Impact of Human Behavior on Social Opportunistic Forwarding
aa r X i v : . [ c s . N I] A ug Impact of Human Behavior on Social OpportunisticForwarding
Waldir Moreira a, ∗ , Paulo Mendes a a COPELABS, University Lusofona, Lisbon, Portugal
Abstract
The current Internet design is not capable to support communications in envi-ronments characterized by very long delays and frequent network partitions. Toallow devices to communicate in such environments, delay-tolerant networkingsolutions have been proposed by exploiting opportunistic message forwarding,with limited expectations of end-to-end connectivity and node resources. Suchsolutions envision non-traditional communication scenarios, such as disaster ar-eas and development regions. Several forwarding algorithms have been investi-gated, aiming to offer the best trade-off between cost (number of message repli-cas) and rate of successful message delivery. Among such proposals, there hasbeen an effort to employ social similarity inferred from user mobility patterns inopportunistic routing solutions to improve forwarding. However, these researcheffort presents two major limitations: first, it is focused on distribution of theintercontact time over the complete network structure, ignoring the impact thathuman behavior has on the dynamics of the network; and second, most of theproposed solutions look at challenging networking environments where networkshave low density, ignoring the potential use of delay-tolerant networking to sup-port low cost communications in networks with higher density, such as urbanscenarios. This paper presents a study of the impact that human behavior has ∗ This is the author’s pre-print version. Personal use of this material is permitted. However,permission to reprint/republish this material for advertising or promotion or for creating newcollective works for resale or for redistribution to thirds must be obtained from the copyrightowner. The camera-ready version of this work is being published at Elsevier Ad Hoc Networks,2014 and is property of Elsevier B.V.
Email address: [email protected] (Waldir Moreira)
Preprint submitted to Ad Hoc Networks August 4, 2014 n opportunistic forwarding. Our goal is twofold: i) to show that performancein low and high density networks can be improved by taking the dynamics ofthe network into account; and ii) to show that the delay-tolerant networkingcan be used to reduce communication costs in networks with higher density bytaking the behavior of the user into account.
Keywords: opportunistic networks, delay/disruption-tolerant networks,social-networking communications, human dynamics, application awareness,challenging environments [2010] 00-01, 99-00
1. Introduction
Wireless devices have become more portable and with increased capabili-ties (e.g., processing, storage), which is creating the foundations for the de-ployment of pervasive wireless networks, encompassing personal devices (e.g.smartphones and tablets). Additionally, wireless technology has been extendedto allow direct communication: vehicle-to-vehicle - for safety information ex-change; device-to-device - aiming at 3G offloading; Wi-Fi direct - overcome theneed for infrastructure entities (i.e., access points).The combination of pervasive wireless devices and direct wireless communi-cation solutions can be used to support the deployment of two major type ofapplications: end-to-end communication in development regions, since today’sInternet routing protocols may operate poorly in such environments, charac-terized by very long delay paths and frequent network partitions; and low costcommunication, namely data sharing, in urban scenarios, to bypass expensivedata mobile communications and the unreliable presence of open Wi-Fi accesspoints.These networking scenarios (from development regional to large urban sce-narios) are characterized by network graphs with different densities, which posedifferent challenges in terms of data forwarding. The challenge that we aim totackle in this paper is to investigate the impact of human behavior on oppor-2unistic forwarding, namely the awareness about users’ social and data similar-ities.Most of the prior art has been studying data transfer opportunities betweenwireless devices carried by humans, by looking at the distribution of the in-tercontact time, which is the time gap separating two contacts between thesame pair of devices [1]. In challenging networking environments, opportunis-tic contacts among mobile devices may improve communications among peersas well as content dissemination, mitigating the effects of network disruption.This gave rise to the investigation of opportunistic networks, of which Delay-Tolerant Networks (DTN) are an example, encompassing different forwardingproposals to quickly send data from one point to another even in the absence ofan end-to-end path between them. Such proposals range from flooding content[2] in the network up to solutions that take into account the social interactionsamong users [3, 4, 5, 6, 7, 8]. In the latter case, wireless contacts are aggregatedinto a social graph, and a variety of metrics (e.g., centrality and similarity) oralgorithms (e.g., community detection) have been proposed to assess the utilityof a node to deliver a content or bring it closer to the destination. Nevertheless,the structure of such graphs is rather dynamic, since users’ social behavior andinteractions vary throughout their daily routines. This brings us to our firstassumption: forwarding algorithms should be able to exploit social graphs thatreflect people’s dynamic behavior. Prior art have studied forwarding algorithmsthat consider only the global network structure, without taking people’s behav-ior into account [9]. In this paper, we show that forwarding algorithms thatexploit social graphs reflecting the variations in people’s daily routines are ableto improve the performance of social-aware opportunistic networking.Our second challenge was to analyze how to expand the deployment of DTNtechnology, which is normally seen only as useful to allow communications inchallenging environments, such as development regions. For this study, we focuson data sharing since this should be the most interesting application to takeadvantage of low cost communication in dense networks, such as in an urbanscenario. In this case, we studied two hypothesis to ensure good performance3hen the density of the network increases: i) forwarding based on social graphs,where aggregation is based only on social similarities; and ii) forwarding basedon behavior graphs, where aggregation is done by combining different aspects ofhuman behavior, such as social similarities and data similarities (derived fromthe interests that users demonstrate in specific type of data).Hence, in this paper we aim to investigate the possibility of developing anopportunistic forwarding system able to support low-cost services in dense net-working scenarios as well as basic services in extreme networking conditions, byexploiting social as well as data similarities among users. Our work shows whichtype of opportunistic forwarding scheme is more suitable for delay-tolerant ap-plications, based on the density of the network in scenarios spanning from devel-oping regions to urban environments. Our findings lead to a new research chal-lenge aiming to expand the impact of DTNs: the investigation of self-awarenessmechanisms able to adapt their forwarding schemes based on the context of theuser, namely the density of the network where he/she is currently.The remainder of the paper is structured as follows. Section 2 aims to moti-vate our work, namely in what concerns the goal to study methods to expand thedeployment of DTNs, and the impact that a better understanding of human be-havior can have in the development of efficient forwarding solutions. In Section3 we present our definition of network density based on the deployment scenariosthat we look at to pursuit our study and experiments. Section 4 presents a set offorwarding algorithms that are considered in our study, including our proposals.In Section 5, we show the performance results of opportunistic forwarding overdifferent network densities. Section 6 concludes our work, and identifies futureresearch challenges to expand the impact of DTNs.
2. Motivation
The growing number of mobile devices equipped with a wireless interfaceand the end-user trend to shift toward wireless technology are opening new pos-sibilities for networking. In particular, opportunistic communication embodies4 feasible solution for environments with scarce or costly infrastructure-basedconnectivity. A lot of attention has been given to the development of oppor-tunistic forwarding solutions for networks with scarce connectivity, which areconsidered a natural fit for DTN technology. However, it is our belief thatopportunistic forwarding can also be applied to more dense networks, were In-ternet communications are expensive, or applications aim to take advantage ofdirect communications among people.In what concerns challenging networks, the most common approach has beento make use of social similarities to improve performance over the overall net-work. In this case, our work is motivated by the fact that such approachesignore the behavior patterns that people present in their daily routines [10],which may lead to further performance improvements.In what concerns the application of DTN and opportunistic forwarding todense networks, most of the prior art aims to implement a store-carry-and-forward communication model that exploits specific devices found in urban sce-narios, such as buses [11] and cars [12]. It is our understanding that the devel-opment of opportunistic forwarding solutions should not depend upon specificequipments only found is some scenarios, since this mitigates the deploymentexpansion of such proposals. It is our belief that the success of the DTN tech-nology depends on its deployment range, which can only be ensured if suchtechnology is based on pervasive wireless devices, such as smartphones: thesedevices are present in development regions as well as urban scenarios. In thelatter case, communications between smartphones can also exploit mobility pat-terns of different vehicles ridden by people.In order to design useful applications, it is vital to have a good understand-ing of the target environment and its users. Different types of user behaviormay result in different network conditions and shall have a huge impact onwhether or not a particular application is of interest to the user. A fair amountof work has been done on studying human mobility traces in order to gain un-derstanding of real life mobility patterns and how those affect the propertiesof the opportunistic networks that are possible in that environment [13, 1, 14].5lthough mobility patterns are important properties of the network, it is alsoimportant to understand the impact the human behavior, such as data interests,have on these networks. Hence, our work aims to tackle this new research trend,expanding social awareness to human behavior awareness. Among the differenthuman behavior metrics that can be considered, we focus our attention on datasimilarities since data sharing is the most common application in the Internet.The study of data similarities depends on which applications are in place in thenetwork and how the users use them. Usage patterns also depend on the users’context, so the same data patterns do not apply to all users. Approximations ofsome use cases might be possible to derive from the way cellular networks areused, but that will most likely not be applicable to all types of applications.Looking at data similarities may improve the performance of opportunisticforwarding [3, 4]. However it is not clear if the improvement is higher thanexploiting social interactions and structure (i.e., communities [5], as well aslevels of social interaction [6, 7]). Thus, combining social and data similaritiesshall bring benefits (i.e., faster, better content reachability) to opportunisticforwarding. Hence, in this work we aim to show when the exploitation of socialsimilarities results in a good performance, and when such performance can beaugmented by combining them with data similarity metrics.
3. Network Scenario Characterization
One can observe that a networking scenario may vary according to its density.Sparse scenarios are characterized by very long delays (e.g., space communica-tions [15]) and communication suffers with frequent disruption mostly due tothe lack of infrastructure and geographic location (e.g., rural areas [16], river-side communities [17, 18]). It is common to see solutions relying on messageferries [19] or data mules [20] as to overcome the missing infrastructure. This isa classic scenario whose challenges are more related to transport protocols (i.e.,dealing with extremely high delays) than routing itself.6 able 1: Network densities of the considered scenarios
Scenario Cambridge MIT SyntheticIdentified density
Shortest Path Map BasedMovement model (i.e., nodes randomly choose destinations and use the shortestpath to reach them). With the Gephi v0.8.2 [23] analysis tool, we accounted forthe network densities of these scenarios summarized in Table 1.7able 1 displays the scenarios in increasing order of density. Thus, it isexpected that the routing solutions have an increasing performance behavior asnetwork density increases. This is due to the different contact opportunities thata node may have, which increase (and therefore can be beneficial) for routingpurposes.
4. Opportunistic Forwarding in Wireless Networks
This section presents the most relevant and latest opportunistic forwardingproposals, considering whether they make use of social and/or data similaritymetrics. Similarity metrics are used to build graphs over which such forward-ing proposals operate [24]. That is, instead of considering the number andfrequency of contacts due to the mobility of hosts, such approaches take intoaccount more stable social (e.g., common social groups and communities, nodepopularity, levels of centrality, social relationships and interactions, user profiles)and/or data (e.g., shared interests, interest of users in the content traversingnetwork, content availability, type of content) aspects, aiming to reduce the costof opportunistic forwarding. Moreover, opportunistic forwarding proposals maytake into account the dynamics of user behavior, i.e., the resulting social graphsmay consider what happens in terms of social interactions throughout the dailyroutine of the users.Table 2 summarizes the type of similarity (i.e., social and/or data) consideredby the opportunistic forwarding proposals and whether (or not) they considerthe observed user behavior to build dynamic social graphs.
Bubble Rap [5],
CiPRO [7],
SocialCast [3], and
ContentPlace [4] belong tothe category that considers social similarity and/or data similarity metrics, butdoes not suitably reflect the dynamism of user behavior in the underlying socialgraph.
Bubble Rap combines node centrality with the notion of community to makeforwarding decisions. The centrality metric identifies hub nodes inside (i.e.,local) or outside (i.e., global) communities. Messages are replicated based on8 able 2: Opportunistic Forwarding Proposals
Proposals Social similaritymetrics Data similaritymetrics Dynamicgraphs
Bubble Rap Communities andcentralityCiPRO User profileSocialCast Shared interestsContentPlace Social relationshipand communities Interest on the contentdLife Social weight andnode importance X SCORP Social weight Content type andinterest on the content X global centrality until they reach the community of the destination host (i.e., anode belonging to the same community). Then, it uses the local centrality toreach the destination inside the community. CiPRO considers the time and place nodes meet throughout their routines.
CiPRO holds knowledge of nodes (e.g., carrier’s name, address, nationality, ...)expressed by means of profiles that are used to compute the encounter prob-ability among nodes in specific time periods. Nodes that meet occasionallyget a copy of the message only if they have higher encounter probability to-wards its destination. If nodes meet frequently, history of encounters is used topredict encounter probabilities for efficient broadcasting of control packets andmessages.
SocialCast considers the interest shared among nodes. It devises a utilityfunction that captures the future co-location of the node (with others sharingthe same interest) and the change in its connectivity degree. Thus, the utilityfunction measures how good message carrier a node can be regarding a giveninterest.
SocialCast functions are based on the publish-subscribe paradigm,where users broadcast their interests, and content is disseminated to interestedparties and/or to new carriers with high utility.
ContentPlace considers information about the users’ social relationships toimprove content availability. It computes a utility function for each data object9onsidering: i) the access probability to each object and the involved cost inaccessing it; ii) the social strength of the user towards the different communitieswhich he/she belongs to and/or has interacted with. The idea is having theusers to fetch data objects that maximize the utility function with respect tolocal cache limitations, and choosing those objects that are of interest to usersand can be further disseminated in the communities they have strong social ties.The next category considers solely social similarity metrics and take intoaccount the dynamism of user behavior while building the underlying socialgraph. dLife [6] is in this category and it takes into account the dynamism ofusers’ behavior found in their daily life routines to aid forwarding. The goal isto keep track of the different levels of social interactions (in terms of contactduration) nodes have throughout their daily activities in order to infer how wellsocially connected they are in different periods of the day. Forwarding takesplace by considering either the social strength (i.e., weight) among users ortheir importance in specific time periods.Finally, the last category comprises both social and data similarity metricsand the dynamism observed in the behavior of users.
SCORP [8] belongs to thiscategory. It considers the type of content and the social relationship between theparties interested in such content type.
SCORP nodes are expected to receiveand store messages considering their own interests as well as interests of othernodes with whom they have interacted before. Data forwarding takes place byconsidering the social weight of the encountered node towards nodes interestedin the message that is about to be replicated.For the remainder of this paper, we consider one representative from each ofthe described categories:
Bubble Rap , for being solely based on social similaritymetrics; dLife and
SCORP , for considering social and data similarity metricsand for being the proposals which satisfactory capture the dynamic user be-havior in the resulting social graphs. These proposals are enough to help usillustrate how the performance of opportunistic routing proposals in networkswith different densities can be further improved by considering the user dynamicbehavior. 10 . Evaluation of Opportunistic Forwarding over Different NetworkDensity Scenarios
In this section, we analyze the performance behavior of
Bubble Rap , dLife, and SCORP over different network density scenarios. With the experiments inthis section, we want i) to show that performance in low (Sec. 5.2) and high(Sec. 5.3) density networks can be improved by taking the dynamics of thenetwork into account; and ii) to show that the delay-tolerant networking can beused to reduce communication costs in networks with higher density by takingthe behavior of the user into account (Sec. 5.4).
Simulations are carried out in the Opportunistic Network Environment (ONE)simulator [25]. Results are presented with a 95% confidence interval and interms of averaged delivery probability (i.e., ratio between the number of deliv-ered messages and the number of messages that should have been delivered),cost (i.e., number of replicas per delivered message), and latency (i.e., timeelapsed between message creation and delivery).The trace scenarios comprise 36 (Cambridge) and 97 (MIT) nodes carryingdevices during their daily activities. The synthetic mobility scenario simulates3 groups ( A , M , and B ) of 50 people each, who carry nodes equipped with 250-Kbps Bluetooth interfaces, and moving with speed up to 1.4 m/s. The reasonfor considering traces and synthetic mobility scenario relates to the fact that: i)with the former, we have a representation of real user behavior; and ii) with thelatter, we are able to have a network density much higher from the perspectiveof the user, as defined in Sec. 3. Most analysis have been based on datasetswith low density, collected in a constrained setting, which is not representativefor realistic use cases of the networks being studied. If one is interested in theproperties of a large scale urban environment, it is probably not meaningful tostudy traces collected from 36 or 97 user at a conference or university campus.Across all experiments, proposals experience the same load and number ofmessages that must reach the destinations. In the Cambridge trace (cf. Sec.11.2), the Bubble Rap / dLife source sends 1, 5, 10, 20 and 35 different messages toeach of the 35 destinations, while the SCORP source creates 35 messages withunique content types, and the receivers are configured with 1, 5, 10, 20, and 35randomly assigned interests. Thus, we have a total of 35, 175, 350, 700, and 1225generated messages. The msg/int notation represents the number of messagessent by
Bubble Rap and dLife sources, or the number of interests of each of the
SCORP receivers. Since
Bubble Rap / dLife sources generate more messages, inthis scenario node 0 (the source) has no buffer restriction and message generationvaries with the load: 35 messages/day rate (load of 1, 5, and 10 messages), and70 and 140 messages/day rates (load of 20 and 35 messages, respectively).As for the synthetic mobility scenario (cf. Sec. 5.3), 200 messages are gen-erated. With Bubble Rap and dLife , node 0 (group A ) generates 100 messagesto nodes in groups B and M , and node 100 (group B ) generates 100 messagesto nodes in groups A and M . For SCORP , each group has different interests:group A ( reading ), group B ( games ), and group M ( reading and games ). Thesource nodes, 0 and 100, generate only one message for each content type, game and reading . This guarantees the same number of messages expected to bereceived, i.e., 200. Also, by varying the node pause times between 100 and100000 seconds, we have different levels of mobility (varying from 3456 to 3.4movements in the simulation). In this scenario, all source nodes have restrictedbuffer, but rate is of 25 messages every 12 hours. This is done so that BubbleRap / dLife do not discard messages prior to even trying exchange/deliver themgiven the buffer constraint.Finally, in Sec. 5.4, the load generated is equivalent to 6000, 78000, and200 messages to be delivered across all experiments for Cambridge, MIT, andsynthetic mobility scenarios, respectively.Regarding message TTL, we set it to be unlimited in order to observe theperformance behavior (i.e., buffer consumption, number of replicas) of the for-warding proposals in networks with high traffic load. Message size ranges from1 to 100 kB. Despite nodes may have plenty of storage, we consider nodes hav-ing different capabilities (i.e., smartphones). Thus, nodes have buffers limited12o 2 MB as we consider that nodes may not be willing to share all their stor-age space. The performance evaluation follows the guidelines of a UniversalEvaluation Framework (UEF) [26 ? ] to guarantee fairness in the assessment.As for proposals, Bubble Rap uses the K-Clique and cumulative windowalgorithms for community formation and centrality computation as in [5]. Asfor dLife and
SCORP , both consider 24 daily samples (i.e., each of one hour)as mentioned in [8].
This section presents the performance of opportunistic forwarding propos-als over a low density network scenario. Fig. 1 presents the average deliveryprobability with different messages and interests being generated. % Bubble RapdLifeSCORP
Figure 1: Delivery under different network loads
In the 1 msg/int configuration, formed communities comprise almost allnodes. This means that each node has high probability to meet any othernode, which is advantageous for
Bubble Rap since most of its deliveries happento nodes sharing communities. Due to the dense properties of the network, dLife and
SCORP take advantage of direct delivery: 57% and 51% of messages,respectively, are delivered directly to destinations.As load increases,
Bubble Rap has an 50% decrease in delivery performance.This occurs since it relies on communities to perform forwardings, and conse-13uently buffer space becomes an issue. To support this claim, we estimate bufferusage for the 5 msg/int configuration: there is an average of 80340.7 forwardings,and if this number is divided by the number of days (12 ) and by the numberof nodes (35, source not included), we get an average of 191.28 replications pernode. Multiplied by the average message size (52kB), the buffer occupancy isroughly 9.94 MB in each node, which exceeds the 2MB allowed (cf. Sec. 4.1).This estimation is for a worst case scenario, where Bubble Rap spreads copiesto every encountered node. Since this cannot happen, as
Bubble Rap also re-lies on local centrality to reduce replication, buffer exhaustion is really an issuegiven that messages are replicated to fewer nodes and not to all as in our es-timation. As more messages are generated, replication increases: this causesthe spread of messages that potentially take over forwarding opportunities fromother messages, reducing
Bubble Rap ’s delivery capability. dLife has a 43% performance decrease when network load increases, as ittakes time to have an accurate view of the social weights. This leads to for-wardings that never reach destinations given the contact sporadicity. For the10 msg/int configuration, dLife also experiences buffer exhaustion: estimatedconsumption is 2.17 MB per node. Still, by considering social weights or nodeimportance allows dLife a more stable behavior than
Bubble Rap .Since content is only replicated to nodes that are interested in it, or have astrong social interaction with other nodes interested in such content, the deliverycapability of SCORP raises as the ability of nodes to become a good carrierincreases (i.e., the more interests a node has, the better it is to deliver contentto others, since they potentially share interests). The maximum estimated bufferconsumption of
SCORP is of 0.16 MB (35 msg/int).Fig. 2 presents the average cost behavior. In the 1 msg/int configuration, allproposals create very few replicas to perform a successful delivery, 7.95 (
BubbleRap ), 14.32 ( dLife ), and 23.46 (
SCORP ), as they rely mostly on shared com-munities and/or direct deliveries. We also observe that
SCORP produces more In simulation it is worth ˜12 days of communications. dLife due to a particularity in its implementation:
SCORP nodeswith interest in a specific content of a message not only process it, but alsoreplicate it to other interested nodes, thus creating extra replicas.For the 5, 10, 20 and 35 msg/int configurations, replication is directly pro-portional to the load. Thus, cost is expected to increase as load increases,as seen with dLife . The same performance behavior was expected for
BubbleRap . However, the observed cost peaks relate to the message creation time andcontact sporadicity: when a message is created in a period of high number ofcontacts, which results in much more replications. This is more evident with
Bubble Rap at the 5 msg/int configuration as it relies on shared communitiesto forward: as mentioned earlier, most of the communities comprise almost allnodes, which increases its replication rate.Despite their efforts, these replications do not improve their delivery prob-abilities, contributing only to the associated cost for performing successful de-liveries. o f r e p li c a s ( x ) Figure 2: Cost under different network loads
With more interests, a
SCORP node can serve as a carrier for a largernumber of nodes. Consequently, the observed extra replicas make the proposalrather efficient:
SCORP creates an average of 6.39 replicas across all msg/int15onfigurations, while
Bubble Rap and dLife produce an average 452.41 and 96replicas, respectively.Fig. 3 shows the average latency that messages experience. The latencypeak in the 1 msg/int configuration refers to the message generation time: somemessages are created during periods where very few contacts (and sometimesnone) take place followed by long periods (12 to 23 hours) with almost nocontact. Consequently, messages are stored longer, contributing to the increaseof the overall latency. This effect is mitigated as the load increases with messagesbeing created almost immediately before a high number of contacts take place. S e c o n d s ( x ) Figure 3: Latency under different network loads
Since latency is in function of the delivered messages, the decrease and vari-able behavior of
Bubble Rap and dLife is due to their delivery rates decreaseand increase, and also to their choices of next forwarders that may take longerto deliver content to destinations.
SCORP experiences latencies up to approx.90.2% and 92.2% less than
Bubble Rap and dLife , respectively. The ability of anode to deliver content increases with the number of its interests. Thus, a nodecan receive more messages when it is interested in their contents, and conse-quently becomes a better forwarder since the probability of coming into contactwith other nodes sharing similar interests is very high, thus reducing latency.16 .3. Performance over High Density Network
This section presents the performance of opportunistic forwarding proposalsover a high density network scenario.Fig. 4 presents the average delivery probability. Given the community for-mation characteristic of this scenario,
Bubble Rap relies mostly on the globalcentrality to deliver content. By looking at centrality [5], we observe very fewnodes (out of the 150) with global centrality that can actually aid in forward-ing, i.e., 19.33% (29 nodes), 10.67% (16 nodes), 21.33% (32 nodes), and 2% (3nodes) for 100, 1000, 10000, and 100000 pause time configurations, respectively.So, these nodes become hubs and given buffer constraint and infinite TTL (i.e.,messages created earlier take the opportunity of newly created ones), messagedrop is certain, directly impacting
Bubble Rap . % P ause time (seconds)Average Delivery Probability Bubble RapdLifeSCORP
Figure 4: Delivery under varied mobility rates
Given the high number of contacts, the computation of social weight andnode importance done by dLife takes longer to reflect reality: thus dLife repli-cates more and experiences buffer exhaustion. Indeed, social awareness is advan-tageous, but still not enough to reach optimal delivery rate in such conditions.Independent of the number of contacts among nodes,
SCORP can still iden-tify nodes that are better related to others sharing similar interests, reachingoptimal delivery rate for 100, 1000, and 10000 pause time configurations. By17onsidering nodes’ interest in content and their social weights,
SCORP does notsuffer as much with node mobility as dLife and
Bubble Rap .With 100000 seconds of pause time, the little interaction happening in asporadic manner (with intervals between 20 and 26 hours) affects
Bubble Rap , dLife and SCORP as they depend on such interactions to compute centrality,node importance, and social weights, as well as to exchange/deliver content.Fig. 5 presents the average cost behavior. As pause time increases, thenumber of contacts among nodes decreases, providing all solutions with the op-portunity to have a stable view of the network in terms of their social metricswith 100, 1000, and 10000 seconds of pause time. This explains the cost re-duction experienced by
Bubble Rap and dLife : both are able to identify thebest next forwarders, which results in the creation of less replicas to perform asuccessful delivery.
70 100 1000 10000 100000 o f r e p li c a s ( x ) Pause time (seconds)Average Cost
Figure 5: Cost under varied mobility rates
SCORP has a very low replication rate (average of 0.5 replicas) given itschoice to replicate based on the interest that nodes have on content and ontheir social weight towards other nodes interested in such content. When theintermediate node has an increased number of interests (i.e., by having differentinterests, the node can potentially deliver more content) as observed in Sec.5.2, replication costs are even lower. Furthermore,
SCORP suitably uses bufferspace with an estimated average occupancy of 0.03 MB per node per day.18ith 100000 seconds of pause time, as cost is in function of delivered mes-sages (and deliveries are very low, due to contact sporadicity), proposals havea low cost. S e c o n d s ( x ) Pause time (seconds)Average Latency
Figure 6: Latency under varied mobility rates
As expected (cf. Fig. 6), latency increases as node mobility decreases:encounters are less frequent, and so content must be stored for longer times.Also, the time that the social metrics take to converge (i.e., a more stableview of the network in terms of centrality, social weight, and node importance)contributes for the increase in the experienced latency. The highest increase inlatency with 100000 seconds of pause time is due to contacts happening in asporadic fashion with intervals between them of up to 26 hours, thus proposalstake much longer to perform a delivery.
This section shows how network density impacts on the performance of op-portunistic forwarding proposals. As mentioned before, we want to bring at-tention to dense scenarios found in urban settings: a panoply of heterogenousdevices that could overcome disruption by interacting directly with one anotherto improve the networking experience of users. Fig. 7 presents the averagedelivery probability with different network densities.19 % Bubble RapdLifeSCORP
Figure 7: Delivery under different network densities
As mentioned in Sec. 3, it was expected that the performance of social-awareopportunistic routing improve with the increase of network density. However, wecan observe that content-oblivious
Bubble Rap and dLife experience a decreasein performance in the MIT scenario despite of its identified density (47.01) beingalmost almost twice as the one identified in Cambridge (26.83). The reasonfor such behavior lies on the characteristics of each scenario, with MIT nodescovering a much bigger area. Despite of having a higher number of contactsbetween nodes, the MIT scenario may lead to messages reaching nodes that arenot the best forwarders, and these messages, given the unlimited TTL, may endup taking the delivery opportunity of newly created messages. This directlyaffects the performance of both
Bubble Rap and dLife.
Yet the content-oriented
SCORP overcomes such features of the MIT sce-nario since it also considers those nodes that are interested in the content beingreplicated or that are strongly related to interested parties. Unlike the content-oblivious solutions,
SCORP has a 6.14% improvement despite the challengingscenario.Performance behavior for all proposal indeed improve with higher networkdensity (148.8, Synthetic scenario). The reason is tied to the fact that a higherdensity indicates more contact opportunities for the exchange of messages.20ig. 8 presents the average cost which is expected to increase with networkdensity. This is because the more contact opportunities the scenario has, themore replicas are created by proposals. This can be easily seen with
BubbleRap , which creates an average of more than 5000 replicas to perform a successfuldelivery. o f r ep li c a s ( x )
40 50 60 Average Cost o f r ep li c a s ( x ) Figure 8: Cost under different network densities
The same cost increasing trend is observed with dLife , but in different ordersand especially for the synthetic scenario. We believe that the much highernumber of copies (up to 98% when compared to other scenarios) is related tothe mobility rate. In the Synthetic experiment nodes move a lot, which results ina high number of contacts with many nodes. Consequently, dLife takes longerto have a stable view of the network in terms of its social weights and nodeimportances, which leads to the creation of unwanted replicas.
SCORP has thebest cost performance (up to 0.5 replicas to perform a successful delivery), sincethe more interests a node has, the better forwarder it is. The M group (cf. Sec.5.1) accounts for 50% of the interests existing in the network, which makes it agreater carrier for messages.Fig. 9 shows the average latency experienced by messages since their creationup to their reception at destination. In the traces experiments, all proposalskeep the same trend: average latency increases. This is due to the fact that21odes in these experiments span different areas and the encounter frequencyhappens at different rates. This can result in messages being forwarded tonodes who may reach a given destination, but delivery time increases with thearea and duration of experiments. SCORP presents a much higher increase inthe MIT experiment, as it takes its time to suitably choose the next forwarders(based on their interest on the message’s content or social relationship to otherinterested parties). This added to fact that interactions among nodes happenaccording to area they move jointly contribute to such latency peak. S e c ond s ( x ) Figure 9: Latency under different network densities
As for the Synthetic mobility experiment, not only the proposals have manydifferent contact opportunities, but also nodes encounter more frequently. Thisconsequently has a positive impact for
Bubble Rap , dLife , and SCORP that areable to deliver content in less time (12486s, 11710s , and 6864s, respectively).
6. Conclusions
Opportunistic forwarding can aid communication in two major applicationscenarios: end-to-end communication in development regions and low cost com-munication in urban scenarios. Such scenarios do have different network den-sities which adds more challenges to opportunistic forwarding. The underlyinggraphs, over which these opportunistic forwarding proposals operate, comprise22e.g., common social groups and communities, node popularity, levels of central-ity, social relationships and interactions, user profiles) and/or data (e.g., sharedinterests, interest of users in the content traversing network, content availabil-ity, type of content) aspects. Additionally, such graphs may (or not) take intoaccount the dynamics of user behavior.Thus, in this paper, we exploit the possibility of having an opportunisticforwarding system that can provide support to i) low-cost services in dense net-working scenarios; and ii) basic services in extreme networking conditions (e.g.,communications in development regions), considering social and data similar-ities among users as well as the dynamic behavior found in the users’ dailyroutines.Our results show that opportunistic forwarding, based on social and datasimilarity metrics and considering the dynamism observed in the users behavior,does answer the communication needs of users in both dense (i.e., urban) andchallenged (i.e., development region) scenarios. Performance improvements goup to 54% regarding delivery capability while latency and cost can be reducedby 45% and 99% respectively, when compared to forwarding solely based ondata similarity and completely agnostic to user behavior.These findings point to a new research challenge regarding the impact ofDTN application: the investigation of self-awareness mechanisms able to adapttheir forwarding schemes based on the context of the user, namely the densityof the network where he/she is currently.