13.07.2015 Views

Social-based autonomic routing in opportunistic networks - Cnr

Social-based autonomic routing in opportunistic networks - Cnr

Social-based autonomic routing in opportunistic networks - Cnr

SHOW MORE
SHOW LESS
  • No tags were found...

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

4 Chiara Boldr<strong>in</strong>i, Marco Conti, Andrea Passarellaploits occasional communication opportunities among the <strong>in</strong>ternets, which mighteither be scheduled over time (e.g., due to the activation of a satellite l<strong>in</strong>k), or completelyrandom. In general, <strong>in</strong> conventional DTNs the po<strong>in</strong>ts of possible disconnectionsare known.Opportunistic <strong>networks</strong> can be seen as a generalisation of DTNs. Specifically, <strong>in</strong><strong>opportunistic</strong> <strong>networks</strong> no a-priori knowledge is assumed about the possible po<strong>in</strong>tsof disconnections, nor the existence of separate <strong>in</strong>ternet-like sub-<strong>networks</strong> is assumed.Opportunistic <strong>networks</strong> are formed by <strong>in</strong>dividual nodes, that are possiblydisconnected for long time <strong>in</strong>tervals, and that <strong>opportunistic</strong>ally exploit any contactwith other nodes to forward messages. The <strong>rout<strong>in</strong>g</strong> approach between conventionalDTNs and <strong>opportunistic</strong> <strong>networks</strong> is therefore quite different. S<strong>in</strong>ce <strong>in</strong> DTNs thepo<strong>in</strong>ts of disconnections (and, sometime, the duration of disconnections) are known,<strong>rout<strong>in</strong>g</strong> can be performed along the same l<strong>in</strong>es used for conventional Internet protocols,by simply consider<strong>in</strong>g the duration of the disconnections as an additional costof the l<strong>in</strong>ks ([22]). S<strong>in</strong>ce <strong>opportunistic</strong> <strong>networks</strong> do not assume the same knowledgeabout the network evolution, routes are computed dynamically while the messagesare be<strong>in</strong>g forwarded towards the dest<strong>in</strong>ation. Each <strong>in</strong>termediate node evaluates thesuitability of encountered nodes to be a good next hop towards the dest<strong>in</strong>ation.Fig. 1 The <strong>opportunistic</strong> network<strong>in</strong>g concept.For example, as shown <strong>in</strong> Figure 1, the user at the desktop <strong>opportunistic</strong>ally transfers,via a Wi-Fi ad hoc l<strong>in</strong>k, a message for a friend to a user pass<strong>in</strong>g nearby, “hop<strong>in</strong>g”that this user will carry the <strong>in</strong>formation closer to the dest<strong>in</strong>ation. This userpasses close to a tra<strong>in</strong> station, and forwards the message to a traveller go<strong>in</strong>g to thesame city where the dest<strong>in</strong>ation user works. At the tra<strong>in</strong> station of the dest<strong>in</strong>ationcity a car driver is go<strong>in</strong>g <strong>in</strong> the same neighbourhood of the dest<strong>in</strong>ation’s work<strong>in</strong>gplace. The driver meets the dest<strong>in</strong>ation user on his way, and the message is f<strong>in</strong>allydelivered.


<strong>Social</strong>-<strong>based</strong> <strong>autonomic</strong> <strong>rout<strong>in</strong>g</strong> <strong>in</strong> <strong>opportunistic</strong> <strong>networks</strong> 7Despite these nice properties, <strong>in</strong> [3] we have shown that the orig<strong>in</strong>al CMM proposalis not able to capture the attraction exerted on users by physical locations.Specifically, we have found that CMM shows a gregarious behaviour, such that allusers <strong>in</strong> a community tend to follow the first user that moves outside the physicallocation where the community is located. The gregarious behaviour does not representsignificant scenarios (e.g., work<strong>in</strong>g places), where users roam around preferredphysical places, besides be<strong>in</strong>g <strong>in</strong>fluenced by social relationships between each other.To address this issue, we propose the Home-cell Community-<strong>based</strong> Mobility Model(HCMM), which jo<strong>in</strong>s the concepts of CMM (for modell<strong>in</strong>g social relationships betweenusers) with the concept of def<strong>in</strong><strong>in</strong>g preferential locations <strong>in</strong> which users tendto spend most of their time. Therefore, HCMM is a first step towards jo<strong>in</strong><strong>in</strong>g togetherthe two promis<strong>in</strong>g mobility modell<strong>in</strong>g approaches discussed above. HCMMstill matches characteristic features of real traces (see [3]). Furthermore, we highlightthat, unlike CMM, it provides very simple knobs to control the time spent byusers <strong>in</strong> their preferred physical locations (Section 3.2).3.1 CMM and HCMM: functional descriptionThe Home-cell Community-<strong>based</strong> Mobility Model (HCMM) (fully described <strong>in</strong> [3])is an evolution of the Community-<strong>based</strong> Mobility Model. Community-<strong>based</strong> (orgroup) mobility models are attract<strong>in</strong>g <strong>in</strong>terest of researches <strong>in</strong> the <strong>opportunistic</strong> network<strong>in</strong>garea, because they are suitable to realistically model the <strong>in</strong>fluence of socialrelationships between people on the user mobility patterns.As <strong>in</strong> CMM, <strong>in</strong> HCMM every node belongs to a social community (group).Nodes that are <strong>in</strong> the same social community are called friends, while nodes <strong>in</strong> differentcommunities are called non-friends. Relationships between nodes are modelledthrough social l<strong>in</strong>ks (each l<strong>in</strong>k has an associated weight). At the system start-upall friends have a l<strong>in</strong>k to each other. Also two nodes that are not friends can havea l<strong>in</strong>k, accord<strong>in</strong>g to the rewir<strong>in</strong>g probability (p r ) parameter. Specifically, for eachnode, each l<strong>in</strong>k towards a friend is rewired to a non-friend with p r probability.<strong>Social</strong> l<strong>in</strong>ks are then used to drive node movements. Nodes move <strong>in</strong> a grid, andeach community is <strong>in</strong>itially randomly placed <strong>in</strong> a square of the grid. Nodes’ movementis made up of two component: first, a node has to select the cell towards whichto move. Node selects the target cell accord<strong>in</strong>g to the social attraction exerted byeach cell on the node. Attraction is measured as the sum of the l<strong>in</strong>ks’ weights betweenthe node and the nodes currently mov<strong>in</strong>g <strong>in</strong> or towards the cell. The targetcell is f<strong>in</strong>ally selected <strong>based</strong> on the probabilities def<strong>in</strong>ed by cells’ attraction (i.e., ifa j is the attraction of cell j, then the probability of select<strong>in</strong>g that cell is a j /∑ j a j ).After select<strong>in</strong>g the target cell, node selects the “goal” with<strong>in</strong> a cell (the precise po<strong>in</strong>ttowards which node will be head<strong>in</strong>g) accord<strong>in</strong>g to a uniform distribution. F<strong>in</strong>ally,speed is also selected accord<strong>in</strong>gly to a uniform distribution with<strong>in</strong> a user-specifiedrange. HCMM (and CMM) also allows for collective group movements. Specifically,once every reconfiguration period nodes of each group select a (different)


<strong>Social</strong>-<strong>based</strong> <strong>autonomic</strong> <strong>rout<strong>in</strong>g</strong> <strong>in</strong> <strong>opportunistic</strong> <strong>networks</strong> 9In [3] we have presented an analytical model to highlight a gregarious behaviourof CMM. To this end, we computed the rema<strong>in</strong><strong>in</strong>g probability (P rem ), def<strong>in</strong>ed as theprobability of no other member of node k’s community to move towards the dest<strong>in</strong>ationcell. When P rem approaches 0, at least one node <strong>in</strong> the start<strong>in</strong>g cell follows nodek. As will be clear from the follow<strong>in</strong>g analysis, this may generate an avalanche effectsuch that all nodes <strong>in</strong> node k’s community follow node k <strong>in</strong> the dest<strong>in</strong>ation cell,thus reveal<strong>in</strong>g the gregarious behaviour. We consider the case of a s<strong>in</strong>gle node (k)hav<strong>in</strong>g l<strong>in</strong>ks outside its community, because it represents the weaker condition forthe gregarious behaviour to take place. Therefore, the P rem formula computed <strong>in</strong> [3]is actually an upper bound of the rema<strong>in</strong><strong>in</strong>g probability achieved <strong>in</strong> the general case.Specifically, the f<strong>in</strong>al expression of P rem is:[[(P rem = (1 − P out ) l] n−1= 1 − w ) ]k/( f n+1) l n−1, (1)w k /( f n+1)+wwhere l is the average number of times each node <strong>in</strong> the start<strong>in</strong>g cell selects a newdest<strong>in</strong>ation while node k is associated with the dest<strong>in</strong>ation cell, 1−P out is the probabilityof each node to select the start<strong>in</strong>g cell for the next step, and n−1 is the numberof nodes <strong>in</strong> the start<strong>in</strong>g cell after node k departure. Furthermore, w k is the averageweight between node k and the other nodes of its community, f n + 1 the numberof nodes <strong>in</strong> the cell towards which node k is travell<strong>in</strong>g, and w is the average weightbetween nodes of node k community.P RemP Rem0.75 10.50.250l2 3 4 5 6 7 8 9 1010 20304050 nFig. 3 P rem as a function of n and l.This model (validated <strong>in</strong> [3] aga<strong>in</strong>st simulation results) allows us to show that thegregarious behaviour occurs basically for all sensible ranges of the model parameters.Just to show an example, Figure 3 illustrates the P rem dependence on n (thenumber of the nodes of k’s community), and l (the ratio between the movementsduration outside and with<strong>in</strong> a community).For small values of l, the grid has few cells and the duration of k’s movementoutside the start<strong>in</strong>g cell is not so different from the duration of nodes’ random move-


<strong>Social</strong>-<strong>based</strong> <strong>autonomic</strong> <strong>rout<strong>in</strong>g</strong> <strong>in</strong> <strong>opportunistic</strong> <strong>networks</strong> 11and CMM lies <strong>in</strong> the expressions of p <strong>in</strong> and p out , that we will derive at the end ofthis section. Otherwise, the analysis of the average time spent <strong>in</strong> the IN and OUTstates is common to CMM and HCMM.First of all, it is straightforward deriv<strong>in</strong>g the stationary distributions, π <strong>in</strong> =p <strong>in</strong> /(p <strong>in</strong> + p out ), and π out = p out /(p <strong>in</strong> + p out ). The average time spent <strong>in</strong> the IN andOUT states can be computed via the conditioned probabilities, as follows:{E [T<strong>in</strong> ] = π out p <strong>in</strong> · E [T <strong>in</strong> |E IN ](2)E [T out ] = π <strong>in</strong> p out · E [T out |E OUT ]where E IN and E OUT denote the events “the node enters the IN state” and “the nodeenters the OUT state”, respectively, while π out p <strong>in</strong> and π <strong>in</strong> p out are the probabilities ofthese events. By recall<strong>in</strong>g that i) the number of steps spent <strong>in</strong> each state is distributedaccord<strong>in</strong>g to a geometric law, ii) the duration of each step both <strong>in</strong> the IN and OUTstate can be approximated with T (<strong>in</strong>) , and iii) the duration of the transitions betweenthe states can be approximated with T (out) , we can compute closed form expressionsfor E [T <strong>in</strong> ] and E [T out ] as follows:{E [T<strong>in</strong> ] = p <strong>in</strong>(1−p out )p <strong>in</strong> +p outE [T out ] = p out(1−p <strong>in</strong> )p <strong>in</strong> +p out· T (<strong>in</strong>)· T (<strong>in</strong>) + p <strong>in</strong> p outp <strong>in</strong> +p out· 2T (out) (3)To specialise Equation 3 to HCMM and CMM we have to compute the transitionprobabilities of the correspond<strong>in</strong>g Markov cha<strong>in</strong>s, hereafter referred to as p (H)out andp (H)<strong>in</strong>, and p(C) out and p (C)<strong>in</strong>respectively. By def<strong>in</strong>ition, p (H)<strong>in</strong>is equal to 1 − p e . For theother parameters, we can use the follow<strong>in</strong>g l<strong>in</strong>e of reason<strong>in</strong>g, common to HCMMand CMM. To compute p out , we should focus on a node <strong>in</strong>side the start<strong>in</strong>g cell,and compute the attractions of the start<strong>in</strong>g and dest<strong>in</strong>ation cells. To compute p <strong>in</strong> , weshould compute the attractions of the start<strong>in</strong>g and dest<strong>in</strong>ation cells on a node outsidethe start<strong>in</strong>g cell. Then, p <strong>in</strong> and p out can be computed as follows:⎧SA⎪⎨ p <strong>in</strong> =(out)startSA (out)dest +SA(out) startSA⎪⎩ p out =(<strong>in</strong>) . (4)destSA (<strong>in</strong>)dest +SA(<strong>in</strong>) startClearly, the difference between HCMM and CMM turns out <strong>in</strong> different expressionsfor SA start and SA dest .The derivation is simpler <strong>in</strong> the case of HCMM. First of all, it is easy to realisethat the attractions of the start<strong>in</strong>g and dest<strong>in</strong>ation cells do not depend on the factthat the node is <strong>in</strong>side or outside the start<strong>in</strong>g cell. The attraction to the start<strong>in</strong>g(dest<strong>in</strong>ation) cell depends only on the relationships with nodes hav<strong>in</strong>g the start<strong>in</strong>g(dest<strong>in</strong>ation) cell as home, and on the number of such nodes. Thus, the attractions<strong>in</strong> HCMM are as follows:


12 Chiara Boldr<strong>in</strong>i, Marco Conti, Andrea Passarella⎧⎨ SA (H)start = ∑n−1 j=1 w i jn≃ w. (5)⎩SA (H)dest= ∑pr(n−1) j=1 w i jf n≃ p r(n−1)wf nClosed form expressions for the average time spent <strong>in</strong> the IN and OUT states <strong>in</strong>HCMM can be derived by replac<strong>in</strong>g Equations 5 and 4 <strong>in</strong> Equation 3.In the case of CMM comput<strong>in</strong>g attractions is more <strong>in</strong>volved. The attraction toa cell dynamically depends on the number of nodes actually be<strong>in</strong>g <strong>in</strong> that cell. Forthe sake of simplicity, we carry on the analysis under the hypothesis that q nodes ofthe start<strong>in</strong>g cell are roam<strong>in</strong>g <strong>in</strong> the dest<strong>in</strong>ation cell, and q′ nodes of the dest<strong>in</strong>ationcell are roam<strong>in</strong>g <strong>in</strong> the start<strong>in</strong>g cell. The attraction of the dest<strong>in</strong>ation cell on a nodecurrently roam<strong>in</strong>g <strong>in</strong> the start<strong>in</strong>g cell (and belong<strong>in</strong>g to the start<strong>in</strong>g cell’s community)are computed <strong>based</strong> on the follow<strong>in</strong>g l<strong>in</strong>e of reason<strong>in</strong>g. The node is attractedto the dest<strong>in</strong>ation cell because nodes of its community are roam<strong>in</strong>g there. S<strong>in</strong>cel<strong>in</strong>ks have been rewired, the node has l<strong>in</strong>ks just towards a fraction of these nodes,i.e., towards (1 − p r )q nodes, result<strong>in</strong>g <strong>in</strong> a contribution to the attraction equal to∑ q(1−p r)j=1w i j ≃ q(1 − p r )w. The node is also attracted by nodes of the dest<strong>in</strong>ation’scommunity, to which it has been rewired. The probability of the node hav<strong>in</strong>g beenrewired to a random node of the dest<strong>in</strong>ation community is (n−1)p rf n, and the numberof nodes exert<strong>in</strong>g such attraction is (n−1)p rf n( f n − q′). Based on the above l<strong>in</strong>e of reason<strong>in</strong>g(applicable also to the attraction of the start<strong>in</strong>g cell) it is possible to derivethe required attractions formulas for CMM, as follows:⎧⎪⎨⎪⎩SA (C,<strong>in</strong>)startSA (C,<strong>in</strong>)destSA (C,out)startSA (C,out)dest= (n−1−q)(1−p r)+ (n−1)prf n q′n−q+q′= q(1−p r)+ (n−1)prf n ( f n−q′)f n−q′+q= (n−q)(1−p r)+ (n−1)prf n q′n−q+q′· w· w· w= (q−1)(1−p r)+ (n−1)prf n ( f n−q′)f n−q′+qIn the case of CMM the closed form expression of E [T <strong>in</strong> ] and E [T out ] is not assimple as <strong>in</strong> HCMM. The key po<strong>in</strong>t is the fact that <strong>in</strong> CMM these figures dependon the dynamic evolution of the users’ movements. Specifically, they depend on qand q′, which are not model parameter, but change <strong>based</strong> on the nodes movements.Therefore, <strong>in</strong> CMM it is very hard to set model parameters to achieve a desirednodes’ behaviour as far as nodes’ physical positions. On the other hand, <strong>in</strong> HCMME [T <strong>in</strong> ] and E [T out ] do not depend on the dynamic evolution of the system, but dependonly on f ,n,p r , and p e . This means that HCMM, while reta<strong>in</strong><strong>in</strong>g the socialtheoretical approach of CMM, also provides simple knobs to control the time spentby nodes <strong>in</strong> the preferred physical locations. These remarks are confirmed by Figure5, which plots E [T <strong>in</strong> ] and E [T out ] for CMM and HCMM as functions of q (timeis normalised with respect to T (<strong>in</strong>) ).· w(6)


<strong>Social</strong>-<strong>based</strong> <strong>autonomic</strong> <strong>rout<strong>in</strong>g</strong> <strong>in</strong> <strong>opportunistic</strong> <strong>networks</strong> 13Normalised Time987654321T <strong>in</strong> CMMT out CMMT <strong>in</strong> HCMMT out HCMM01 2 3 4 5 6 7 8 9qFig. 5 Average time <strong>in</strong> the IN and OUT states as functions of q.4 Rout<strong>in</strong>g <strong>in</strong> <strong>opportunistic</strong> <strong>networks</strong>In all the case studies described <strong>in</strong> Section 2, <strong>rout<strong>in</strong>g</strong> is one of the most compell<strong>in</strong>gchallenge. The design of efficient <strong>rout<strong>in</strong>g</strong> strategies for <strong>opportunistic</strong> <strong>networks</strong> isgenerally a complicated task due to the absence of knowledge about the topologicalevolution of the network. Rout<strong>in</strong>g performance improves when more knowledgeabout the expected topology of the network can be exploited ([22]). Unfortunately,this k<strong>in</strong>d of knowledge is not easily available, and a trade-off must be met betweenperformance and knowledge requirement. A key piece of knowledge to design efficient<strong>rout<strong>in</strong>g</strong> protocols is <strong>in</strong>formation about the context <strong>in</strong> which the users communicate.Context <strong>in</strong>formation, such as the users’ work<strong>in</strong>g address and <strong>in</strong>stitution, theprobability of meet<strong>in</strong>g with other users or visit<strong>in</strong>g particular places, can be exploitedto identify suitable forwarders <strong>based</strong> on context <strong>in</strong>formation about the dest<strong>in</strong>ation.In the follow<strong>in</strong>g of this section we classify the ma<strong>in</strong> <strong>rout<strong>in</strong>g</strong> approaches proposed<strong>in</strong> the literature <strong>based</strong> on the amount of knowledge about the context of users theyexploit. We specifically identify three classes, correspond<strong>in</strong>g to context-oblivious,partially context-aware, and fully context-aware protocols.4.1 Context-oblivious <strong>rout<strong>in</strong>g</strong>Rout<strong>in</strong>g techniques <strong>in</strong> this class basically exploit some form of flood<strong>in</strong>g. The heuristicbeh<strong>in</strong>d this policy is that, when there is no knowledge of a possible path towardsthe dest<strong>in</strong>ation nor of an appropriate next-hop node, a message should be dissem<strong>in</strong>atedas widely as possible. Protocols <strong>in</strong> this class might be the only solution whenno context <strong>in</strong>formation is available. Clearly, they generate a high overhead (as wealso highlight <strong>in</strong> the performance evaluation section), may suffer high contentionand potentially lead to network congestion ([23]). To limit this overhead, the commontechnique is to control flood<strong>in</strong>g by either limit<strong>in</strong>g the number of copies allowed


14 Chiara Boldr<strong>in</strong>i, Marco Conti, Andrea Passarellato exist <strong>in</strong> the network, or by limit<strong>in</strong>g the maximum number of hops a message cantravel. In the latter case, when no relay<strong>in</strong>g is further allowed, a node can only senddirectly to the dest<strong>in</strong>ation when (<strong>in</strong> case) it is met.The most representative protocol of this type is Epidemic Rout<strong>in</strong>g (Epidemicfor short) ([38]). Whenever two nodes come <strong>in</strong>to communication range they exchangesummary vectors that conta<strong>in</strong> a compact unambiguous representation of themessages currently stored <strong>in</strong> the local buffers. Then, each node requests from theother the messages it is currently miss<strong>in</strong>g. The dissem<strong>in</strong>ation process is somehowbounded because each message is assigned a hop count limit giv<strong>in</strong>g the maximumnumber of hops it is allowed to traverse till the dest<strong>in</strong>ation. When the hop countlimit is set to one, the message can only be sent directly to the dest<strong>in</strong>ation node.Dissem<strong>in</strong>ation-<strong>based</strong> algorithms also <strong>in</strong>clude network-cod<strong>in</strong>g-<strong>based</strong> <strong>rout<strong>in</strong>g</strong> ([39]),which takes an orig<strong>in</strong>al approach to limit message flood<strong>in</strong>g. Just to give a classicalexample, let A, B, and C, be the only three nodes of a str<strong>in</strong>g network, such as anymessage travell<strong>in</strong>g between A and C has to be relayed by B. Let node A generatemessage a addressed to node C, and node C generate the message c addressed tonode A. In a conventional forward<strong>in</strong>g scheme node B has to relay message a to Cand message c to A. In network cod<strong>in</strong>g, node B broadcasts a s<strong>in</strong>gle packet conta<strong>in</strong><strong>in</strong>ga ⊕ c. Once received a ⊕ c, both nodes A and C can decode the messages. Ingeneral, network cod<strong>in</strong>g-<strong>based</strong> <strong>rout<strong>in</strong>g</strong> outperforms flood<strong>in</strong>g, as it is able to deliverthe same amount of <strong>in</strong>formation with fewer messages <strong>in</strong>jected <strong>in</strong>to the network. Amore extended survey about network cod<strong>in</strong>g techniques can be found <strong>in</strong> [33].An alternative, drastic way of reduc<strong>in</strong>g the overhead of Epidemic without rely<strong>in</strong>gon network cod<strong>in</strong>g is implemented by Spray&Wait ([35]). Message delivery is subdivided<strong>in</strong> two phases: the spray phase and the wait phase. Dur<strong>in</strong>g the spray phasemultiple copies of the same message are spread over the network both by the sourcenode and those nodes that have first received the message from the source node itself.This phase ends when a given number of copies, say L, have been dissem<strong>in</strong>ated<strong>in</strong> the network. Then, <strong>in</strong> the wait phase each node hold<strong>in</strong>g a copy of the message(i.e., each relay node) stores its copy and eventually delivers it to the dest<strong>in</strong>ationwhen (<strong>in</strong> case) it comes with<strong>in</strong> reach. The analytical model derived <strong>in</strong> [35] showsthat L can be chosen <strong>based</strong> on a target average delay. The spray phase may be performed<strong>in</strong> many ways. Under the assumption that nodes movements are i.i.d., theB<strong>in</strong>ary Spray and Wait policy is the best one <strong>in</strong> terms of delay. Any node (<strong>in</strong>clud<strong>in</strong>gthe sender) hold<strong>in</strong>g n copies (n > 1) of the message hands over ⌊ n2⌋copies to thefirst encountered node, and keeps the rema<strong>in</strong><strong>in</strong>g copies for itself. When a node isleft with only one copy of the message, it switches to direct transmission and onlytransmits the message to the f<strong>in</strong>al dest<strong>in</strong>ation node when (if) it is met.4.2 Partially context-aware <strong>rout<strong>in</strong>g</strong>Partially context-aware protocols exploit some particular piece of context <strong>in</strong>formationto optimise the forward<strong>in</strong>g task. The ma<strong>in</strong> difference with fully context-aware


<strong>Social</strong>-<strong>based</strong> <strong>autonomic</strong> <strong>rout<strong>in</strong>g</strong> <strong>in</strong> <strong>opportunistic</strong> <strong>networks</strong> 15protocols is the fact that the latter usually provide a full-fledged set of algorithms togather and manage any type of context <strong>in</strong>formation, while the former are customisedfor a specific type of context <strong>in</strong>formation.Probabilistic Rout<strong>in</strong>g Protocol us<strong>in</strong>g History of Encounters and Transitivity(PROPHET [29]) is one of the most popular examples of protocols fall<strong>in</strong>g <strong>in</strong> thisclass. PROPHET is an evolution of Epidemic that <strong>in</strong>troduces the concept of deliverypredictability. The delivery predictability is the probability for a node to encounter acerta<strong>in</strong> dest<strong>in</strong>ation. The delivery predictability for a dest<strong>in</strong>ation <strong>in</strong>creases when thenode meets the dest<strong>in</strong>ation, and decreases (accord<strong>in</strong>g to an age<strong>in</strong>g function) betweenmeet<strong>in</strong>gs. A transitivity law is also <strong>in</strong>cluded <strong>in</strong> the algorithm, such that if node Afrequently meets node B, and node B frequently meets node C, then nodes A andC have high delivery predictability to each other. The PROPHET forward<strong>in</strong>g algorithmis similar to Epidemic except that, dur<strong>in</strong>g a contact, nodes also exchange theirdelivery predictability to the dest<strong>in</strong>ations of the messages they store <strong>in</strong> their buffers,and messages are requested only if the delivery predictability of the request<strong>in</strong>g nodeis higher than that of the node currently stor<strong>in</strong>g the message.The context <strong>in</strong>formation used by PROPHET is the frequency of meet<strong>in</strong>gs betweennodes. The same type of context <strong>in</strong>formation is also used by MV ([8]) andMaxProp ([7]), which, <strong>in</strong> addition, also exploit <strong>in</strong>formation about the frequencyof visits to specific physical places. Other protocols use the time lag from the lastmeet<strong>in</strong>g with a dest<strong>in</strong>ation to estimate the probability of deliver<strong>in</strong>g the messages.The bottoml<strong>in</strong>e idea (thoroughly <strong>in</strong>vestigated <strong>in</strong> [17]) is that the decreas<strong>in</strong>g gradientof the time lag identifies a suitable path towards the dest<strong>in</strong>ation. Examples of protocolsexploit<strong>in</strong>g this piece of context <strong>in</strong>formation are Last Encounter Rout<strong>in</strong>g ([17])and Spray&Focus ([35]).In MobySpace Rout<strong>in</strong>g ([26]) the mobility pattern of nodes is the context <strong>in</strong>formationused for <strong>rout<strong>in</strong>g</strong>. The protocol builds up a high dimensional Euclidean space,named MobySpace, where each axis represents a possible contact between a coupleof nodes and the distance along an axis measures the probability of that contactto occur. Two nodes that have similar sets of contacts, and that experience thosecontacts with similar frequencies, are close <strong>in</strong> the MobySpace. The best forward<strong>in</strong>gnode for a message is the node that is as close as possible to the dest<strong>in</strong>ation node<strong>in</strong> this space. Obviously, <strong>in</strong> the virtual contact space just described, the knowledgeof all the axes of the space also requires the knowledge of all the nodes that arecirculat<strong>in</strong>g <strong>in</strong> the space. This full knowledge, however, might not be required forsuccessful <strong>rout<strong>in</strong>g</strong>.The f<strong>in</strong>al example we mention is Bubble Rap ([20]), <strong>in</strong> which the context <strong>in</strong>formationis the social community users belong to. In Bubble Rap communities areautomatically detected via the patterns of contacts between nodes. It is assumed thatcommunities are labelled. Messages orig<strong>in</strong>at<strong>in</strong>g <strong>in</strong> a community different from thedest<strong>in</strong>ation’s one are forwarded as follows. Assume node A is carry<strong>in</strong>g a messageaddressed to D, and meets node B. The message is handed over to B if the communityof B is the same as the community of D, or if B has a higher rank<strong>in</strong>g withrespect to node A. The rank<strong>in</strong>g is measured <strong>based</strong> on the set of peers a node is usually<strong>in</strong> touch with, and is thus a measure of the “sociability” of nodes. Basically,


16 Chiara Boldr<strong>in</strong>i, Marco Conti, Andrea PassarellaBubble Rap looks for nodes belong<strong>in</strong>g to the same community of the dest<strong>in</strong>ation. Ifsuch nodes are not found, it forwards the message to <strong>in</strong>creas<strong>in</strong>gly sociable nodes,which have more chances to get <strong>in</strong> touch with the community of the dest<strong>in</strong>ation.Exploit<strong>in</strong>g context <strong>in</strong>formation related to the social behaviour of people is one ofthe most promis<strong>in</strong>g research directions <strong>in</strong> the area.4.3 Fully context-aware <strong>rout<strong>in</strong>g</strong>Fully context-aware protocols not only exploit context <strong>in</strong>formation to optimise <strong>rout<strong>in</strong>g</strong>,but also provide general mechanisms to handle and use context <strong>in</strong>formation.The advantage of this approach is to be much more general than the approachesmentioned <strong>in</strong> Section 4.2. Indeed, these <strong>rout<strong>in</strong>g</strong> protocols can be used with any setof context <strong>in</strong>formation, thus allow<strong>in</strong>g the system to be customised to the particularenvironment it has to operate <strong>in</strong>. To the best of our knowledge, two protocols onlyfall <strong>in</strong> this category, i.e., Context-Aware Rout<strong>in</strong>g (CAR, [30]) and History BasedOpportunistic Rout<strong>in</strong>g (HiBOp, [1]).CAR assumes an underly<strong>in</strong>g MANET <strong>rout<strong>in</strong>g</strong> protocol that connects togethernodes <strong>in</strong> the same MANET cloud. To reach nodes outside the cloud, a sender looksfor the node <strong>in</strong> its current cloud with the highest probability of deliver<strong>in</strong>g the messagesuccessfully to the dest<strong>in</strong>ation. This node temporarily stores the message, wait<strong>in</strong>geither to get <strong>in</strong> touch with the dest<strong>in</strong>ation itself, of to enter a cloud with othernodes with higher probability of meet<strong>in</strong>g the dest<strong>in</strong>ation. Therefore, nodes <strong>in</strong> CARcompute delivery probabilities proactively, and dissem<strong>in</strong>ate them <strong>in</strong> their ad hoccloud. Note that context <strong>in</strong>formation is exploited to evaluate probabilities just forthose dest<strong>in</strong>ations each node is aware of (i.e., that happen to have been co-located<strong>in</strong> the same cloud at some time). The ma<strong>in</strong> focus of CAR is on def<strong>in</strong><strong>in</strong>g algorithmsto comb<strong>in</strong>e context <strong>in</strong>formation (which is assumed available <strong>in</strong> some way) to computedelivery probabilities. Specifically, a multi-attribute utility-<strong>based</strong> framework isdef<strong>in</strong>ed to this end. The framework is general enough to accommodate for differenttypes of context <strong>in</strong>formation. As an example, <strong>in</strong> [30] authors use residual batterylife, the rate of connectivity change and the probability of meet<strong>in</strong>g between nodesas context <strong>in</strong>formation.With respect to CAR, HiBOp is more general, as it does not necessarily requirean underly<strong>in</strong>g MANET <strong>rout<strong>in</strong>g</strong> protocol, and is able to exploit context <strong>in</strong>formationalso for those nodes that have never been with<strong>in</strong> the same cloud. Furthermore, thedef<strong>in</strong>ition and management of context <strong>in</strong>formation is not addressed <strong>in</strong> CAR, whileit is a core part of HiBOp. F<strong>in</strong>ally, and most importantly, CAR does not capture, <strong>in</strong>the context def<strong>in</strong>ition, any <strong>in</strong>formation about the users social behavior, which results<strong>in</strong> [1] demonstrate be<strong>in</strong>g a particularly valuable piece of <strong>in</strong>formation to design anefficient <strong>rout<strong>in</strong>g</strong> scheme.S<strong>in</strong>ce the performance analysis presented <strong>in</strong> this chapter focuses on the HiBOpprotocol, we describe its mechanisms <strong>in</strong> more details <strong>in</strong> the follow<strong>in</strong>g section.


<strong>Social</strong>-<strong>based</strong> <strong>autonomic</strong> <strong>rout<strong>in</strong>g</strong> <strong>in</strong> <strong>opportunistic</strong> <strong>networks</strong> 174.4 The History-<strong>based</strong> Opportunistic Rout<strong>in</strong>g protocolHiBOp is a fully context-aware <strong>rout<strong>in</strong>g</strong> protocol completely described <strong>in</strong> [1]. HiBOp<strong>in</strong>cludes mechanisms to handle any type of context <strong>in</strong>formation. As a particular <strong>in</strong>stance,<strong>in</strong> [1] the context is assumed to be a collection of <strong>in</strong>formation that describesthe community <strong>in</strong> which the user lives, and the history of social relationships amongusers. At each node, basic data used to build the context can be personal <strong>in</strong>formationabout the user (e.g. name), about her residence (e.g. address), about her work(e.g. <strong>in</strong>stitution), etc. In HiBOp nodes share their own data dur<strong>in</strong>g contacts, andthus learn the context they are immersed <strong>in</strong>. Messages are forwarded through nodesthat share more and more context data with the message dest<strong>in</strong>ation. S<strong>in</strong>ce users ofHiBOp have possibly to share personal <strong>in</strong>formation, privacy issues should be considered.Privacy management <strong>in</strong> <strong>opportunistic</strong> <strong>networks</strong> is – <strong>in</strong> general – a topic stilllargely not addressed, and it is not the target of this chapter to provide completeprivacy solutions for HiBOp. It should be noted that the set of <strong>in</strong>formation that isconsidered <strong>in</strong> [1] (and that we also consider hereafter) is equivalent to personal <strong>in</strong>formationpeople advertise on their public web pages (e.g., the work<strong>in</strong>g <strong>in</strong>stitutionand address) which are, therefore, not perceived as sensitive <strong>in</strong>formation from a privacystandpo<strong>in</strong>t. Design<strong>in</strong>g complete privacy solutions for HiBOp is one of the ma<strong>in</strong>subjects of future work.Table 1 Identity TablePersonal InformationResidenceName John Doe City PisaEmail j.doe@iit.cnr.it Street Via Garibaldi, 2. . . . . .More <strong>in</strong> detail, HiBOp assumes that each node locally stores an Identity Table(IT), that conta<strong>in</strong>s personal <strong>in</strong>formation on the user that owns the device (an exampleis reported <strong>in</strong> Table 1). Nodes exchange ITs when gett<strong>in</strong>g <strong>in</strong> touch. At each node, itsown IT, and the set of current neighbours’ ITs, represent the Current Context, whichprovides a snapshot of the context the node is currently <strong>in</strong>.The current context is useful <strong>in</strong> order to evaluate the <strong>in</strong>stantaneous fitness of anode to be a forwarder. But even if a node is not a good forwarder because of itscurrent location/neighbors, it could be a valid carrier because of its habits and pastexperiences. Under the assumption that humans are most of the time “predictable”,it is important to collect <strong>in</strong>formation about the context data seen by each node <strong>in</strong>the past, and the recurrence of these data <strong>in</strong> the node’s Current Context. To this end,each context attribute seen <strong>in</strong> the Current Context (i.e., each row <strong>in</strong> neighbours’ ITs)is recorded <strong>in</strong> a History Table (HT), together with a Cont<strong>in</strong>uity Probability <strong>in</strong>dex,that represents the probability of encounter<strong>in</strong>g that attribute <strong>in</strong> the future (actuallymore <strong>in</strong>dices are used, as described <strong>in</strong> [1]).


18 Chiara Boldr<strong>in</strong>i, Marco Conti, Andrea PassarellaThe ma<strong>in</strong> idea of HiBOp forward<strong>in</strong>g is look<strong>in</strong>g for nodes that show <strong>in</strong>creas<strong>in</strong>gmatch with known context attributes of the dest<strong>in</strong>ation. High match means high similaritybetween node’s and dest<strong>in</strong>ation’s contexts and, therefore, high probability forthe node to br<strong>in</strong>g the message <strong>in</strong> the dest<strong>in</strong>ation’s community (possibly, to the dest<strong>in</strong>ation).Therefore, a node wish<strong>in</strong>g to send a message through HiBOp specifies (anysubset of) the dest<strong>in</strong>ation’s Identity Table <strong>in</strong> the message header. Any node <strong>in</strong> thepath between the sender and the dest<strong>in</strong>ation asks encountered nodes for their matchwith the dest<strong>in</strong>ation attributes, and hands over the message if an encountered nodeshows a greater match than its own. The detailed algorithms to evaluate matchesare described <strong>in</strong> [1]. It is worth recall<strong>in</strong>g here that matches are evaluated as deliveryprobabilities, and dist<strong>in</strong>ct probabilities are computed <strong>based</strong> on the Current Context(P CC ) only, and on the History (P H ) only. The f<strong>in</strong>al probability is evaluated via standardsmoothed average, as P = α · P H +(1 − α) · P CC ,0 ≤ α ≤ 1. The α parameterallows HiBOp to tune the relative importance of the Current Context and History.In HiBOp just the source node is allowed to replicate the message, <strong>in</strong> order totightly control the trade-off between reliability and message spread. Specifically,the source node replicates the message until the jo<strong>in</strong>t loss probability of nodesused for replication is below a system-def<strong>in</strong>ed threshold (p maxl). Specifically, ifp (i) is the delivery probability of the i-th node used for replicat<strong>in</strong>g the message,and k is{the number of nodes used } for replication, the follow<strong>in</strong>g equation holds:k = m<strong>in</strong> j|∏ ji=0 (1 − p (i)) ≤ p maxl.5 Performance of <strong>opportunistic</strong> <strong>rout<strong>in</strong>g</strong> approaches under socialmobility patternsThe goal of this section is to compare the different <strong>opportunistic</strong> <strong>rout<strong>in</strong>g</strong> approaches<strong>in</strong> realistic human mobility scenarios. Specifically, we <strong>in</strong>vestigate the protocols’ behaviourwith respect to a number of parameters that describe user movement patterns.The performance evaluation is carried out by consider<strong>in</strong>g the two oppositeends of the spectrum presented <strong>in</strong> Section 4. Specifically, we compare a contextoblivious<strong>rout<strong>in</strong>g</strong> protocol (Epidemic) with a fully context-aware <strong>rout<strong>in</strong>g</strong> protocol(HiBOp).5.1 Performance evaluation strategyIn the follow<strong>in</strong>g of the chapter we highlight how the different <strong>rout<strong>in</strong>g</strong> approachesare able to <strong>autonomic</strong>ally react and adapt to the dynamically evolv<strong>in</strong>g conditionsof the operat<strong>in</strong>g scenario. To this end, we exploit several control knobs providedby HCMM to highlight the different <strong>autonomic</strong> properties of Epidemic and Hi-BOp. Specifically, we identify three ma<strong>in</strong> reference cases for our study. In the first


<strong>Social</strong>-<strong>based</strong> <strong>autonomic</strong> <strong>rout<strong>in</strong>g</strong> <strong>in</strong> <strong>opportunistic</strong> <strong>networks</strong> 19one (Section 5.2), we analyse the reactivity of <strong>rout<strong>in</strong>g</strong> protocols to sudden contactsamong groups. Specifically, we focus on closed groups (i.e., p r = 0), and then weforce groups to collectively move with vary<strong>in</strong>g frequency. Messages addressed tonodes outside the group can be delivered only dur<strong>in</strong>g contacts between differentgroup members dur<strong>in</strong>g collective movements 2 . This analysis allows us to understandif <strong>rout<strong>in</strong>g</strong> protocols are able to exploit even those few chances to f<strong>in</strong>d goodroutes. We analyse this aspect by vary<strong>in</strong>g the reconfiguration <strong>in</strong>terval parameter.In the second scenario, (Section 5.3), we analyse the effect of social relationshipsbetween users. We want to understand how <strong>rout<strong>in</strong>g</strong> protocols react to different levelsof users’ sociability, measured as the probability of users hav<strong>in</strong>g relationshipsoutside their reference group. We clearly achieve this by vary<strong>in</strong>g the rewir<strong>in</strong>g parameter(p r ). The higher p r , the more nodes are “social”, the lesser groups are closedcommunities.In the third scenario, we look at how protocols work <strong>in</strong> completely closed groups.In this case no rewir<strong>in</strong>g nor reconfigurations are allowed, and we place a differentgroup <strong>in</strong> each cell of the grid. Therefore, the only chance of deliver<strong>in</strong>g messages betweengroups is by exploit<strong>in</strong>g contacts between nodes at the borders of the cells. Westudy the <strong>rout<strong>in</strong>g</strong> protocols’ performance as a function of the nodes’ transmissionrange. Basically, this scenario allows us to understand how protocols can exploitcontacts that are not related to social relationships, but just happen because of physicalco-location (e.g., contacts between people work<strong>in</strong>g for different companies <strong>in</strong>the same floor of a build<strong>in</strong>g).We test <strong>rout<strong>in</strong>g</strong> performance <strong>in</strong> terms of QoS perceived by users, and resourceconsumption. The user QoS is evaluated <strong>in</strong> terms of message delay and packet loss.Message delay is evaluated <strong>based</strong> on the first replica reach<strong>in</strong>g the dest<strong>in</strong>ation, whilewe count a packet loss if all replicas get lost. To highlight some specific differentbehaviour between Epidemic and HiBOp, <strong>in</strong> some cases we also show the averagenumber of hops required by messages to be delivered, and we separate the delayfor messages addressed to friend and non-friend nodes. Resource consumption isevaluated <strong>in</strong> terms of buffer occupation and bandwidth overhead. Specifically, thebandwidth overhead is computed as the ratio between the number of bytes generated<strong>in</strong> the whole network dur<strong>in</strong>g a simulation run, and the number of bytes generated bythe senders. Note that we count <strong>in</strong> all overheads related to <strong>rout<strong>in</strong>g</strong> and forward<strong>in</strong>g,such as the exchanges of Identity Tables, requests for delivery probabilities, etc. Tohighlight specific differences, <strong>in</strong> a few cases we also show the number of copiesspread <strong>in</strong> the network, and we separately highlight the bandwidth overhead relatedto data and non-data messages.To highlight the effect of human mobility patterns only, we assume i) <strong>in</strong>f<strong>in</strong>itebuffers, ii) an ideal MAC level that completely avoids congestion impairments, iii)an ideal physical channel where nodes experience 0% packet loss with<strong>in</strong> a circulartransmission range and 100% packet loss outside; and iv) “<strong>in</strong>f<strong>in</strong>ite” bandwidth(<strong>in</strong> the sense that messages can be always exchanged when nodes get <strong>in</strong> touch). Asthoroughly discussed <strong>in</strong> [1], this setup tends to favour dissem<strong>in</strong>ation-<strong>based</strong> schemes2 The probability of contacts due to groups choos<strong>in</strong>g adjacent cells is typically low due to the highnumber of cells with respect to the number of groups.


20 Chiara Boldr<strong>in</strong>i, Marco Conti, Andrea Passarellasuch as Epidemic. More specifically, <strong>in</strong> this configuration HiBOp best results wouldbe to approach the delay and packet loss achieved by Epidemic, while significantlyreduc<strong>in</strong>g the resource consumption. F<strong>in</strong>ally, unless otherwise stated, our setup consistsof 30 nodes evenly divided <strong>in</strong> three groups. We assume a square simulationarea 1250mx1250m large, divided <strong>in</strong> a 5x5 grid. The default transmission range is125m. Unless otherwise stated 2 nodes <strong>in</strong> each group generate messages, with an<strong>in</strong>ter-generation time exponentially distributed (with average 300s). Each messageis addressed to a friend or to a non-friend node with 50% probability. Messages expireafter 18000s. Each simulation run for 90000s (of simulated time). For particularsetups we <strong>in</strong>creased the run lenghts so as to achieve a m<strong>in</strong>imum amount of characteristicevents <strong>in</strong> each run (e.g. reconfiguration runs with reconfiguration <strong>in</strong>tervalequal to 36000s last for 397000s). To make sure that messages still not deliveredat the end of a run will never be delivered (so as to achieve a correct measure ofthe packet loss <strong>in</strong>dex), dur<strong>in</strong>g the last 18000s senders do not generate any new message.Furthermore, statistics are collected elim<strong>in</strong>at<strong>in</strong>g the <strong>in</strong>itial and f<strong>in</strong>al transitoryregimes, i.e., us<strong>in</strong>g the steady-state phase of simulation runs only. Each setup wasreplicated 50 times: statistics presented hereafter are averaged over the 50 replicas,with confidence <strong>in</strong>terval at 95% confidence level.5.2 Impact of collective groups’ movements (reconfigurations)It is worth recall<strong>in</strong>g that <strong>in</strong> this scenario the rewir<strong>in</strong>g probability is 0, and thus, exceptfor reconfigurations, nodes do not have chances to meet. The reconfiguration<strong>in</strong>terval varies between 2250s, 9000s, and 36000s. Table 2 shows the QoS performanceas a function of the reconfiguration <strong>in</strong>terval. As expected, both packet lossand delay <strong>in</strong>crease with this parameter, because messages addressed outside thegroup of the sender are forced to wait for a reconfiguration. The performance <strong>in</strong>terms of delay can be better highlighted by separately focus<strong>in</strong>g on delay towardsfriend and non-friend nodes. Specifically, Figures 6, 7, and 8 show the delay distributiontowards friend nodes (left-hand-side plots) and non-friend nodes (right-handsideplots) for the three reconfiguration periods. First of all, delays towards friendsbasically do not depend on the reconfiguration <strong>in</strong>terval, s<strong>in</strong>ce friends are always colocated<strong>in</strong> the same group. While only a small amount of messages dest<strong>in</strong>ed to friendnodes experiences a delay greater than 10s, most (between 60% and 70% depend<strong>in</strong>gon the reconfiguration <strong>in</strong>terval) of the messages addressed to non-friend nodes experiencea delay greater than 10 3 . Furthermore, note that depend<strong>in</strong>g on the frequencyof reconfigurations, distributions’ tails are more or less “heavy”. The worst case isclearly for a reconfiguration <strong>in</strong>terval equal to 36000s, where about 50% of messagestowards non-friend dest<strong>in</strong>ations expire. Also note that, even though HiBOp provideshigher packet loss and delay, the difference with Epidemic is quite th<strong>in</strong>. Note that, asbuffers and bandwidth are not limited, Epidemic gives a reference upper bound onthe performance achievable by any <strong>rout<strong>in</strong>g</strong> protocol. These results clearly show that


<strong>Social</strong>-<strong>based</strong> <strong>autonomic</strong> <strong>rout<strong>in</strong>g</strong> <strong>in</strong> <strong>opportunistic</strong> <strong>networks</strong> 21HiBOp is able to identify very good paths even dur<strong>in</strong>g sporadic, sudden contactsdur<strong>in</strong>g reconfigurations among nodes belong<strong>in</strong>g to different groups.Table 2 Users QoS (focus on the reconfiguration parameter)reconf (s) HiBOp Epidemic2250 0 ± 0 0 ± 0ploss (%) 9000 8.16 ± 1.68 5.52 ± 1.4636000 25.64 ± 1.30 24.12 ± 1.312250 1202.52 ± 91.09 907.10 ± 67.08delay (s) 9000 3651.68 ± 295.05 3204.58 ± 278.7036000 5615.43 ± 225.93 5445.11 ± 161.53The good performance <strong>in</strong> terms of user QoS shown by HiBOp comes along with adrastic reduction <strong>in</strong> resource usage. Figure 9 shows the buffer occupation over timeshown as a percentage of duration of a simulation run (po<strong>in</strong>ts are average valuesover the replicas). HiBOp is much less greedy <strong>in</strong> spread<strong>in</strong>g messages, and thereforethe buffer occupation is drastically reduced. This is a general difference betweenEpidemic and HiBOp, which is confirmed <strong>in</strong> all scenarios we have tested. The extentof this reduction depends on the scenario, and can be as high as an order ofmagnitude.Figure 10 compares Epidemic and HiBOp with respect to the number of copiesgenerated (recall that the number of nodes <strong>in</strong> the network is 30, thus the maximumnumber of copies is 29). High resource consumption for Epidemic is due to thefact that each node copies all its messages to all nodes it encounters. Therefore, themore the contacts between nodes, the more the spread<strong>in</strong>g of messages. Figure 10shows that approximately 50% of messages (correspond<strong>in</strong>g to the messages witha non-friend dest<strong>in</strong>ation) are spread by Epidemic across the whole network, whenthe reconfiguration <strong>in</strong>terval is equal to 2250s and 9000s. The performance <strong>in</strong> termsof delay and packet loss shows that <strong>in</strong> this particular scenario flood<strong>in</strong>g yields nosignificant advantages. As contacts dur<strong>in</strong>g reconfigurations <strong>in</strong>volve entire groups,a fully replication <strong>in</strong>side each group is not more convenient than replicat<strong>in</strong>g themessage on a s<strong>in</strong>gle node of each group. HiBOp, due to its reliability rule, tends toreplicate the message <strong>in</strong>side the sender’s group, but does not flood the other groupsupon reconfigurations, thus result<strong>in</strong>g <strong>in</strong> lower number of copies.F<strong>in</strong>ally, Figure 11 shows the bandwidth overhead of the two protocols. It allowsus to highlight a ma<strong>in</strong> difference between HiBOp and Epidemic, related to how theyreact to movement patterns. Reduc<strong>in</strong>g the reconfiguration <strong>in</strong>terval (from 36000sdown to 2250s) means <strong>in</strong>creas<strong>in</strong>g the forward<strong>in</strong>g opportunities, because nodes get<strong>in</strong> touch with more peers more frequently. Epidemic does not use these additional“connectivity resources” wisely, as it is <strong>based</strong> on flood<strong>in</strong>g. Therefore, the bandwidthoverhead greatly <strong>in</strong>creases. HiBOp behaves <strong>in</strong> a different way. When groups do notmix (reconfiguration <strong>in</strong>terval equal to 36000s) paths for messages go<strong>in</strong>g outsidethe sender’s group are seldom available. HiBOp realises this, because context <strong>in</strong>-


22 Chiara Boldr<strong>in</strong>i, Marco Conti, Andrea PassarellaP(delay>x)10.90.80.70.60.50.40.30.20.1010 -210 -1Delay CCDF - Friend dest.1 10 10 2 10 3Time [s]10 410 52250 E 2250 H(a) friendsP(delay>x)10.90.80.70.60.50.40.30.20.1010 -210 -1Delay CCDF - Non-friend Dest.110 10 2Time [s]10 310 410 52250 E 2250 H(b) non-friendsFig. 6 Delay distributions with reconfigurations every 2250sformation about nodes outside the group is rarely available, and avoids consum<strong>in</strong>gresources uselessly. As nodes mix more and more (reconfiguration <strong>in</strong>tervals equalto 9000s and 2250s), also HiBOp (as Epidemic) generates more overhead, becausemore contacts become available, which may possibly lead to paths towards the dest<strong>in</strong>ation.However, the rate of <strong>in</strong>crease of the HiBOp’s overhead is significantly lowerthan the one of Epidemic, thus show<strong>in</strong>g a much more judicious use of the availablenetwork resources. These results <strong>in</strong>dicate that exploit<strong>in</strong>g context <strong>in</strong>formation makesHiBOp much more efficient than flood<strong>in</strong>g-<strong>based</strong> protocols, despite the additionalresources needed for context management purposes.


<strong>Social</strong>-<strong>based</strong> <strong>autonomic</strong> <strong>rout<strong>in</strong>g</strong> <strong>in</strong> <strong>opportunistic</strong> <strong>networks</strong> 23P(delay>x)10.90.80.70.60.50.40.30.20.10-0.110 -210 -1Delay CCDF - Friend dest.1 10 10 2 10 3Time [s]10 410 59000 E 9000 H(a) friendsP(delay>x)10.90.80.70.60.50.40.30.20.10-0.110 -210 -1Delay CCDF - Non-friend Dest.110 10 2Time [s]10 310 410 59000 E 9000 H(b) non-friendsFig. 7 Delay distributions with reconfigurations every 9000s5.3 Impact of User SociabilityTo understand the impact of user sociability on <strong>rout<strong>in</strong>g</strong> performance we vary therewir<strong>in</strong>g parameter (p r ). When a node goes to a cell different from its home itshows to nodes <strong>in</strong> the “foreign” cell context <strong>in</strong>formation related to its home cell,thus becom<strong>in</strong>g a good next hop for messages dest<strong>in</strong>ed to its friends. On the otherhand, it roams <strong>in</strong> the foreign cell for a number of rounds and collects context dataabout nodes <strong>in</strong> that cell. When it then comes back to the home cell, this knowledgecan effectively be used for send<strong>in</strong>g messages to that particular foreign cell. Indeed,that node is likely to go back to the same foreign cell after a while, because the sociall<strong>in</strong>ks towards nodes <strong>in</strong> that cell are still active. Clearly, the <strong>rout<strong>in</strong>g</strong> performance


24 Chiara Boldr<strong>in</strong>i, Marco Conti, Andrea PassarellaP(delay>x)10.90.80.70.60.50.40.30.20.10-0.110 -210 -1Delay CCDF - Friend dest.1 10 10 2 10 3Time [s]10 410 536000 E 36000 H(a) friendsP(delay>x)10.90.80.70.60.50.40.30.20.10-0.110 -2Delay CCDF - Non-friend Dest.10 -1 1 10 10 2 10 3 10 4Time [s]10 536000 E 36000 H(b) non-friendsFig. 8 Delay distributions with reconfigurations every 36000sis sensitive to the user sociability, because users hav<strong>in</strong>g social relationships withother groups are the only possible way of gett<strong>in</strong>g messages out of the orig<strong>in</strong>at<strong>in</strong>ggroup. This sensitiveness impacts differently on the resource usage of HiBOp andEpidemic, as shown by Figure 12. Similar remarks drawn with respect to reconfiguration<strong>in</strong>tervals apply also here. The higher the users sociability (high p r ), the higherthe mix between nodes and the forward<strong>in</strong>g opportunities. While Epidemic naivelyuses all these resources spread<strong>in</strong>g messages, HiBOp leverages nodes’ mix<strong>in</strong>g (andthe result<strong>in</strong>g spread of context <strong>in</strong>formation) to identify good paths more and moreaccurately.Figure 13 shows how data and non-data traffic contributes to the bandwidth overhead.As already said, Epidemic exploits all the possibilities of reach<strong>in</strong>g the desti-


<strong>Social</strong>-<strong>based</strong> <strong>autonomic</strong> <strong>rout<strong>in</strong>g</strong> <strong>in</strong> <strong>opportunistic</strong> <strong>networks</strong> 25[# of messages]Buffer Evolution807060504030201000 20 40 60 80 100% of simulation run2250 E2250 H9000 E9000 H36000 E36000 HFig. 9 Buffer occupation (focus on the reconfiguration parameter)nation by copy<strong>in</strong>g the messages on nodes as much as possible. This results <strong>in</strong> ahigh overhead, which is useless particularly for highly connected scenarios wherethere are a lot of forward<strong>in</strong>g opportunities. Note that the high Epidemic’s overheadessentially comes from the aggressive replication of messages (i.e., from datatraffic). Indeed, Figure 13(b) shows that the traffic related to forward<strong>in</strong>g (i.e., thetraffic related to the exchange of summary vectors) actually decreases when moreconnectivity opportunities are available. The buffer occupation curves (Figure 14)<strong>in</strong>dicate that for higher rewir<strong>in</strong>g, the buffers under Epidemic are less full, becausemessages can be delivered more quickly to the dest<strong>in</strong>ations. Therefore, the size ofsummary vectors decreases, and this expla<strong>in</strong>s the trend of Figure 13(b). However,the reduction <strong>in</strong> terms of forward<strong>in</strong>g traffic is overwhelmed by the aggressive spreadof message, which results <strong>in</strong> an <strong>in</strong>crease of the overhead related to the data traffic(Figure 13(a)) and, ultimately, to the overall overhead <strong>in</strong>crease Figure 12. UnlikeEpidemic, HiBOp “learns” the degree of connectivity of the network and uses thisknowledge for adjust<strong>in</strong>g the load. More specifically, HiBOp learns the current stateof the network through the exchange of context messages. As context <strong>in</strong>formationis spread more and more widely (rewir<strong>in</strong>g equal to 0.1 and 0.5) paths become moreand more known, and HiBOp reduces the exchanges of both data and non-data messages.Epidemic’s high resources consumption is confirmed by Figure 15. With Epidemic,between 50% and 70% of messages are spread through the whole network.Epidemic tends to exploit all opportunities, regardless of the sociality of users.Therefore, when nodes are more mixed (higher rewir<strong>in</strong>g), Epidemic floods the networkmore aggressively. As we will show when present<strong>in</strong>g the QoS performancefigures, this is basically useless and thus results <strong>in</strong> wast<strong>in</strong>g memory and bandwidthresources. HiBOp, <strong>in</strong>stead, is aware of the current state of the network and adjuststhe number of replicas of each packet <strong>based</strong> on the sociality of the network. Note


26 Chiara Boldr<strong>in</strong>i, Marco Conti, Andrea PassarellaP(copies>x)Copies CCDF10.90.80.70.60.50.40.30.20.10-0.10 5 10 15 20 25 30# of copies2250 E 9000 E 36000 E(a) EpidemicP(copies>x)Copies CCDF10.90.80.70.60.50.40.30.20.10-0.10 5 10 15 20 25# of copies2250 H 9000 H 36000 H(b) HiBOpFig. 10 Copies Distribution (focus on the reconfiguration parameter)that, even with the lowest sociality (rewir<strong>in</strong>g = 0.03), only about 30% of messagesare copied to more than ten nodes. Note also that, unlike Epidemic, this percentagedecreases to zero with higher levels of sociality.As far as the QoS performance figures (Table 3), aga<strong>in</strong> the packet loss is negligible(so we do not show it), while – as expected – the average delay decreasesas users become more social. The performance of HiBOp are still not far from thebound represented by Epidemic. It is also <strong>in</strong>terest<strong>in</strong>g to note (Figure 16) that thedelay of messages towards friend nodes tends to slightly <strong>in</strong>crease as users becomemore social, because they spend (on average) more time outside their home group.However, as shown by Table 3, the advantage of connect<strong>in</strong>g more efficiently users


<strong>Social</strong>-<strong>based</strong> <strong>autonomic</strong> <strong>rout<strong>in</strong>g</strong> <strong>in</strong> <strong>opportunistic</strong> <strong>networks</strong> 27[byte tx/byte snt]3432302826242220181614Bandwidth Overhead2250 9000 36000Reconfiguration Interval [s]HiBOpEpidemicFig. 11 Bandwidth overhead (focus on the reconfiguration parameter)60Bandwidth Overhead[byte tx/byte snt]402000.030.1Rewir<strong>in</strong>g Probability0.5HiBOpEpidemicFig. 12 Bandwidth overhead (focus on the rewir<strong>in</strong>g parameter)between groups as users become more social overwhelms the slight performancereduction experienced by friends.Table 3 Average delay (focus on the rewir<strong>in</strong>g parameter)p r HiBOp Epidemic0.03 170.86 ± 25.86 130.28 ± 20.59delay (s) 0.1 129.42 ± 12.51 83.20 ± 8.570.5 104.91 ± 8.87 73.69 ± 7.16


28 Chiara Boldr<strong>in</strong>i, Marco Conti, Andrea Passarella[byte tx/byte snt]4540353025201510Bandwidth Overhead - Data0.030.10.5Rewir<strong>in</strong>g ProbabilityHiBOp(a) Data messagesEpidemic[byte tx/byte snt]4.543.532.521.510.50Bandwidth Overhead - Non-data0.030.10.5Rewir<strong>in</strong>g ProbabilityHiBOpEpidemic(b) Non-data messagesFig. 13 Bandwidth overhead (focus on the rewir<strong>in</strong>g parameter)Mobility affects also the number of hops a message passes through before reach<strong>in</strong>gits f<strong>in</strong>al dest<strong>in</strong>ation (see Figure 17). As our setup simulates a social network,nodes belong<strong>in</strong>g to the same community are expected to meet more frequently andfor a longer time. This results <strong>in</strong> better QoS performances for messages dest<strong>in</strong>edto friends. As the network becomes more mixed, nodes tend to spend more timeoutside their community, thus becom<strong>in</strong>g good forwarders for messages dest<strong>in</strong>edoutside. The proximity between friends reduces as rewir<strong>in</strong>g <strong>in</strong>creases and more forward<strong>in</strong>ghops are needed <strong>in</strong> order to reach the dest<strong>in</strong>ation (Figure 17(a)). On theother hand, the proximity between non-friend nodes <strong>in</strong>creases and the number ofhops a message passes through decreases (Figure 17(b)).


<strong>Social</strong>-<strong>based</strong> <strong>autonomic</strong> <strong>rout<strong>in</strong>g</strong> <strong>in</strong> <strong>opportunistic</strong> <strong>networks</strong> 29[# of messages]Buffer Evolution7060504030201000 18000 36000 54000 72000 90000Simulated time [s]E 0.03H 0.03E 0.1H 0.1E 0.5H 0.5Fig. 14 Buffer occupation (focus on the rewir<strong>in</strong>g parameter)5.4 Break<strong>in</strong>g Closed GroupsIn this set of simulations we use a 3x3 grid with 9 groups of 5 nodes each. Justone node, located <strong>in</strong> the upper left cell sends messages, dest<strong>in</strong>ed to a node <strong>in</strong> thelower right cell. Recall that the only way a message can reach its f<strong>in</strong>al dest<strong>in</strong>ationis through edge contacts with nodes between which no social relation exists. Byvary<strong>in</strong>g nodes’ transmission range we can analyse how this edge effect impacts onforward<strong>in</strong>g. We use three values for the transmission range, i.e. 62.5m, 125m and250m. Therefore, nodes cover – on average – less than half a cell, slightly less thana cell, and one and a half cell.The bottoml<strong>in</strong>e of the results is that HiBOp is not suitable for <strong>networks</strong> with nosociability. At very small transmission ranges (62.5m) HiBOp is not able to deliveracceptable QoS (Table 4). HiBOp needs a m<strong>in</strong>imum number of contacts betweenusers to spread context <strong>in</strong>formation around. Indeed, at 125m HiBOp restores acceptableQoS at least <strong>in</strong> terms of packet loss, and is fully effective at 250m. Also<strong>in</strong> this case Epidemic and HiBOp behave differently with respect to the bandwidthoverhead (Figure 18). At 62.5m HiBOp seldom forwards messages. As context datais not circulat<strong>in</strong>g, nodes <strong>in</strong> the sender’s group are almost all equally fit to carry themessages closer to the dest<strong>in</strong>ation. At a high transmission range the context data iscirculat<strong>in</strong>g effectively, and therefore good paths can be identified soon. In the <strong>in</strong>termediatecases (e.g., transmission range equal to 125m) HiBOp is not (yet) ableto correctly learn the status of the network, and this results <strong>in</strong> a higher overheadwith respect to Epidemic. However, note that these results confirm that Epidemic isnot able to exploit rich connectivity scenarios without flood<strong>in</strong>g the network, s<strong>in</strong>ce it<strong>in</strong>creases its overhead at high transmission ranges.Figure 19 shows the average number of hops (recall that <strong>in</strong> this configurationstatistics are related to non-friend nodes only). We can see that Epidemic generates


30 Chiara Boldr<strong>in</strong>i, Marco Conti, Andrea PassarellaP(copies>x)Copies CCDF10.90.80.70.60.50.40.30.20.100 5 10 15 20 25 30# of copies0.03 E 0.1 E 0.5 E(a) EpidemicP(copies>x)Copies CCDF10.90.80.70.60.50.40.30.20.10-0.10 5 10 15 20 25# of copies0.03 H 0.1 H 0.5 H(b) HiBOpFig. 15 Copies distribution (focus on the rewir<strong>in</strong>g parameter)44 copies of each message, i.e. it replicates messages on all nodes, as it is not awareof the current state of the network. In HiBOp, the number of copies <strong>in</strong>creases ascontext <strong>in</strong>formation spreads, i.e., for <strong>in</strong>creas<strong>in</strong>g transmission ranges. This is becausewhen the transmission range is low there is no reason to replicate messages, s<strong>in</strong>ceno good path can be found <strong>in</strong> a context-aware scheme if context <strong>in</strong>formation cannotspread. As soon as context <strong>in</strong>formation can be exploited, paths can be found andHiBOp starts replicat<strong>in</strong>g messages. F<strong>in</strong>ally, Figure 20 shows the average number ofhops. In both cases this figure decreases with higher transmission ranges, as morecontact opportunities become available, and a s<strong>in</strong>gle hop is able to br<strong>in</strong>g messagescloser to the dest<strong>in</strong>ation.


<strong>Social</strong>-<strong>based</strong> <strong>autonomic</strong> <strong>rout<strong>in</strong>g</strong> <strong>in</strong> <strong>opportunistic</strong> <strong>networks</strong> 31120100HiBOpEpidemicAverage delay - Friend Dest.Delay [s]8060402003503000.030.1Rewir<strong>in</strong>g Probability(a) friends0.5Average delay - Non-friend Dest.HiBOpEpidemicDelay [s]250200150100500.03 0.1 0.5Rewir<strong>in</strong>g Probability(b) non-friendsFig. 16 Average delay (focus on the rewir<strong>in</strong>g parameter)Table 4 Users QoS (closed groups)range (m) HiBOp Epidemic62.5 61.41 ± 10.16 0 ± 0ploss (%) 125 0 ± 0 0 ± 0250 0 ± 0 0 ± 062.5 14732.57 ± 1242.74 535.50 ± 14.05delay (s) 125 576.40 ± 177.56 102.83 ± 1.82250 1.77 ± 0.55 23.58 ± 0.80


32 Chiara Boldr<strong>in</strong>i, Marco Conti, Andrea Passarella[# of hops]32.82.62.42.221.81.6Average number of hop - Friend Dest.HiBOpEpidemic0.030.1Rewir<strong>in</strong>g Probability(a) friends0.5[# of hops]4.44.243.83.63.43.232.82.62.42.2Average number of hop - Non-friend Dest.0.030.1HiBOpEpidemicRewir<strong>in</strong>g Probability(b) non-friends0.5Fig. 17 Average number of hops (focus on the rewir<strong>in</strong>g parameter)6 ConclusionsIn this chapter we have jo<strong>in</strong>tly presented results <strong>in</strong> two key research areas of the<strong>autonomic</strong> <strong>opportunistic</strong> network<strong>in</strong>g research field. Specifically, we have consideredmobility models <strong>based</strong> on users social relationships and behaviour, and contextaware<strong>rout<strong>in</strong>g</strong>. Mobility models are a cornerstone to design and evaluate <strong>rout<strong>in</strong>g</strong>protocols, as users mobility is one of the key enabler of end-to-end communication<strong>in</strong> <strong>opportunistic</strong> <strong>networks</strong>.We have discussed social-<strong>based</strong> mobility models <strong>in</strong> which users movements are<strong>based</strong> on social ties between people. We have highlighted that this <strong>in</strong>formation aloneis not sufficient to model relevant scenarios, and that it should be complemented


<strong>Social</strong>-<strong>based</strong> <strong>autonomic</strong> <strong>rout<strong>in</strong>g</strong> <strong>in</strong> <strong>opportunistic</strong> <strong>networks</strong> 33140Bandwidth Overhead[byte tx/byte snt]1201008060402062.5125Transmission Range [m]250HiBOpEpidemicFig. 18 Bandwidth overhead (closed groups)[# of copies]454035302520151050Average number of copies62.5 125 250Transmission Range [m]HiBOpEpidemicFig. 19 Average number of copies (closed groups)with <strong>in</strong>formation about the physical places where people spend their time due totheir social behaviour. We have presented a mobility model exploit<strong>in</strong>g both types of<strong>in</strong>formation, and shown its advantages through an analytical model.We have then considered <strong>rout<strong>in</strong>g</strong> issues, and how the social aspects of peoplebehaviour impact on context-aware <strong>rout<strong>in</strong>g</strong> protocols. Specifically, we have highlightedhow different approaches to <strong>rout<strong>in</strong>g</strong> <strong>in</strong> <strong>opportunistic</strong> <strong>networks</strong> are able to<strong>autonomic</strong>ally adapt to the dynamic scenarios result<strong>in</strong>g from humans’ mobility patterns.We have framed this work <strong>in</strong> the ongo<strong>in</strong>g research on <strong>rout<strong>in</strong>g</strong> for <strong>opportunistic</strong><strong>networks</strong>, and we have compared the performance figures of two protocols at the oppositeends of the spectrum as far as the use of context <strong>in</strong>formation is concerned,namely Epidemic and HiBOp.


34 Chiara Boldr<strong>in</strong>i, Marco Conti, Andrea Passarella109Average number of hop - Non-friend Dest.HiBOpEpidemic[# of hops]8765462.5 125 250Transmission Range [m]Fig. 20 Average number of hops (closed groups)With respect to the <strong>rout<strong>in</strong>g</strong> aspects, the results we have presented can be summarisedas follows. Context-<strong>based</strong> <strong>rout<strong>in</strong>g</strong> actually provides an effective congestioncontrol mechanism, and, with respect to dissem<strong>in</strong>ation-<strong>based</strong> <strong>rout<strong>in</strong>g</strong>, providesacceptable QoS with drastically lower overhead, unless <strong>in</strong> very adverse scenarios.Indeed, HiBOp is able to automatically learn the connectivity opportunities determ<strong>in</strong>edby users movement patterns, and exploit them efficiently. This <strong>autonomic</strong>,self-learn<strong>in</strong>g feature is completely absent <strong>in</strong> dissem<strong>in</strong>ation-<strong>based</strong> <strong>rout<strong>in</strong>g</strong> schemes.The results also suggest a hybrid scheme for <strong>networks</strong> with vary<strong>in</strong>g levels of user sociability.When groups are very isolated, context data cannot circulate, and cannot beused for tak<strong>in</strong>g effective forward<strong>in</strong>g decisions. In such cases, dissem<strong>in</strong>ation-<strong>based</strong>schemes seem the only way to enable communication between groups. As soon asusers become more social, context <strong>in</strong>formation spreads <strong>in</strong> the network, and context<strong>based</strong><strong>rout<strong>in</strong>g</strong> becomes a preferable solution. An <strong>in</strong>terest<strong>in</strong>g follow-up of this workis how to exploit context <strong>in</strong>formation to dist<strong>in</strong>guish these different scenarios andselect the appropriate <strong>rout<strong>in</strong>g</strong> scheme.From a complementary standpo<strong>in</strong>t, our results show that <strong>in</strong> <strong>opportunistic</strong> <strong>networks</strong>user sociability helps <strong>rout<strong>in</strong>g</strong>: users’ relationships outside their “home” communityallow context <strong>in</strong>formation to spread <strong>in</strong> the network, and make forward<strong>in</strong>gmore and more efficient. These results open <strong>in</strong>terest<strong>in</strong>g research directions. Actually,s<strong>in</strong>ce <strong>opportunistic</strong> <strong>networks</strong> build the network by exploit<strong>in</strong>g mobile devicespeople carry with them, look<strong>in</strong>g at social network theories to model users’ socialrelationships and exploit these models for design<strong>in</strong>g network protocols is a very <strong>in</strong>terest<strong>in</strong>gresearch direction. Indeed, the EC FET-PERADA SOCIALNETS project(started <strong>in</strong> February 2008) will be look<strong>in</strong>g at these aspects. Other <strong>in</strong>terest<strong>in</strong>g researchdirections <strong>in</strong>clude provid<strong>in</strong>g privacy and security support through distributedand scalable systems <strong>in</strong> <strong>opportunistic</strong> <strong>networks</strong>. Also, another challeng<strong>in</strong>g researchdirection is how to <strong>in</strong>tegrate purely <strong>in</strong>frastructure-less <strong>opportunistic</strong> <strong>networks</strong> (likethe ones we have considered <strong>in</strong> this chapter) with access po<strong>in</strong>ts to the Internet <strong>in</strong>fras-


<strong>Social</strong>-<strong>based</strong> <strong>autonomic</strong> <strong>rout<strong>in</strong>g</strong> <strong>in</strong> <strong>opportunistic</strong> <strong>networks</strong> 35tructure. F<strong>in</strong>ally, design<strong>in</strong>g data-management systems (built on top of <strong>opportunistic</strong><strong>rout<strong>in</strong>g</strong> schemes) to improve data availability <strong>in</strong> <strong>opportunistic</strong> <strong>networks</strong> is anotherdirection we f<strong>in</strong>d extremely important.Acknowledgements This work was partially funded by the European Commission under theHAGGLE (027918) FET-SAC Project, and under the SOCIALNETS (217141) FET-PERADAProject.References1. Boldr<strong>in</strong>i, C., Conti, M., Iacop<strong>in</strong>i, I., Passarella, A.: HiBOp: a History Based Rout<strong>in</strong>g Protocolfor Opportunistic Networks. Proc. of the IEEE International Symposium on a World ofWireless, Mobile and Multimedia Networks (WoWMoM 2007) (2007)2. Boldr<strong>in</strong>i, C., Conti, M., Passarella, A.: Impact of <strong>Social</strong> Mobility on Rout<strong>in</strong>g Protocols forOpportunistic Networks. In: Proc. of the First IEEE WoWMoM Workshop on Autonomic andOppoortunistic Network<strong>in</strong>g (AOC 2007) (2007)3. Boldr<strong>in</strong>i, C., Conti, M., Passarella, A.: Users Mobility Models for Opportunistic Networks:the Role of Physical Locations. Proc. of the IEEE Wireless Rural and Emergency CommunicationsConference (WRECOM 2007) (2007)4. Boldr<strong>in</strong>i, C., Conti, M., Passarella, A.: Autonomic behaviour of <strong>opportunistic</strong> network <strong>rout<strong>in</strong>g</strong>.Int. J. Autonomous and Adaptive Communications Systems 1(1) (2008)5. Boldr<strong>in</strong>i, C., Conti, M., Passarella, A.: Exploit<strong>in</strong>g users social relations to forward data <strong>in</strong><strong>opportunistic</strong> <strong>networks</strong>: The HiBOp solution. Pervasive and Mobile Comput<strong>in</strong>g (2008)6. Borgia, E., Conti, M., Delmastro, F., Pelusi, L.: Lessons from an Ad-Hoc Network Test-Bed:Middleware and Rout<strong>in</strong>g Issues. Ad Hoc & Sensor Wireless Networks, An International Journal1(1-2) (2005)7. Burgess, J., Gallagher, B., Jensen, D., Lev<strong>in</strong>e, B.: MaxProp: Rout<strong>in</strong>g for Vehicle-BasedDisruption-Tolerant Networks. Proc. of the 25th IEEE Annual Jo<strong>in</strong>t Conference of the IEEEComputer and Communications Societies (INFOCOM 2006) (2006)8. Burns, B., Brock, O., Lev<strong>in</strong>e, B.: MV <strong>rout<strong>in</strong>g</strong> and capacity build<strong>in</strong>g <strong>in</strong> disruption tolerant<strong>networks</strong>. Proc. of the 24th IEEE Annual Jo<strong>in</strong>t Conference of the IEEE Computer and CommunicationsSocieties (INFOCOM 2005) (2005)9. Cha<strong>in</strong>treau, A., Hui, P., Crowcroft, J., Diot, C., Gass, R., Scott, J.: Impact of Human Mobilityon Opportunistic Forward<strong>in</strong>g Algorithms. IEEE Transactions on Mobile Comput<strong>in</strong>g 6(6),606–620 (2007)10. Conti, M., Giordano, S.: Multihop Ad Hoc Network<strong>in</strong>g: The Reality. IEEE CommunicationsMagaz<strong>in</strong>e 45(4), 88–95 (2007)11. Conti, M., Giordano, S.: Multihop Ad Hoc Network<strong>in</strong>g: The Theory. IEEE CommunicationsMagaz<strong>in</strong>e 45(4), 78–86 (2007)12. D.J.Watts: Small Worlds The Dynamics of Networks between Order and Randomness. Pr<strong>in</strong>cetonStudies on Complexity, Pr<strong>in</strong>ceton University Press (1999)13. Doria, A., Uden, M., Pandey, D.: Provid<strong>in</strong>g connectivity to the Saami nomadic community.In: Proc. of the 2nd International Conference on Open Collaborative Design for Susta<strong>in</strong>ableInnovation (DYD 2002) (2002)14. Fall, K.: A delay-tolerant network architecture for challenged <strong>in</strong>ternets. Proc. of the 2003ACM conference on Applications, technologies, architectures, and protocols for computercommunications (SIGCOMM 2003) pp. 27–34 (2003)15. Ghosh, J., Beal, M.J., Ngo, H.Q., Qiao, C.: On profil<strong>in</strong>g mobility and predict<strong>in</strong>g locations ofwireless users. In: Proc. of ACM REALMAN (2006)


36 Chiara Boldr<strong>in</strong>i, Marco Conti, Andrea Passarella16. Grossglauser, M., Tse, D.N.C.: Mobility <strong>in</strong>creases the capacity of ad hoc wireless <strong>networks</strong>.IEEE/ACM Trans. Netw. 10(4), 477–486 (2002). DOI http://dx.doi.org/10.1109/TNET.2002.80140317. Grossglauser, M., Vetterli, M.: Locat<strong>in</strong>g nodes with EASE: last encounter <strong>rout<strong>in</strong>g</strong> <strong>in</strong> ad hoc<strong>networks</strong> through mobility diffusion. Proc. of the 22nd IEEE Annual Jo<strong>in</strong>t Conference of theIEEE Computer and Communications Societies (IEEE INFOCOM 2003) (2003)18. Guo, S., Falaki, M.H., Oliver, E.A., Rahman, S.U., Seth, A., Zaharia, M.A., Keshav, S.: Verylow-cost <strong>in</strong>ternet access us<strong>in</strong>g kiosknet. SIGCOMM Comput. Commun. Rev. 37(5), 95–100(2007). DOI http://doi.acm.org/10.1145/1290168.129018119. j. Hsu, W., Spyropoulos, T., Psounis, K., Helmy, A.: Model<strong>in</strong>g Time-Variant User Mobility <strong>in</strong>Wireless Mobile Networks. In: Proc. of IEEE Infocom (2007)20. Hui, P., Crowcroft, J., Yoneki, E.: Bubble rap: social-<strong>based</strong> forward<strong>in</strong>g <strong>in</strong> delay tolerant <strong>networks</strong>.In: Proceed<strong>in</strong>gs of the 9th ACM <strong>in</strong>ternational symposium on Mobile ad hoc network<strong>in</strong>gand comput<strong>in</strong>g, pp. 241–250. ACM New York, NY, USA (2008)21. Hui, P., Yoneki, E., Chan, S., Crowcroft, J.: Distributed community detection <strong>in</strong> delay tolerant<strong>networks</strong>. Proceed<strong>in</strong>gs of the 2nd ACM International Workshop on Mobility <strong>in</strong> the Evolv<strong>in</strong>gInternet Architecture (MobiArch 2007) (2007)22. Ja<strong>in</strong>, S., Fall, K., Patra, R.: Rout<strong>in</strong>g <strong>in</strong> a delay tolerant network. In: Proc. of the 2004 ACMConference on Applications, technologies, architectures, and protocols for computer communications(SIGCOMM 2004), pp. 145–158. ACM, New York, NY, USA (2004). DOIhttp://doi.acm.org/10.1145/1015467.101548423. J<strong>in</strong>dal, A., Psounis, K.: Contention-aware analysis of <strong>rout<strong>in</strong>g</strong> schemes for mobile <strong>opportunistic</strong><strong>networks</strong>. Proc. of the 1st <strong>in</strong>ternational ACM MobiSys workshop on Mobile <strong>opportunistic</strong>network<strong>in</strong>g (MobiOpp 2007) pp. 1–8 (2007)24. Juang, P., Oki, H., Wang, Y., Martonosi, M., Peh, L., Rubenste<strong>in</strong>, D.: Energy-efficient comput<strong>in</strong>gfor wildlife track<strong>in</strong>g: design tradeoffs and early experiences with ZebraNet. ACMSIGPLAN Notices 37(10), 96–107 (2002)25. Kim, M., Kotz, D., Kim, S.: Extract<strong>in</strong>g a mobility model from real user traces. In: Proc. ofInfocom (2006)26. Leguay, J., Friedman, T., Conan, V.: Evaluat<strong>in</strong>g mobility pattern space <strong>rout<strong>in</strong>g</strong> for dtns. Proc.of the 25th IEEE Annual Jo<strong>in</strong>t Conference of the IEEE Computer and Communications Societies(INFOCOM 2006) pp. 1–10 (2006). DOI 10.1109/INFOCOM.2006.29927. Leguay, J., L<strong>in</strong>dgren, A., Scott, J., Friedman, T., Crowcroft, J.: Opportunistic content distribution<strong>in</strong> an urban sett<strong>in</strong>g. Proc. of the 2006 SIGCOMM workshop on Challenged <strong>networks</strong>(CHANTS 2006) pp. 205–212 (2006)28. Lelescu, D., Kozat, U.C., Ja<strong>in</strong>, R., Balakrishnan, M.: Model T++: an empirical jo<strong>in</strong>t spacetimeregistration model. In: Proc. of ACM MobiHoc (2006)29. L<strong>in</strong>dgren, A., Doria, A., Schelen, O.: Probabilistic <strong>rout<strong>in</strong>g</strong> <strong>in</strong> <strong>in</strong>termittently connected <strong>networks</strong>.ACM Mobile Comput<strong>in</strong>g and Communications Review 7(3), 19–20 (2003)30. Musolesi, M., Hailes, S., Mascolo, C.: Adaptive <strong>rout<strong>in</strong>g</strong> for <strong>in</strong>termittently connected mobilead hoc <strong>networks</strong>. Proc. of the IEEE International Symposium on a World of Wireless, Mobileand Multimedia Networks (WoWMoM 2005) pp. 183–189 (2005)31. Pelusi, L., Passarella, A., Conti, M.: Beyond manets: Dissertation on <strong>opportunistic</strong> network<strong>in</strong>g.Tech. rep., IIT-CNR, http://bruno1.iit.cnr.it/˜andrea/tr/commag06_tr.pdf (2006)32. Pelusi, L., Passarella, A., Conti, M.: Opportunistic Network<strong>in</strong>g: Data Forward<strong>in</strong>g <strong>in</strong> DisconnectedMobile Ad Hoc Networks. IEEE Communications Magaz<strong>in</strong>e 44(11) (2006)33. Pelusi, L., Passarella, A., Conti, M.: Handbook of Wireless Ad hoc and Sensor Networks,chap. Encod<strong>in</strong>g for Efficient Data Distribution <strong>in</strong> Multi-hop Ad hoc Networks. Wiley andSons Publisher (2007)34. Pentland, A., Fletcher, R., Hasson, A.: DakNet: reth<strong>in</strong>k<strong>in</strong>g connectivity <strong>in</strong> develop<strong>in</strong>g nations.IEEE Computer 37(1), 78–83 (2004)35. Spyropoulos, T., Psounis, K., Raghavendra, C.: Efficient <strong>rout<strong>in</strong>g</strong> <strong>in</strong> <strong>in</strong>termittently connectedmobile <strong>networks</strong>: The multiple-copy case,”. ACM/IEEE Transactions on Network<strong>in</strong>g 16(2007)


<strong>Social</strong>-<strong>based</strong> <strong>autonomic</strong> <strong>rout<strong>in</strong>g</strong> <strong>in</strong> <strong>opportunistic</strong> <strong>networks</strong> 3736. T.Camp, J., V.Davies: A survey of mobility models for ad hoc network research. WirelessCommunication and Mobile Comput<strong>in</strong>g 2(5) (2002)37. Tuduce, C., Gross, T.: A mobility model <strong>based</strong> on WLAN traces and its validation. In: Proc.of IEEE Infocom (2005)38. Vahdat, A., Becker, D.: Epidemic Rout<strong>in</strong>g for Partially Connected Ad Hoc Networks. Tech.Rep. CS-2000-06, CS. Dept. Duke Univ. (2000)39. Widmer, J., Le Boudec, J.: Network cod<strong>in</strong>g for efficient communication <strong>in</strong> extreme <strong>networks</strong>.Proc. of the ACM SIGCOMM 2005 Workshop on Delay Tolerant Network<strong>in</strong>g (WDTN 2005)pp. 284–291 (2005)40. Yoneki, E., Hui, P., Chan, S., Crowcroft, J.: A socio-aware overlay for publish/subscribe communication<strong>in</strong> delay tolerant <strong>networks</strong>. Proceed<strong>in</strong>gs of the 10th ACM Symposium on Model<strong>in</strong>g,analysis, and simulation of wireless and mobile systems (MSWiM 2007) pp. 225–234(2007)41. Zhang, Z.: Rout<strong>in</strong>g <strong>in</strong> <strong>in</strong>termittently connected mobile ad hoc <strong>networks</strong> and delay tolerant<strong>networks</strong>: overview and challenges. IEEE Communications Surveys & Tutorials 8(1), 24–37(2006)

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!