci of nodes that Ni has encountered. Each node Ni belongs to a single community, called its home community [2] so that it has a label [27] that indicates its home community, denoted by Hi, where Hi is one of {1, 2, …, r} and r is the number of communities. Each node moves freely from its own home community to other communities and is aware of its own speed and current location. Each node periodically assesses its location. To determine the speed and location, each node is assumed to have positioning system equipment. For simplicity, we do not consider resources such as buffer size, bandwidth, or power. Although such measurements would result in additional computational costs, we do not consider such computational costs in this paper because the computation can be performed with simple equations. Moreover, the focus of this paper is not on computational cost reduction. The following table summarizes key notations used in describing HCS; some notations are introduced in the following sections (Table 1).

4 Proposed scheme 4.1 Information exchange in the warm-up period The greater is the number of common associates between a pair of persons, the more likely they are to be friends with each other, and the more frequently they are likely to meet each other. Such a social concept is naturally suited for the ‘‘common neighbor similarity’’ in the network [32] if we interpret a neighbor of a node as a commonly encountered node. A pair of nodes becomes more ‘‘similar’’ to each other as the number of common neighbor nodes increases. During the warm-up period, nodes exchange and update their information, including their own similarity scores. The purpose of updating the similarity scores is to

accumulate the node contact information so that each node is able to obtain the estimated global information on the entire network. The similarity score between a pair of nodes, Ni and Nj, i = j, can be computed with |Ci \ Cj| that is the number of nodes encountered by both Ni and Nj. Each node Ni maintains the information vector of (Ni, Hi, Ci, Si), where Si is the similarity score lists. During the warm-up period, whenever Ni meets other nodes, Ni exchanges its information vector with each of the encountered nodes and updates both Ci and Si. Figure 1 shows a data structure of Si, where |Ci \ Cj| is the number of nodes encountered by both Ni and Nj. Observe that Si contains not only the contact information of Ni itself but also the contact information of other nodes; for example, in Fig. 1, Si also contains the contact information of Nj and Nk. If any of Nj and Nk have met other nodes before encountering Ni, there would be more entries in Si for their contact information. We now explain how two nodes update their own similarity scores after exchanging the information vectors each other. Assume that two nodes, say N1 and N2, are approaching each other at time t1 with their similarity score lists S1 and S2 as shown in Fig. 2(a). Assume that N1 had encountered N3 and N4 met N3 for the first time, respectively, and later N1 has encountered N4. Therefore, |C1 \ C4| = 1, indicating that both N1 and N4 commonly encountered N3 before they meet each other. All such encounters have been recorded at S1. Similarly, N2 is assumed to have its contact information as in S2. At time t2, N1 and N2 exchange their information vectors and update their similarity score lists as shown in Fig. 2(b). In the first and second entries of S1 and S2, (N2,1) and (N1,1) have been created with score |C1 \ C2| = 1, indicating that both N1 and N2 had encountered N3 earlier than t2. In Fig. 2(b), S1 has the shaded entries that came from S2 at time t1, while S2 has the unshaded entries that came from S1 at time t1. However, when S1 and S2 have a common entry such that one entry has larger similarity

Table 1 Key notations used in HCS Notation

Definition

m

The number of nodes in the network area

r

The number of communities in the network area

Ni

Node i, where i = 1, 2, …, m

Hi

Home community of Ni

Ci

A set of nodes that Ni has encountered

Si

The similarity score lists of Ni

Li[j]

Nj’s level in the hierarchy of clusters constructed by Ni

Ki[j]

A set of nodes in the cluster at level j in the hierarchy of clusters constructed by Ni A threshold of level to quit the clustering

d

Fig. 1 Data structure of Si

123

Wireless Netw (2015) 21:1699–1711

1703

Fig. 2 Update of the similarity score lists. (a) Before the encounter (b) During the encounter

score than the other, the entry with smaller score is updated with the larger one. Notice that S1 and S2 are equivalent after updating. By the end of the warm-up period, each node accumulates the contact history with similarity scores to extract the global information on the network. After the warm-up period, each node builds hierarchical clusters based on the obtained similarity score lists.

4.2 Hierarchical clustering Hierarchical clustering is widely used for finding community structures in a network. HCS adopts bottom-up hierarchical clustering based on common neighbor similarity. In bottom-up hierarchical clustering, also known as agglomerative clustering, each node serves as a single cluster in the beginning of the warm-up period, and clusters are iteratively merged until a certain condition is satisfied. In HCS, each node hierarchically builds based on its similarity score lists at the end of the warm-up period. We modify bottom-up hierarchical clustering in two ways. First, clustering ceases at some predetermined level of the hierarchy. Second, more than two clusters can be merged in a larger cluster. Each node performs clustering greedily in terms of similarity scores; that is, clusters containing the nodes with the highest similarity score are first merged into a cluster. Note that, once the nodes with the highest score belong to

the same cluster, their similarity scores will no longer be considered for future clustering steps. Figure 3 illustrates how a node, say N1, performs clustering with a threshold d = 4. In Fig. 3, {N4} and {N5} are merged first at level 1, because |C4 \ C5|, the similarity score between N4 and N5, is assumed to be the largest. We store such clustering information in an 1-dimensional array K1, where Ki[j] denotes a set of nodes in the cluster at level j in the hierarchy of clusters constructed by Ni. Hence K1[1] = {N4} [ {N5} = {N4, N5}. We also store the level number ‘‘1’’ for each of N4 and N5 in an 1-dimensional array L1; L1[4] = L1[5] = 1. Then, {N4, N5} and {N3} become one cluster at level 2, assuming that |C3 \ C4| is the next largest; K1[2] = {N4, N5} [ {N3} = {N3, N4, N5}. We assign level ‘‘2’’ to N3; L1[3] = 2. In the figure, at level 3 {N3, N4, N5}, {N1}, and {N2} are merged into a single cluster, because |C1 \ C3| = |C2 \ C3| are assumed to be the next largest. Hence, K1[3] = {N3, N4, N5} [ {N1} [ {N2} = {N1, N2, N3, N4, N5} and L1[1] = L1[2] = 3. Finally, {N6} and {N7} are merged at level 4 in the above example; K1[4] = {N6} [ {N7} = {N6, N7} and L1[6] = L1[7] = 4. The reason why N6 and N7 are merged into a different cluster from K1[3] is that |C6 \ C7| is assumed to be the next largest similarity score and neither N6 nor N7 belongs to K1[3] = {N1, N2, N3, N4, N5}. In the above hierarchical clustering, the higher the similarity is, the lower level the nodes are merged at. HCS stops merging clusters at a certain level d, which is

123

1704

Wireless Netw (2015) 21:1699–1711

Fig. 3 An example of hierarchical clustering

determined through extensive simulation. Algorithm 1 formally describes the hierarchical clustering algorithm for HCS.

The message forwarding process in level-based forwarding mimics the way people forward messages in virtual clusters with similar movement patterns. Within such a cluster, it is

4.3 Hierarchical clustering-based forwarding scheme

highly likely that some people could travel around the destination of a message. Note that HCS does not compute the delivery probabilities for forwarding messages. Instead, it exploits shortcuts on the path toward the destination node with the help of social similarity and node movement patterns. To implement this concept, each node in HCS creates its own hierarchical clusters in which nodes can frequently interact with each other. Even though each node independently builds its own hierarchical clusters, we expect that the clustering results of the nodes are quite similar to each other, because the nodes continuously exchange their information vectors and update similarity score lists whenever encountering other nodes during the warm-up period. Such clustering results can be viewed as the global network information. In HCS, clustering in each node is not updated during the forwarding process, because it causes additional communication overheads and does not improve the overall performance noticeably. Figure 4 illustrates how message forwarding is achieved in HCS. Figure 4(a) shows how level-based forwarding is performed. Suppose that N2 is the source node and N5 is the destination node. A solid arrow indicates a possible

In this section, we describe the proposed HCS forwarding scheme. In HCS, each node Ni uses either level-based forwarding or similarity-based forwarding. In level-based forwarding, Ni forwards the message to Nj if Li[j] \ Li[i]. Note that, when nodes are clustered at lower levels, they are more likely to be similar; that is, they have encountered more common nodes. Hence, in the future, they are expected to meet ‘‘valuable’’ nodes, including the destination node. By valuable we mean that a node is more likely to reach the destination node. On the other hand, if Ni encounters only Nj such that Li[j] C Li[i], then Ni stops using the clustering information, and instead simply uses the similarity scores in its Si. In other words, Ni forwards the message to a node with the highest similarity score with respect to the destination node. Such forwarding is called similarity-based forwarding. Hence, similarity-based forwarding serves to compensate for cases when the level of the destination node is relatively higher.

123

Wireless Netw (2015) 21:1699–1711

1705

Fig. 4 Examples of forwarding process. (a) Level-based forwarding (b) Similarity-based forwarding

forwarding, while a dotted arrow denotes that there is no message forwarding. N2 forwards the message only if the level of the encountered node is lower than that of N2. Here, N2 can forward to N3, N4, or directly N5, if they are within the communication range of N2. If N3 receives the message from N2, N3 first looks into its level information, and then forwards the message to the encountered nodes with lower level than the level of N3. If N4 receives the message from N2, N4 can send the message to N5 when they encounter each other.

N7 whenever they are encountered, because they have higher similarity scores than N5; that is, |C5 \ C2| \ |C4 \ C2| \ |C3 \ C2|. Note that they can check their similarity scores with the destination N2 in their similarity score lists without additional communication with N2. However, N5 never sends the message to N1 nor N6 because their scores are lower than that of N5. In general, once node Ni has a message, it can forward the message to a node that has higher similarity score than that of Ni with respect to the destination.

Figure 4(b) illustrates how similarity-based forwarding is accomplished. We assume that the source node is N5 and the destination node is N2. The integer next to each node is the similarity score with respect to the destination node. In this case, because the level of the destination node is higher than that of the source node, HCS exploits the similarity scores. Therefore, N5 can forward the message to N3, N4, or

At the end of the warm-up period, each node independently constructs its own clusters and generates a message to send. When each node Ni encounters a set of nodes within its communication range, it executes Algorithm 2. Let E be the set of (encountered) nodes within the communication range of Ni. Assume that Ns is the source node and Nd is the destination node. In Line 2, when Nd is in E,

123

1706

Wireless Netw (2015) 21:1699–1711

Ni simply forwards the message, and the forwarding process is done. Otherwise, Ni sends the message with levelbased forwarding if Li[d] \ Li[s]; that is, if the level of the destination node is lower than that of the source node. Therefore, Ni attempts to find node Nj, whose level is lower than Li[i]. If Li[d] C Li[s], Ni sends the message with similarity-based forwarding. Therefore, it tries to finds node Nj whose similarity score is higher than that of Ni with respect to Nd. In either case, Ni sends the message to Nj if found in E in Line 9. If Ni cannot find Nj in Lines 6 or 8, Ni does not transmit the message at this time.

Table 2 Simulation parameters Parameter

Value (default)

Network area

450 9 450 m2

Community size

150 9 150 m2

Number of grids

9

Number of communities

4

Number of nodes

40, 50, 60, 70, (40)

Communication range

10, 20, 30, 40, 50, (10) m

Velocity of nodes

1 * 9 m/s

Warm-up period

1,000 s

Cluster level

1, 2, 3, 4, 5, 6, 7, 8, 9, 10, (5)

Simulation time

8,000 s

5 Experimental results 5.1 Simulation environments In the experiment, we use the network simulator NS-2 v2.35 [33] for the simulations. The network area is set to 450 m 9 450 m, and the community size is 150 m 9 150 m. The number of grids is 9. The number of communities among the grids is 4, and each community has 10 or more nodes. The number of nodes is set to 40, 50, 60 or 70. The communication range varies from 10 to 50 m. The movement of a node follows HCMM [15], which is a frequently used movement pattern in MSN simulations. Because the node contact information is not available in HCMM, a warm-up period is required to obtain estimated social (global) information in order to utilize the socialaware forwarding schemes. The velocity of a node ranges from 1 to 9 m/s, which is appropriate for either people or vehicles. In our simulator, a mobile node issues one message to a random destination right after the warm-up period. After a source node transmits its message to other nodes, it still keeps the message. The warm-up period is set to 1,000 s for collecting enough node contact information. Both SimBet and PRoPHET also need the warm-up period to obtain the information required. In particular, the cluster level varies from 1 to 10 for constructing hierarchical clusters based on our extensive experiments. The total simulation time is 8,000 s. We run each scheme 20 times and determine the average results. Table 2 summarizes the parameters used in our simulation. All simulation environments are as in [23]. We evaluated the proposed scheme with the following performance metrics: 1. 2. 3.

Delivery ratio: Ratio of the number of delivered messages to the total number of messages issued. Network traffic: Total number of messages sent and received. Delay: Time required for a message to travel from the source to destination nodes.

123

We do not consider the network traffic during the warmup period, since most schemes in MSNs ignore network traffic [24, 29]. We found that the amount of traffic in the warm-up period is not big enough to affect the total amount of traffic. For 8,000 s, we simulate and compare the proposed scheme HCS with non-clustering schemes such as Epidemic, Wait, SimBet, PRoPHET, and common neighbor similarity schemes. During this period, all the schemes except Wait achieve 1.0 delivery ratio. Epidemic and Wait, which are typical social-oblivious schemes, Epidemic has the highest network traffic and the lowest transmission delay, while Wait shows the highest transmission delay and the lowest network traffic. In the rest of the schemes, SimBet, which is the centrality/similarity-based scheme, uses the betweenness centrality. PRoPHET, which is the predictability-based scheme, takes advantages of contact probability. During the warm-up period, SimBet collects common neighbor similarities and betweenness centralities of nodes, obtains the SimBet utility values by combining both metrics. PRoPHET calculates the delivery predictability P(a, b) during the warm-up period. In both schemes, each node forwards a message to another node with higher SimBet utility or higher delivery predictability for the destination, respectively. Common neighbor similarity scheme uses only common neighbor similarity that HCS uses for clustering. Hence, we compare these schemes against HCS. 5.2 Simulation results 5.2.1 Effect of cluster level We examine the performance of HCS at different cluster levels when the communication range is 10 m. Figure 5 shows network traffic and transmission delay of HCS with various cluster levels. The cluster level varies from 1 to 10. As shown in Fig. 5(a), when the level is low, HCS shows higher network traffic. However, as the level increases,

Wireless Netw (2015) 21:1699–1711

1707

Fig. 5 Effect of cluster level.(a) Network traffic (b) Transmission delay

Fig. 6 Delivery ratios and network traffic by time. (a) Delivery ratio (b) Network traffic

HCS increasingly reduces network traffic. Such a phenomenon arises because HCS utilizes clustering information very well at high clustering levels. On the other hand, as shown in Fig. 5(b), as the level becomes higher, HCS shows longer transmission delay because it requires longer time to find the nodes with lower levels. However, because HCS appropriately uses both level-based and similarity-based forwarding, its traffic amount is still smaller than those of the non-clustering schemes. It is evident that a proper value of d in hierarchical clustering for this environment is 5, because we get relatively lower traffic as well as acceptable delay. However, d should be appropriately chosen according to the given environment.

5.2.2 Delivery ratios and network traffic by time Figure 6 shows the delivery ratios and network traffic as the simulation time reaches 8,000 s. The results are shown after 1,000 s in order not to include the results of the warm-up period. The number of nodes is set to 40. In Fig. 6(a), the non-clustering schemes, except Wait, achieve the maximum delivery ratio that is faster than that of HCS, because they allow multiple copies of messages. Because Wait does not distribute a message but instead waits for the message to encounter its destination, it requires much longer time to reach the 1.0 delivery ratio. However, HCS uses an appropriate number of copies of a message, thereby resulting in a somewhat slower time in

123

1708

Wireless Netw (2015) 21:1699–1711

Fig. 7 Effect of the number of nodes. (a) Delivery ratio. (b) Network traffic. (c) Transmission delay

reaching the 1.0 delivery ratio compared with other schemes. For network traffic shown in Fig. 6(b), each scheme experiences higher network traffic as time passes. HCS shows much lower traffic than the non-clustering schemes except Wait. Note that HCS distributes the messages to the nodes with lower levels or higher similarity scores. 5.2.3 Effect of number of nodes We evaluate the performance of the schemes as the number of nodes increases. Figure 7(a) shows the average delivery ratio with 40–70 nodes. Figure 6(a) compares the delivery ratios of the schemes; the non-clustering schemes, except Wait, reach a 1.0 delivery ratio faster than HCS, because HCS maintains only an appropriate number of copies of a

123

message. Figure 7(b) shows the network traffic when the number of nodes increases. As expected, the amount of traffic in the non-clustering schemes explosively increase as the number of nodes increases. In particular, the difference between HCS and each of the non-clustering schemes is large. HCS shows the lowest traffic except for Wait. In HCS, as the number of nodes increases, so does the number of ‘‘valuable’’ nodes that play a critical role in message delivery. Figure 7(c) shows transmission delay. The delays of most schemes decrease as the number of nodes increases. However, Wait demonstrates similar patterns regardless of the number of nodes. The delay in HCS is higher than those of the non-clustering schemes because of their multiple message copies. However, HCS well maintains the balance between network traffic and transmission delay.

Wireless Netw (2015) 21:1699–1711

1709

Fig. 8 Effect of communication range. (a) Delivery ratio, (b) Network traffic, (c) Transmission delay

5.2.4 Effect of communication range Finally, we compare the effect of the communication range of the nodes for each scheme. The number of nodes is set to 40. Figure 8(a) shows the average delivery ratios with 10–50 m of the communication range. In Figs. 6(a) and 7(a), the delivery ratios of HCS increase little bit slower than those of other schemes except Wait because HCS maintains a smaller number of copies even if the communication range increases. As shown in Fig. 8(b), as the communication range becomes wider, all the schemes except Wait experience increased network traffic. PRoPHET shows a moderate increase because only the nodes with higher contact probability to the destination participated in the delivery of the messages. However, HCS shows the lowest traffic except Wait when the

communication range is 10 because the nodes that are involved in message delivery are properly chosen. Such a result confirms that HCS can be well implemented in a sparse network. Figure 8(c) shows the transmission delays of the schemes. It is natural that most schemes incur shorter delays as the communication range increases. Epidemic shows the shortest delay. SimBet and Common exhibit similar results. However, PRoPHET suffers from a longer delay compared with HCS, except when the communication range is 10. The transmission delays of HCS and Wait significantly decrease as the range increases. In HCS, the percentages of level-based forwarding and similarity-based forwarding during the entire forwarding processes are 91.3 and 8.7 %, respectively. It is evident that hierarchical clustering on social similarity in HCS is well constructed; therefore, relay nodes are properly chosen with the

123

1710

clustering information. Hence, the transmission delay of HCS becomes significantly lower as the communication range becomes larger. The results in both figures demonstrate that HCS maintains a well-balanced performance between network traffic and transmission delay.

6 Conclusion In MSNs, non-clustering schemes suffer from higher network traffic. To reduce network traffic, we proposed a forwarding scheme using hierarchical clustering based on social information. The proposed scheme effectively distributes the messages to the nodes via shortcuts by exploiting clustering results. Experimental results demonstrated that the scheme reduced network traffic compared to non-clustering schemes, while delay was acceptable compared to the Wait scheme. In future work, we plan to study more enhanced dynamic forwarding schemes with varying resources in consideration of continuous information updating in MSNs. Acknowledgments This research was supported by the Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education, Science and Technology (2013R1A1A2011114).

References 1. Conti, M., Giordano, S., May, M., & Passarella, A. (2010). From opportunistic networks to opportunistic computing networks. IEEE Communications Magazine, 48(9), 126–139. 2. Hui, P., Chaintreau, A., Scott, J., Gass, R., Crowcroft, J., & Diot, C. (2005). Pocket switched networks and human mobility in conference environments. In Proceedings of ACM SIGCOMM Workshop on Delay-Tolerant Networking (WDTN 2005) (pp. 244–251). 3. Lindgren, A., & Doria, A. (2007). Experiences from deploying a reallife dtn system. In Proceedings of IEEE Consumer Communications and Networking Conference (CCNC) (pp. 217–221). 4. Juang, P., Oki, H., Wang, Y., Maronosi, M., Peh, L., & Rubenstein, D. (2002). Energy-efficient computing for wildlife tracking: Design tradeoffs and early experiences with zebraNet. In Proceedings of Architectural Support for Programming Languages and Operating Systems (ASPLOS) (pp. 96–107). 5. Small, T., & Haas, Z. J. (2003). The shared wireless infestation model: A new ad hoc networking paradigm (or where there is a whale, there is a way). In Proceedings of MobiHoc (pp. 234–244). 6. Zeng, Y., Xiang, K., Li, D., & Vaslakos, A. V. (2013). Directional routing and scheduling for green vehicular delay tolerant networks. Wireless Networks, 19(2), 161–173. 7. Zafar, H., Alhamahmy, N., Harle, D., & Andonovic, I. (2011). Survey of reactive and hybrid routing protocols for mobile ad hoc networks. Communication Networks and Information Security, 3(3), 193–202. 8. Spyropoulos, T., Rais, R. N., Turletti, T., Obraczka, K., & Vasilakos, A. (2010). Routing for disruption tolerant networks: Taxonomy and design. Wireless Networks, 16(8), 2349–2370.

123

Wireless Netw (2015) 21:1699–1711 9. Hui, P., Xu, K., Li, V., Crowcroft, J., Latora, V., & Lio, P. (2009). Selfishness, altruism and message spreading in mobile social networks. In Proceedings of IEEE INFOCOM Workshop (pp. 1–6). 10. Conti, M., & Kumar, M. (2010). Opportunities in opportunistic computing. IEEE Computer, 43(1), 42–50. 11. C. Bordrini, M. Conti, and A. Passarella (2011). Modeling socialaware forwarding in opportunistic networks. Performance Evaluation of Computer and Communication Systems, 6821 (LNCS), 141–152. 12. Dang, H., & Wu, H. (2010). Clustering and cluster-based routing protocol for delay-tolerant mobile networks. IEEE Transactions on Wireless Communications, 9(6), 1874–1881. 13. Liu, C., & Wu, J. (2009). Scalable routing in delay tolerant networks. In Proceedings of MobiHoc (pp. 51–60). 14. Cambridge University (2009). Online edition (c), chapter hierarchical clustering, pp. 377–401. 15. Boldrini, C., & Passarella, A. (2010). HCMM: Modelling spatial and temporal properties of human mobility driven by users’ social relationships. Computer Communications, 33(9), 1056–1074. 16. Wu, J., Xiao, M., & Huang, L. (2013). Homing spread: Community home-based multi-copy routing in mobile social networks. In Proceedings of IEEE INFOCOM (pp. 138–146). 17. Kim, J. H., Lee, K. J., Kim, T. H., & Yang, S. B. (2011). Effective routing schemes for double-layered peer-to-peer systems in manet. Journal of Computing Science Engineering, 5(1), 19–31. 18. Vahdat, A., & Becker, D. (2000). Epidemic routing for partiallyconnected ad hoc networks. Technical Report CS-2000-06. 19. Spyropoulos, T., Psounis, K., & Raghavendra, C. (2005). Spray and wait: An efficient routing scheme for intermittently connected mobile networks. In Proceedings of ACM SIGCOMM (pp. 252–259). 20. Hua, D., Du, X., Qian, Y., & Yan, S. (2009). A DTN routing protocol based on hierarchy forwarding and cluster control. In Proceedings of IEEE Computational Intelligence and Security (pp. 397–401). 21. Alresaini, M., Sathiamoorthy, M., Krishnamachari, B., & Neely, M. J. (2012). Backpressure with adaptive redundancy. In Proceedings of IEEE INFOCOM (pp. 2300–2308). 22. Dvir, A., & Vasilakous, A. V. (2010). Backpressure-based routing protocol for DTNs. In Proceedings of ACM SIGCOMM (pp. 405–406). 23. Kim,S. K., Choi, J. H., & Yang, S. B. (in press). Hotspot: Location-based forwarding scheme in an opportunistic network. Ad hoc & Sensor Wireless Networks. 24. Daly,E. H., & Haahr, M. (2007). Social network analysis for routing in disconnected delay-tolerant manets. In Proceedings of ACM MobiHoc (pp. 32–40). 25. Hui, P., Crowcroft, J., & Yoneki, E. (2011). Bubble rap: Socialbased forwarding in delay tolerant networks. IEEE Transactions on Mobile Computing, 10(11), 1576–1589. 26. Mei, A., Morabitto, G., Santi, P., & Stefa, J. (2011). Social-aware stateless forwarding in pocket switched networks. In Proceedings of IEEE INFOCOM (pp. 251–255). 27. Hui, P., & Crowcroft, J. (2007). How small labels create big improvements. In Proceedings of Pervasive Computing and Communications Workshop (pp. 65–70). 28. Boldrini, C., Conti, M., & Passarella, A. (2008). Exploiting users’ social relations to forward data in opportunistic networks: The HiBop solution. Pervasive and Mobile Computing, 4(5), 633–657. 29. Lindgren, A., Doria, A., & Schelen, O. (2004). Probabilistic routing in intermittently connected networks. Service Assurance with Partial and Intermittent Resources, 3126 (LNCS), 239–254.

Wireless Netw (2015) 21:1699–1711 30. Mtibaa, A., May, M., Diot, C., & Ammar, M. (2010). PeopleRank: Social opportunistic forwarding. In Proceedings of IEEE INFOCOM (pp. 1–5). 31. Leguay, J., Friedman, T., & Conan, V. (2007). Evaluating MobySpace based routing strategies in delay tolerant networks. Wireless Communications and Mobile Computing, 7(10), 1171–1182. 32. Newman, M. E. J. (2001). Clustering and preferential attachment in growing networks. Physical Review Letters, E 64, 025102. 33. Network simulator-2. http://www.isi.edu/nsnam/ns/

Sun-Kyum Kim received his M.S. in computer science from Yonsei University in Korea in 2012. He is currently a Ph.D. candidate at Yonsei University. His research interests include mobile social networks, delay tolerant networks and social network analysis.

1711 Junyeop Lee is currently an M.S. candidate in computer science at Yonsei University in Korea. His research interests include mobile social networks, delay tolerant networks and social network analysis.

Sung-Bong Yang received his M.S. and Ph.D. from the Department of Computer Science at the University of Oklahoma in 1986 and 1992, respectively. He has been a professor at Yonsei University since 1994. His research interests include graph algorithms, mobile computing, and social network analysis.

Ji-Hyeun Yoon is currently an Ph.D. candidate in computer science at Yonsei University in Korea. His research interests include mobile social networks, delay tolerant networks and social network analysis.

123