DelQue: A Socially-Aware Delegation Query Scheme in Delay Tolerant Networks

Transcription

1 DelQue: A Socially-Aware Delegation Query Scheme in Delay Tolerant Networks Jialu Fan Jiming Chen Yuan Du Ping Wang Youxian Sun State Key Lab. of Industrial Control Technology, Zhejiang University, China School of Computer Engineering, Nanyang Technological University, Singapore {jlfan, jmchen, ydu, yxsun}@iipc.zju.edu.cn; wangping@ntu.edu.sg 1 Abstract In Delay Tolerant Networks (DTNs), information search is an important topic, which has not been widely investigated yet. Recently, pervasive and ubiquitous computing scenarios are pushing situations where it is the recipient (instead of the sender) who determines whether and how to seize data flowing in the network. Most existing information search schemes only focus on publish-subscribe paradigm or consider query and response separately, which are all in the form of hop-by-hop. In this paper, we propose DelQue, a novel socially-aware onehop delegation query scheme considering query and response integratedly. DelQue does not require the selected relays to forward query to other users, thus the network overhead can be lowered and source can ensure system performance with only local knowledge. We also explore the geographic regularity on mobility pattern of sociallyrelated people, and further employ a semi-markov model to predict user mobility. Such technique does not require the storage of the entire past history of the system and is computationally lightweight, making it suitable for a resource-scarce mobile setting. Extensive realistic trace-driven simulations show that DelQue allows for maintaining a very high and steady information query rate with extremely low overhead. Index Terms Delay tolerant networks, social networks, user mobility, information search, information query. I. INTRODUCTION Thanks to the popularity of personal hand-held mobile devices (e.g., MP3 players, PDAs and smartphones), people of similar interests or commonalities in intermittently-connected human networks can cooperate to establish network connectivity and communicate with each other in the absence of network infrastructure [1], [2]. Due to such ad hoc characteristics, Delay/Disruption Tolerant Networks (DTNs) [3] can be deployed in a number of critical areas, including battlefield operations, vehicular ad hoc networks and disaster response scenarios [4], [5]. The initial work in DTNs always focuses on routing schemes [6], [7], [8]. However, the ability to access information rapidly and conveniently is also an important feature that DTNs should have since

2 2 the ultimate goal of having such networks is to allow mobile users to access information quickly and efficiently. For example, soldiers in a battlefield need to access information related to commands from the general, detailed geographical maps, information about enemy locations, weather information etc. In order to bring these applications to the domain of DTNs, a search scheme is therefore required that works despite the unreliable network conditions. The traditional approach to map search queries to resources is mainly based on centralized search engines, e.g., Google, MSN, and Yahoo. The problem of such centralized schemes is not only the high costs of index maintenance, but also the prerequisite that the clients are able to access both of the index and the locations in which the search content resides. Recently, decentralized search approaches in peerto-peer environments have also been developed, but they still depend on good connectivity among the users. Otherwise, intermittent connections may result in a high maintenance cost of the index. The main challenge of designing information search schemes in DTNs is to overcome the intermittent connectivity among users, which is also a difficult problem in routing scheme design. However, data flowing in routing schemes is in the form of unidirectional, because data is forwarded from source to destination hop by hop. The design of routing schemes focuses on how to facilitate the data forwarding rapidly (low latency) and efficiently (low cost, i.e., minimal relay usage). In some social-related schemes, data is always greedily forwarded to users with higher social metrics (e.g., contact opportunity to destination, users willingness to forward etc.) through multiple hops [7], [9]. In an information search scheme, however, information feedback is ultimately needed to match the query. Therefore, data flowing in search schemes is in the form of bidirectional, which makes the solutions more complex. The existing information access schemes in DTNs always employ publish-subscribe (pub-sub) paradigm [1], [11], in which publisher aims to publish message to the interested subscribers, at the same time subscriber also intends to retrieve the relevant message from the network. However, pub-sub aims to match the interest messages to the relevant subscribers in a distributed manner thanks to its decoupling properties, which applies to the environment that publishers and subscribers are both expecting to find each others. In this paper, we focus on the information search scheme that mobile users use to initialize the interestbased query and then delegate neighbors to get the relevant information back. We refer to this as DelQue

3 3 (delegation query). [12] and [13] also consider such application scenario, but they consider query and response separately. In either step, data (query or message) is forwarded greedily by multi-hops. However, DelQue integratedly considers information query and response by devising a corresponding utility metric to users in the network. The novelty and contributions of this paper are two-fold. 1) DelQue is in a one-hop and closed-loop form, i.e., relays take charge of both querying the relevant interest information and getting it back to the source. Therefore, the selection of relays is based on their utility to fulfill the whole task. DelQue does not require the selected relays to forward the query to other users, hence the network overhead can be lowered and source can ensure system performance with only local knowledge. 2) DelQue presents a spatio-temporal prediction method of user mobility based on semi-markov model to identify the best query relays. Users in the network only need to maintain several critical parameters instead of the entire past mobility history. Therefore, the technique is computationally lightweight, which makes it suitable for a resource-scarce mobile setting. The remainder of this paper is structured as follows. Section II provides an overview about the background of DTNs, the big picture and basic idea of DelQue. Then we elaborate the details of DelQue scheme in Section III. Section IV presents the spatio-temporal prediction of user mobility based on semi- Markov model. We can compute users utilities for query using such predict method. Section V evaluates the performance of spatio-temporal prediction and DelQue scheme via realistic trace-driven simulation. The last two sections present related work and conclusions, respectively. A. geo-community in DTNs II. OVERVIEW From the social network perspective, people with mutual interesting properties (e.g., common hobbies, social functions, and occupations) always tightly link to each other to form a community [7], [9]. On campus, faculty working in the same department interact more frequently with each other; members affiliating with the same club such as baseball or salsa, have such heavy interactions. In an academic conference, scholars having mutual research interests also make up a community.

4 4 Interests often highly relate to geography in human society [14]: Officemates contact each other in the office; baseball lovers play baseball together in gyms; scholars share and communicate their research interests in conferences. In this paper, we focus on geography-related DTNs and borrow geography-related community (geo-community) from [14]. 1 Since people have varying roles in society, a user can be a mutual member of several different communities, and move around those well-visited geo-communities instead of moving randomly. For example, consider a mother in family, who loves volleyball and swimming, and works as an accountant in an IT corporation. She therefore affiliates to and moves around the above four geo-communities: Family, Volleyball Court, Natatorium and IT Corporation. Furthermore, the user s dwell time at each geocommunity is fairly regular, because their social behavior patterns usually remain stable in a relatively long interval. This means a user has his/her own mobility schedule and he/she generally moves among geo-communities according to that schedule, subject to few random deviations. In this paper, we consider a DTN with a finite number of mobile users that move mostly among a set of geo-communities. Suppose geo-community and interest have one-to-one correspondence relationship with each other. A geo-community is a place where users can communicate directly [9], i.e., any two users that are located at a geo-community at the same time can establish contact to exchange messages. Users at different geo-communities cannot establish contact. B. Big picture We consider the following scenario in DTNs: Eric takes charge of security surveillance of a university campus. Today, a dangerous place is the College Theater, since where will perform an opera. Unfortunately, Eric has been trapped in an urgent task so that he cannot go there himself. Hence, he has to delegate his neighbors who can get there and respond to him somewhere within working time to help. However, Eric aims to delegate as little number of friends as possible to accomplish the query task. Fig. 1 gives the DelQue solution to the above example. Located at the same geo-community (Office) with Eric are Rachel, Kris, Jerry and Thomas. Based on the history mobility information, Eric predicts 1 We will use the terms community and geo-community interchangeably in subsequent sections.

5 5 Theater Dormitory Theater Dormitory Theater Dormitory Jerry Thomas Classroom Store Classroom Thomas Eric Store Classroom Jerry Eric Store Office Office Rachel Office Kris Jerry Rachel Eric Thomas Kris Rachel Kris (a) (b) (c) Fig. 1. An example to illustrate the process of DelQue. The dashed line stands for forwarding packets. that within working time Jerry and Thomas have better query and response performance than all other users in the Office. Therefore, Eric forwards the query to Jerry and Thomas. In this scenario, Jerry goes to the College Theater geo-community later (Fig. 1(b)) and then meets Eric in Classroom for information response (Fig. 1(c)). Note that the above stated story is a representative example of how DelQue scheme works. It does not mean DelQue only applies to this scenario. As a matter of fact, DelQue can be easily used in data query from a special data center, score query from a certain department, etc. C. The basic idea The basic idea of our approach is to develop social-based metrics based on the probabilities of users arriving destination geo-community to query the interest information and then getting it back to the source. Based on the social-based metrics, we formulate the relay selections as a knapsack problem: N min s.t. m=1 x m N u m x m U m=1 (1)

6 6 where x m {, 1} represents whether User m is selected as the relay, and the constraint U indicates that the selected relays should satisfy the performance requirements of the source in terms of query ratio and actual delay. Note that cost-aware relay selection schemes always employ such knapsack problem [14], [15]. Social-based metrics will be developed to calculate the utility u m associated with each User m in the network, and the total required utility U is determined by the performance requirements of the source. The remainder of this paper therefore focuses on answering the following questions: 1) What are the appropriate social-based metrics for DelQue? 2) How to calculate the utilities u m of individual users? 3) How can the source calculate the total required utility U? III. DELQUE SCHEME Assuming the following setting: a collection of geo-communities S = {1, 2,..., J} over a domain of mobile users M = {1, 2,..., N}. Users generate queries over time, and each query has a certain lifetime (i.e., TTL). A query item is in the form of < σ, S(t), D, T >, where σ is the source user ID, S(t)( t T ) indicates the geo-community where σ stays at time t, and S(t) is always a piecewise function over S; D is the destination geo-community ID corresponding to the query interest, and T is the deadline. σ only cares the neighbors query performance before T, which also represents the expiration time that query and data item are removed from storage users when they expire. At random times users come into the same geo-community with σ, meaning that σ is capable of delegating them to query. In our analysis we make no assumptions about the time instants when queries are generated or the time needed for transmission. Users do not possess any a priori knowledge of the number of users in the system or knowledge of any properties of the other users. The metrics we are concerned with are: (1) cost, which is the number of replicas per generated query in the network, i.e., the number of relays delegated to query the interest information; (2) query rate, which is the fraction of generated queries for which at least one replica is eventually responded; and (3) actual delay, which is the average duration between a query s generation and the first response of one of its replicas. By high performance we mean high query rate, low cost and low average delay.

7 7 Suppose p m be the probability that the neighbor User m of the source σ can fulfill the delegation task before deadline T, i.e., query the relevant interest information and then get it back to σ. Therefore, for User m, let the random variable T 1 be the time for User m to move to the destination geo-community D, and T 2 be the time for User m to get information back to the source σ, then the probability that σ can query the information through User m before T is P (T 1 + T 2 T ). Assuming that the probability density functions (PDF) of T 1 and T 2 are f 1 (t) and f 2 (t) (t ), respectively, P (T 1 + T 2 T ) is calculated through the convolution f 1 (t) f 2 (t) as p m = P (T 1 + T 2 T ) = T f 1 (t) f 2 (t)dt = T ( t f 1 (τ)f 2 (t τ)dτ)dt. (2) The computation of p m is detailed in Sec. IV-C and based on predicting user mobility over a finite time horizon. In contrast with prior predicting algorithms, that mainly focus on estimating the probability of contact regardless of the contact time, DelQue presents a spatio-temporal prediction of user mobility based on a time-homogeneous semi-markov process to determine the probabilities of contact for each time unit. The semi-markov modeling does not require the storage of the entire history but only several critical parameters of the system, which is computationally lightweight and suitable for a resource-scarce mobile setting. To ensure that the query ratio for σ is higher than p, the probability that all the delegated neighbors cannot fulfill the task should be lower than 1 p, i.e., N (1 p m ) xm 1 p (3) m=1 where x m {, 1} indicates whether User m is selected as the relay. Such problem can be transformed to the knapsack formulation in Eq. (1) by taking logarithms on both sides of the inequality, where u m = log 1 1 p m and U = log 1 1 p. Since σ is in contact with its neighbors, he/she can request each neighbor to calculate p m locally. Note that when σ generates the query, it is possible that all the neighbors in contact can not accomplish the query task, then he/she will forward the query replica to the immediate neighbors with p m > λ and

8 8 wait for the new neighbors until the performance requirement is achieved. We introduce λ to exclude the neighbors with lower query ratio to be selected as relays. The value of λ could be different depending on the application scenarios, and the effect of λ on DelQue performance is studied in Sec. V-E. Note that only best-effort solution is available if N m=1 u m < U. In this paper, we study two cases: (a) Querying for Static Source (QSS): The source σ stays at the original geo-community until deadline T (i.e., S(t) = S(), for t T ), which is a special case of querying for mobile source. We study this scenario firstly because this situation happens frequently in real world, e.g., Eric will be trapped in the office for the urgent task within the remaining working time. At the same time, the computation overhead of this situation is lower than that of querying for mobile source. QSS is then generalized as follows. (b) Querying for Mobile Source (QMS): The source σ moves according to a given schedule after delegation, and he/she exactly knows where he/she will be at any time (i.e., S(t) is known in advance by σ for t T ). That is reasonable in real world, because people are usually aware of their own schedule. However, σ does not have the knowledge of other users exact schedule due to privacy etc. Hence, σ has to send his/her schedule to neighbors and ask them to compute the probability that they can fulfill the query task before deadline T. IV. COMPUTING UTILITIES FROM SOCIAL PATTERNS AND MOBILITY To compute users utilities for query, we should first model users mobility in DTNs. As users in DTNs always belong to several communities, they usually move around those well-visited geo-communities. Hence, we can model user mobility as Markov process, where the state space can be represented by the set of geo-communities. Specifically, we use semi-markov processes rather than standard Markov chains, because the sojourn time during which a user is associated with a community can take various forms, besides the exponential or geometric distributions [16], [17]. Semi-Markov process was first introduced by Lee et al. [18] to model user mobility in WLANs. They focused on analyzing load balance among APs, and explored the characteristics of user mobility using the trace data collected from laptops but not hand-held cellphones, which weakened the effectiveness of modeling real-time user mobility. A similar model has also been used in [19] to present a routing protocol in DTNs, where they employed a synthetic

9 9 model to analyze the performance of the presented protocol. Moreover, [18] and [19] did not consider any social network concept. [14] has presented an active data dissemination scheme and focused on the steady-state analysis of semi-markov model. In this section, we describe the semi-markov processes we employ to model user mobility in DTNs, present a spatio-temporal prediction method based on the devised semi-markov model, and further propose the computation methods of social utilities for the relays selection. For simplicity, we assume the renewal process is time homogeneous during the period in which the mobility model is built. A. Time-Homogeneous semi-markov Model We consider user s mobility as a Markov renewal process {(X n, T n ) : n }, where T n is the time instant of the n-th transition (T = ) and X n S is the state at the n-th transition. The state space is represented by the set of geo-communities S = {1, 2,..., J}. A user that moves between two geocommunities transits in the markov process between the corresponding states. We assume the transition probabilities between states have the markov memoryless property, which means that the probability of a user to transit from state X n to state X n+1 is independent of state X n 1. Thus, process (X n ) is a standard markov chain. Random variable T n+1 T n describes the geo-communities sojourn time. Then, the associated time homogeneous semi-markov kernel Q is defined by: Q ij (t) = P r(x n+1 = j, T n+1 T n t X n = i) = p ij H ij (t), i, j S (4) Suppose P = [p ij ] is the transition probability matrix of the embedded Markov chain, where the transition probability from state i to state j is p ij = lim t Q ij (t) = P r(x n+1 = j X n = i) (5) We also derive the sojourn time distribution in state i when the next state is j. H ij (t) = P r(t n+1 T n t X n+1 = j, X n = i) (6) Then the sojourn time probability distribution in state i regardless of the next state is

10 1 D i (t) P r(t n+1 T n t X n = i) Note that the distribution of the sojourn time, D i (t), during which the user is associated with geo- Community i before his/her next transition takes place can be expressed as J D i (t) = Q ij (t). j=1 Now we define the time-homogeneous semi-markov process as X = (X t, t R + ), with the transient distributions ϕ ij (t) P r(x t = j X = i) J =(1 D i (t))δ ij + =(1 D i (t))δ ij + t l=1 t where δ ij represents the Kronecker δ function defined by {, for i j δ ij = 1, for i=j. J l=1 ϕ lj (t τ)dq il (τ) Q il (τ)ϕ lj (t τ)dτ, (7) (8) B. Spatio-Temporal Prediction of User Mobility Exploring spatio-temporal characteristics of user mobility is useful for computing user s utility for a certain interest query. In order to obtain the probability distribution of user s future location, we have to compute the transient distributions, ϕ ij (t), of the semi-markov model in the evolution equations Eq. (7). A numerical solution has already been proposed in [2]. Specifically, the evolution equations can be re-written for the discrete-time homogeneous semi-markov process as ϕ ij (k) = d ij (k) + J k v il (τ)ϕ lj (k τ), (9) l=1 τ=1

11 11 where d ij (k) = (1 D i (kh))δ ij, v ij (k) = h Q ij (kh), and h is the discretization step. Since Q ij (t) is not given in a closed form, but is obtained in the form of histogram distribution from experimental data, we need to further approximate Q ij (kh) in the expression of v ij as v ij (k) = { Qij (h), for k=1 Q ij (kh) Q ij ((k 1)h), for k>1, (1) Using the assumption that the geo-community sojourn time random variables are independent from the embedded state transition process (X ij ), we derive: Q ij (k) =P (X n+1 = j, T n+1 T n k X n = i) =P (X n+1 = j X n = i) P (T n+1 T n k X n+1 = j, X n = i) (11) =p ij H ij (k) where p ij represents the transition probability from state i to state j, H ij (k) is the probability that User m will move from geo-community i to geo-community j at, or before time k. The values of p ij and H ij (k) can be obtained from the trace data based on the method proposed in [14], and we can further compute D i (kh) = J j=1 Q ij(k). Therefore, all the values required in Eq. (9) have been computed, and User m can readily compute the transient distributions ϕ ij (k). Based on the Markov property of the underlying processes, if the state i of a user is known at time, then at time k >, the probability of that user being in state j is ϕ ij (k). Therefore, distributions ϕ ij (k) give the probability that the future location at time k of User m will be geo-community j considering that at time the location was geo-community i. Upon receiving the request from σ, User m can predict his/her own future location at time k with only two parameters, sojourn time probability distributions H ij and the transition probability matrix P. Therefore, each user in the network only need to calculate and maintain the above two parameters based on its own mobility history. Obviously, DelQue is a lightweight information search scheme. C. Computation of Utilities In this section we propose the computation of p m defined in Sec. III for QSS and QMS, respectively.

12 12 1) Computing Utilities for QSS: As introduced before, the source σ stays at the same geo-community from time to T, we set the geo-community ID as S, and the interest destination geo-community ID is D. We employ the discrete-time homogeneous semi-markov model, hence the convolution equations can be re-written as T k 1 p m = [f 1 (τ)f 2 (k τ)], (12) k=2 τ=1 To compute f 1 and f 2, we introduce ˆϕ ij (k), which represents the probability that User m is in geo- Community j at time k given his current location is geo-community i, without going back to Community i during time [, k], whose computation is similar with ϕ ij (k) and detailed as follows J k ˆϕ ij (k) = v il (τ)ϕ lj (k τ), r, j i (13) l=1 τ=1 Different from Eq. (9), here we omit d ij (k) because d ij (k) = when j i. Once a semi-markov mobility model has been constructed based on the association patterns of a user, we can calculate ˆϕ ij (k) by the discrete-time evolution equations given in Eq. (13). With ˆϕ ij (k) s, we can calculate f 1 (k) = ˆϕ SD (k), f 2 (k) = ˆϕ DS (k) and further calculate neighbors p m to select relays according to Eq. (3), i.e., calculate their utility u m and base on Eq. (1) to choose relays. 2) Computing Utilities for QMS: In QMS, we define the related parameters as p m, f 1(k) and f 2(k). They also have the following relationships, T k 1 p m = [f 1(τ)f 2(k τ)], (14) k=2 τ=1 Then, the computation of f 1(k) is the same as f 1 (k) in QSS, i.e., f 1(k) = ˆϕ S()D (k) (15) The computation method of f 2(k) is more complex due to the introduced movement of source σ, compared with f 2 (k) in QSS.

13 13 We use ˆϕ ij (k) to compute f 2(k). Assuming that trajectories of the source σ is known, then we can further compute the probability ˆϕ S()S(k) (k) that User m and σ will meet again based on the devised spatio-temporal prediction method, where S(k) indicates the community where the source σ will stay at time k, which is known as a priori for σ according to the assumption, and is included in the utility computation request from σ. For our study, we care about the probability that σ and User m re-encounter at the first time. Note that when we talk about the first re-encounter anywhere at time k, it means that they had no contacts between (, k). Assuming that user trajectories are independent, the probability f 2(k) of the first re-encounter at time k is defined as, f 2(k) = ˆϕ k 1 S()S(k) (k) (1 ˆϕ S()S(k) (k)), k > (16) t= Plugging Eq. (15) and Eq. (16) together into Eq. (14), we obtain the query probability p m in QMS. A. Experimental Traces V. PERFORMANCE EVALUATION We use two experimental traces collected from realistic DTNs to study the accuracy of spatio-temporal prediction method based on semi-markov model, and to evaluate the performance of DelQue schemes. The MIT Reality dataset [21] comes from an experimental trace involving 97 people for the duration of 9 months. Each participant carries a Nokia66 smart phone with a software that periodically detects his/her peers or neighboring Access Points (APs) via the Bluetooth interfaces, and a contact is recorded when the device moves close to other users or APs. The Infocom 6 dataset [22] contains opportunistic Bluetooth contacts between 98 imotes, 78 of which were distributed to Infocom6 participants and 2 of which with external antennas (providing longer range) were deployed at several places at the conference venue to act as APs. As summarized in Table. I, the two traces differ in their scale, detection period, as well as the contact density and duration. Concretely, we use the syslog data for mobile users association patterns to APs. Each syslog message contains a timestamp in seconds, the client s MAC address, the AP id, and the event type. From these

14 14 TABLE I TRACE SUMMARY Trace Infocom 6 MIT Reality Device imote Smart Phones Network type Bluetooth Bluetooth Duration (days) Granularity (seconds) 12 3 No. of devices Pairwise contact frequency (per day) syslog messages, the mobility of each user is extracted in the form of a series of two tuples (AP id, the timestamp when the association with this AP occurs). In our work, the static APs can help identify geocommunities thanks to their geography-related property. Specifically, we use the neighborhood of an AP to represent a geo-community, and consider the sojourn time of a participant spending at a geo-community as the time interval of his/her two consecutive contacts with different APs, the former of which is the corresponding geo-community. B. geo-community Clustering A major problem with the MIT Reality dataset is that sometimes users experience frequent re-associations between two APs in a short period of time. This phenomenon is also addressed in the Wi-Fi networks referred to as ping-pong effect. However, such phenomenon seldomly occurs in the Infocom 6 dataset because the sparse APs deployment. Community i Community j Fig. 2. Examples of oscillating transitions We employ the method of noise cancelation in filter technology to eliminate the negative effect of oscillating AP-association in geo-community identification. Oscillating transitions occur when a user is geographically located among two neighboring APs. We define such oscillation as follows: For any APs

15 15 i and j, if a user makes a sequence of transitions i j i j in a short period of time. Fig. 2 depicts the corresponding scenario in which oscillating transitions described in the above case may occur. Fig. 3 gives the cumulative distribution of users against their oscillating ratios, which is defined as the ratio of the number of oscillating transitions over the total number of transitions made by the user. When the oscillating ratio is.5, the cumulative distribution of users is also.5, which means more than 5% of the transitions have oscillating phenomenon in half of the users. The high value of oscillating ratio means this is a pervasive phenomenon in MIT Reality dataset, hence it should be eliminated before modeling user mobility. In our experiment scenario, we classify the two APs together into a cluster when the frequency of transitions exceeds 4 round-trip during 8 minutes. C. Spatio-Temporal Prediction Evaluation Before evaluating the performance of our DelQue scheme, we firstly study the predict accuracy of the spatio-temporal prediction of user mobility, which also evaluates to what extent semi-markov process is appropriate to model user mobility in DTNs. We have carried out experiments using the MIT Reality dataset with larger network scale, and compare 1-community prediction with 2-communities prediction. In 1-community prediction, the geo-community with the highest probability among the corresponding ϕ ij (k) of Eq. (9) is chosen as the future location at the next prediction time. In 2-communities prediction, the geo-community with the next highest probability is also chosen as the second potential location at the next prediction time. As mentioned in Sec. IV, the spatio-temporal prediction of user mobility based on semi-markov model has to specify the time intervals at which predictions are made. The parameters for the devised prediction algorithm are as follows 1) h: the discretization step in time; 2) T p : the period at which predictions are made. In general, higher prediction accuracy will be achieved with a smaller discretization step h, because a smaller discretization step implies more refined characterization of the distribution functions. Due to limited space, we leave comparison of prediction accuracy under various combinations of parameters

16 16 Cumulative fraction of users Oscillating ratio Fig. 3. Cumulative distribution of users against their oscillating ratios. (For 97 active users during the period between July 26, 24 and May 5, 25.) for future work. Fig. 4 shows the cumulative fraction of users against prediction accuracy under the combination of h = 18 and T p = 18, where the fraction of users whose prediction accuracy lower than 5% is less than 1% in 2-communities prediction and 5% in 1-community prediction. Overall, the prediction accuracy for 1-community prediction with the highest probability is reasonable (mean =.49), and 2-communities prediction (mean =.63) has slightly better performance than 1-community prediction. D. Performance Evaluation of DelQue Recent user-initializing information query schemes always consider query and response separately. They employ well-known DTN routing schemes in query step and let the response be routed back along the query traces to save forwarding cost. In this section, we compare the performance of DelQue with such query framework with the flooding-based approach Epidemic routing [23], the mobility-based approach PROPHET [24], and the social-based approach BUBBLE Rap [9]. Since PROPHET and BUBBLE Rap are not designed for information query scheme, they are hard to apply to QMS scenario. Therefore, we compare the performance of the four algorithms only in QSS scenario, and then experimentally explore the performance comparison between QMS and QSS. Epidemic routing relies on flooding the network with information query. In PROPHET, the delivery predictability to destination geo-community is calculated at each relay candidate by using encounter

17 17 history, then query is forwarded in a greedy fashion to maximize such probability. In Bubble Rap, each query is forwarded to the destination geo-community greedily considering user centrality in a hierarchical manner based on social community knowledge. Cumulative fraction of users community prediction 2 communities prediction Accuracy Fig. 4. Cumulative fraction of users against the prediction accuracy with h = 18 and T p = 18 In our simulation, we focus on three metrics stated in Sec. III: query ratio, actual delay and average cost, which are key characteristics in data dissemination and access schemes of DTNs. We generate queries according to a Poisson process. Each simulation is repeated 5 times with random query sources and destination geo-communities for statistical convergence. In the performance comparison, we choose p = 9% and λ =.1. As seen from Fig. 5(a), the delivery ratio is tightly related to the time constraint (i.e., deadline T ). As a matter of fact, the selected relays may not have the chance to fulfill the query tasks if the time constraint is short, due to the low contact rates among users in DTNs. Such ratio increases remarkably as the time constraint becomes longer, since the selected relays have more chances to query, and the actual delay increases accordingly. Since the required query ratio p cannot be achieved when the time constraint is shorter, it is possible that all the relay candidates with λ.1 together cannot satisfy the performance requirements in Eq. (3), then the source hence can only forward the query to all the qualified relays with best effort. However, DelQue scheme performs as excellent as Epidemic in query ratio under multiple time constraints, and outperforms PROPHET and BUBBLE Rap by more than 15%. DelQue has slightly worse performance

18 18 Query ratio(%) 1% 8% 6% 4% DelQue Epidemic Bubble Rap Prophet Actual delay(h) DelQue Epidemic Bubble Rap Prophet 2% Time constraint(h) (a) Time constraint(h) (b) Average cost DelQue Epidemic Bubble Rap Prophet Time constraint(h) (c) Fig. 5. Performance of DelQue compared with other search schemes: (a) Query ratio, (b) Actual delay, (c) Average cost on actual delay, since in which the relays take charge of both query and response. As seen from Fig. 5(b), the actual delay of DelQue is 5% longer than that of PROPHET and BUBBLE Rap, and 1% longer than Epidemic. Fig. 5(c) shows that our approach has much less cost than the other three well-known routing mechanisms. When the time constraint is 8 hours, the cost of DelQue is only around 5% of the cost of PROPHET and BUBBLE Rap, and 2% of the cost of Epidemic. Cost is a significant metric in DTNs, since the devices in which are always energy-aware, and the device will be useless if the battery has exhausted. We also summarize the average number of messages each user has carried when the time constraint is 8 hours in Fig. 6. Apparently, the cost of each user in DelQue is much smaller than other algorithms. As a consequence, DelQue can deal with much more query requests compared to other approaches with the same devices. From the experiment results, we can see that DelQue can reach the comparable query ratio and actual

19 19 delay with some well-known DTN routing mechanisms with much lower average cost. Therefore, we can argue that DelQue is an effective information access scheme. E. Impact of p and λ on DelQue Performance We have conducted experiments to investigate the effects of different p and λ on the query performance of DelQue scheme. p (as seen from Eq. (3)) is the query requirement, which is determined by the source. Upon the query requirement reaches p, the source stops choosing relays. λ is the lower-limit of the user s query requirement, i.e., the source will only choose the users whose query probabilities higher than λ. In Fig. 7, the effects of different query requirements p on DelQue are investigated when λ =.1. Obviously, increasing p leads to an improvement of the query ratio, as shown by Fig. 7(a). When p increases from.5 to.7 and.7 to.9, the query ratios both increases by 15%-2%. Accordingly, higher p requires the source to select more query relays. When p is low, increasing p leads to a slight increment of the average cost, as shown in Fig. 7(c). When p increases from.5 to.7, the average cost increases by 15%-25%. Such increment becomes larger when p increases from.7 to.9, since the source will choose much more query relays with his/her best-effort to accomplish the required high query ratio. However, the actual delay remains stable among the three values of p as seen from Fig. 7(b), it is possibly because the query relays who finally respond the interest information back to the source are always selected. The impact of different λ on DelQue are studied in Fig. 8 when p =.9. As seen from Fig. 8(a), the query ratio increases slowly with the decrease of λ. On the other hand, the average cost increases dramatically as λ decreases. Therefore, introducing λ makes DelQue perform more efficiently. Moreover, through tuning λ we can balance the query ratio and average cost of DelQue to match different application scenarios. As seen from Fig. 7(c) and Fig. 8(c), the cost increases dramatically as the time constraint changes from 1h to 2h, and remains stable afterwards. The fact is that the source continues to choose relays to accomplish the query ratio requirement with best effort before T = 2h. One might think the cost should decrease as the time constraint becomes larger, as the relays have more chance to query the interest information. However, a user s query probability for the source increases slightly as the time constraint

20 2 becomes longer, since a user in society doesn t move around his/her affiliated communities frequently. Hence, the cost remains stable at an extremely low value (less than 1). 5 DelQue 5 Epidemic 4 4 Cost 3 2 Cost Cost User id Bubble Rap User id Cost User id Prophet User id Fig. 6. Number of messages each user has carried when time constraint T = 8h F. Performance Comparison of QSS and QMS Since other well-known DTN routing schemes cannot be easily extended to information query for mobile source, we compare their performance with DelQue in QSS scenario in Sec. V-D. In this section, we compare the performance of QMS with QSS to study the flexibility and robustness of our DelQue scheme. In QMS scenario, source user is randomly selected as a real user in the Infocom 6 trace. Fig. 9 shows the performance comparison between QSS and QMS. As seen from Fig. 9(a) and Fig. 9(b), our DelQue scheme performs almost the same in QSS and QMS scenarios, since the query ratio requirement in them are both 9%. In terms of the average cost, QMS is higher than QSS by 1%, that s because a user in the

21 21 1% 8% p =.5 p =.7 p = p =.5 p =.7 p =.9 Query ratio(%) 6% 4% Actual delay(h) 3 2 2% Time constraint(h) (a) Query Ratio Time constraint(h) (b) Actual Delay 1 8 Average cost p =.5 p =.7 p = Time constraint(h) (c) Average Cost Fig. 7. Performance of DelQue on different p value dataset may always belong to and move among several geo-communities, then the users query probability for a mobile source decreases, and the source in QMS needs to select more relays to accomplish the same query ratio with QSS. VI. RELATED WORK Information searching and retrieval is an active research in the computer science community. There are a wide range of publication with specializations from algorithms and operations of large scale databases to search in various application scenarios, e.g., centralized search engines, decentralized search approaches in DTNs. A. Internet Search Engines The traditional approach to map search queries to resources through Internet is mainly based on centralized search engines (e.g., Google, MSN, and Yahoo), where the mapping is implemented by

22 22 1% 8% λ = λ =.1 λ = λ = λ =.1 λ =.2 Query ratio(%) 6% 4% Actual delay(h) 3 2 2% Time constraint(h) (a) Query Ratio Time constraint(h) (b) Actual Delay Average cost λ = λ =.1 λ = Time constraint(h) (c) Average Cost Fig. 8. Performance of DelQue on different λ value maintaining an index about the resources. The index contains keywords and maps them to sets of related resources. Centralized search engines enable one to access information by collecting and maintaining mappings of content to location. However, a hypothesis of such search engines is user s real-time Internet access authority, which can not be satisfied in people s daily life. Therefore, decentralized information search schemes emerge as the times require. B. Decentralized search schemes in unstructured ad-hoc networks Recently, decentralized search schemes in unstructured ad-hoc networks have also been studied. There are three categories of approaches for searching in such networks: flooding, random walk, and using probabilistic paths. Most flooding-based schemes use TTL (Time-to-live) to control the spread of the queries [25]. However, search-schemes based on random walks [26], [27] avoid the high cost of flooding, but cannot ensure the expected performance metric, e.g., query ratio and actual delay, due to the random-

23 23 1% 8% QSS QMS 5 4 QSS QMS Query ratio(%) 6% 4% Actual delay(h) 3 2 2% Time constraint(h) (a) Query Ratio Time constraint(h) (b) Actual Delay Average cost QSS 1 QMS Time constraint(h) (c) Average Cost Fig. 9. Performance comparison between QSS and QMS ness. Specifically, we summarize several search-approaches using probabilistic paths to inspire related research in DTNs: Hara considers the data replica allocation problem in ad hoc networks in order to improve data accessibility [28]. Three replica allocation methods have been proposed in [28], they are Static Access Frequency (SAF), Dynamic Access Frequency and Neighborhood (DAFN), and Dynamic Connectivity based Group (DCG). However, all these three methods assume users have global knowledge of access frequencies, which may not be applicable in distributed networks. In [29], Stillerman has proposed flooding queries using limited radius and also replicating objects to improve the information retrieval latency. However, no detailed algorithms have been provided on how to replicate the objects. BubbleStorm [3] is a probabilistic search strategy, which is based on a combination of replication and probabilistic distribution of queries. Moreover, Yang and Hurson [31] present probabilistic schemes for locating content in wireless

24 24 ad-hoc networks. Based on knowledge about the query history, they use heuristics and Bayesian probability calculations guide the dissemination of queries. The common disadvantage of these approaches is that they always place tight demands on connectedness of the users or even consider users in networks have fixed position. However, well-connectivity cannot be ensured in DTNs, then the above approaches may not be applicable in such opportunistic networks due to the high maintenance cost of the index. C. Information search schemes in DTNs There are a few studies on web search in DTNs relying on the aid of Internet gateways. In [32], a system called Thedu has been proposed to enable efficient web search from a city bus. Thedu can be interpreted as a web proxy and collect search queries from a mobile user at the time of disconnections. Ott et al. [33] has investigated web resource retrieval in mobile DTNs via Internet gateways using well-known DTN routing mechanisms. However, we restrict our considerations to searching the mobile intermittent-connectivity environment and do not discuss interaction with infrastructure networks. Some researches on information access scheme in DTNs employ pub-sub paradigm [1], [11], in which the publisher aims to publish the message to the interested subscribers, at the same time the subscriber also intends to retrieve the relevant message from the network. In some application scenarios such as distributed photo or music sharing, users need to be able to find content even when they do not know an unambiguous identifier. The pub-sub paradigm is appropriate to these applications in DTNs. The user-initializing information query schemes originates from Osmosis [34], a search mechanism for pocket switched networks (PSN), in which the queries are spread based on epidemic forwarding, and the results are routed back based on the traces that queries left while traveling between requesting node and the responding node. Most later work tries to avoid the high cost of epidemic routing at the same time to approach its query ratio. Adamic et al. [35] present search strategies that exploit power-law degree distributions. Their method passes a search request from one node to another, choosing the neighbor with the highest degree. After the node with the maximum degree has been reached, it will be avoided, thus the search descends in the degree sequence. In [36], Pitkanen et al. have shown how caching improves efficiency with shared interest on resources using well-known DTN routing mechanisms.

25 25 One might intuitively think that evolving efficient DTN routing mechanisms will achieve high performance for information query. Most existing DTN routing algorithms aim to find a middle ground between epidemic and wait-for-destination scheme by relying on information that can be learned during contacts. PROPHET [24] is a mobility-based routing scheme that calculates the delivery predictability at each user by using encounter history. Some DTN routing schemes take advantage of the social behaviors of mobile users. SimBet routing [7] uses egocentric betweenness as forwarding metric and only forwards data to nodes with higher ones. BUBBLE Rap [9] considers social community as well as node centrality knowledge. Wei et al. study multicast routing schemes from a social perspective in [15]. However, the above stated multi-hop routing schemes introduce much overhead for the network, because they evolve lots of users as relays. Our work differs from the above stated search schemes that DelQue is the first to consider query and response integratedly, i.e., the chosen relays take charge of both querying the relevant interest message and getting it back to the source. Therefore, DelQue can ensure the query performance with only local knowledge. At the same time, DelQue can lower the network cost remarkably thanks to its one-hopdelegation pattern, compared with other existing multi-hop information search schemes. Furthermore, DelQue is designed for information query, hence it is more flexible than other forwarding-mechanismbased query scemes, especially dealing with mobile-source scenarios. VII. CONCLUSION In this paper, we have proposed a socially-aware information search scheme in DTNs that mobile users use to initialize the interest-based query and then delegate neighbors to get the relevant information back. We refer to this scheme as DelQue (short for delegation query). Concretely, we consider two cases: the source is static (QSS) or moving after delegation (QMS). As a matter of fact, some existing researches also consider information search application scenarios. However, to the best of our knowledge, none of the (few) works consider the steps of query and response integratedly. DelQue is in a one-hop and closedloop forms, i.e., relays take charge of the tasks of both querying and response. Since DelQue does not require the selected neighbors to forward the query to other users, the network overhead can be lowered dramatically and source can ensure system performance with only local knowledge. We also employ a

26 26 semi-markov model to compute the utility values, which are linked to their ability to fulfill the whole task. Semi-Markov model does not require the storage of the entire past history of the system and is computationally lightweight, making it suitable for a resource-scarce mobile device like PDAs. Extensive realistic trace-driven simulation results show that DelQue can reach almost the same query ratio and actual delay with flooding-based Epidemic schemes using less than 2% of its costs (in terms of the number of relays used). While the investigations in this paper provide a starting point in delegation query or even information search schemes in DTNs, numerous directions become apparent which deserve further attention: We currently consider the source users only request single interest-based message, a user however may aim to request multiple interest-based messages. In this work, we also assume that query items have a fixed expiration time. In real life scenarios, the expiration times of the query items may depend on the application scenarios. In addition, data centric security design needs to be done to ensure that only legitimate users are allowed to query and respond data in an information search system operating in challenging network environments. Last but not least, we intend to build a medium size testbed that demonstrates our DelQue scheme. REFERENCES [1] A. Balasubramanian, B. Levine, and A. Lindgren. Dtn routing as a resource allocation problem. In Proc. ACM SIGCOMM, Kyoto, Japan, Aug. 27. [2] S. Ioannidis, A. Chaintreau, and L. Massoulie. Optimal and scalable distribution of content updates over a mobile social network. In Proc. IEEE INFOCOM, Rio de Janeiro, Brazil, Apr. 29. [3] K. Fall. A delay-tolerant network architecture for challenged internets. In Proc. ACM SIGCOMM, Karlsruhe, Germany, Aug. 23. [4] S. He, J. Chen, Y. Sun, D. K. Y. Yau, and N. K. Yip. On optimal information capture by energy-constrained mobile sensors. IEEE Trans. Veh. Technol., 59(5), Jun. 21. [5] J. Zhao and G. Cao. Vadd: vehicle-assisted data delivery in vehicular ad hoc networks. IEEE Trans. Veh. Technol., 57(3), May 28. [6] Q. Li, S. Zhu, and G. Cao. Routing in socially selfish delay tolerant networks. In Proc. IEEE INFOCOM, San Diego, California, USA, Mar. 21. [7] E. Daly and M. Haahr. Social network analysis for routing in disconnected delay-tolerant manets. In Proc. ACM MobiHoc, Montreal, Quebec, Canada, Sep. 27. [8] W. Gao and G. Cao. On exploiting transient contact patterns for data forwarding in delay tolerant networks. In Proc. IEEE ICNP, Kyoto, Japan, Oct. 21. [9] P. Hui, J. Crowcroft, and E. Yoneki. Bubble rap: social-based forwarding in delay tolerant networks. In Proc. ACM MobiHoc, Hong Kong SAR, China, May 28. [1] E. Yoneki, P. Hui, S. Chan, and J. Crowcroft. A socio-aware overlay for publish/subscribe communication in delay tolerant networks. In Proc. ACM MSWiM, Chania, Crete Island, Greece, Oct. 27. [11] P. Costa, C. Mascolo, M. Musolesi, and G. P. Picco. Socially-aware routing for publish-subscribe in delay-tolerant mobile ad hoc networks. IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, 26(5):748 76, May 28. [12] M. J. Pitkanen, T. Karkkainen, J. Greifenberg, and J. Ott. Searching for content in mobile dtns. In Proc. IEEE PerCom, Galveston, Texas, USA, Mar. 29. [13] P. Yang and M. Chuah. Performance evaluations of data-centric information retrieval schemes for dtns. Computer Networks, 53(4): , 29.