A Content-Centric Framework for Effective DataDissemination in Opportunistic Networks

ABSTRACTIn this paper, we address the challenges of content transferin opportunistic networks, and propose techniques to betterfacilitate content dissemination based on the characteristicsof the content. To investigate this problem from its origins,we propose three message scheduling algorithms (HEC-FI,HEC-SF, and HEC-BI). Each algorithm is embedded in aspecially tailored content dissemination technique to evalu-ate the benefits of applying it to different types of contentand content dissemination methods. Three types of content(file, video and web) are considered and evaluated, and thedissemination methods considered are Layered MDC-basedand file-based. Using simulations as well as realistic net-work scenarios, we evaluate the proposed schemes in termsof latency and user perceived quality, and demonstrate howthe schemes can achieve much better latency performancefor file transfers. Furthermore, we show that using LayeredMDC-based techniques leads to higher user perceived qual-ity, since the end user is allowed to “preview” video fileor web content, even before the data has been completelytransferred. The effectiveness and robustness of our mes-sage scheduling algorithms and their corresponding contentdissemination techniques make them ideal solutions that cango a long way toward effective data dissemination in oppor-tunistic networks.

1. INTRODUCTIONWith wireless networking technologies extending into ev-

ery part of our working and living environments, proper han-dling of intermittent wireless connectivity and network dis-ruptions is important. As the foreseeable need for data com-munication escalates in challenged network environments,an increasing amount of effort is being invested in devel-oping techniques that can address these anticipated require-ments. Applications of these techniques are wide ranging.For instance, it would be quite advantageous to be able to in-terconnect mobile search and rescue nodes in disaster areas(where communication infrastructures have been disabled byearthquakes, hurricanes, wildfires, or flooding), allow mes-sage exchanges in underdeveloped areas (remote towns andvillages interconnected by wireless networks, but not guar-anteed an always-on Internet connection), and permit scien-

tific monitoring of wilderness areas (remote monitoring ofvarious forms of wildlife).

Formally, an opportunistic network is a type of challengednetwork that satisfies the following conditions: (1) networkcontacts (i.e., communication opportunities) are intermittent,(2) an end-to-end path between the source and the destina-tion has never existed, (3) disconnection and reconnectionis common, and (4) link performance is highly variable orextreme. Because of various disruptions and long delays,traditional MANET and Internet routing techniques can notbe applied directly to opportunistic networks. Meanwhile,with the development of numerous opportunistic network-ing applications, such as wireless sensor networks (WSN)[6] [20] [40], underwater sensor networks (UWSN) [15],pocket switched networks (PSN) [10] [11] [21], people net-works [36] [38], and transportation networks [3] [8] [23], itis both desirable and necessary to develop an effective con-tent dissemination framework that can better accommodatedata content and the various characteristics of opportunisticnetworks.

Several data forwarding schemes have been proposed foropportunistic networks [8] [18] [24] [25] [28] [37] [39] [42].Such schemes can be divided into two main categories, replication-based and coding-based, according to their basic technicalstrategies. In general, coding-based schemes are more ro-bust than replication-based schemes when the network con-nectivity is extremely poor, but they result in inefficiencywhen the network is fairly connected (due to the additionalcode blocks embedded in the data). Yet, all of the aboveschemes only consider data delivery performance (in termsof delivery latency and success ratio), and neglect the innatecharacteristics (e.g., content types) of applications when de-signing dissemination mechanisms. With the growing diver-sity of opportunistic network applications, a well-designedand content-centric solution for effective data disseminationremains highly desirable.

In this study, we propose techniques to better facilitatecontent dissemination based on the characteristics of the con-tent. To investigate this problem from its origins, we pro-pose three message scheduling algorithms: HEC-FI, HEC-SF, and HEC-BI. Each algorithm is embedded in a speciallytailored content dissemination technique to evaluate the ben-

1

efits of applying it to different types of content and contentdissemination methods. Three types of content (file, videoand web) are considered and evaluated, and the dissemina-tion methods considered are Layered MDC-based and file-based. Using simulations as well as realistic network sce-narios, we evaluate the proposed schemes in terms of la-tency and user perceived quality. We demonstrate how theschemes can achieve much better latency performance forfile transfers. Specifically, the results show that HEC-BIand HEC-SF achieve a good performance for networks withgood connectivity, while HEC-FI yields a more resilient per-formance in cases of poor network connectivity. Further-more, we show that our proposed techniques enable the enduser to “preview” video or web content at a lower qual-ity, even before the data has been completely transferred,thereby improving the overall viewing experience. Our eval-uations provide a deeper understanding of how to take ad-vantage of content types and message scheduling to moreeffectively disseminate data across opportunistic networks1.

The remainder of the paper is organized as follows. InSection 2, we review related works on opportunistic rout-ing, and define some commonly used terminologies. Section3 provides a comprehensive overview of the H-EC routingscheme and describes the proposed enhancements in detail.In Section 4, a Layered MDC scheme with an unequal era-sure protection scheme is presented, and the objective of thescheme is to facilitate more effective data transfer for videoand web content delivery over opportunistic networks. Sec-tion 5 presents a comprehensive set of simulation results forvarious opportunistic network scenarios; the results are alsoanalyzed and explained in detail. We then present our con-clusions in Section 6.

2. RELATED WORKRouting in an opportunistic network is challenging and

completely different to routing in a conventional network.An ideal routing scheme for opportunistic networks has toprovide reliable data delivery, even when the network con-nectivity is intermittent or when an end-to-end path is tem-porally unavailable. Moreover, since ‘contacts’ in an oppor-tunistic network may appear arbitrarily without prior infor-mation, neither scheduled optimal routing (e.g., linear pro-gramming routing in delay tolerant networks of scheduledcontacts [22]) nor mobile relay approaches (e.g., MessageFerrying [43] [44]) can be applied.

Currently, replication is the most popular design choicefor opportunistic routing schemes. For instance, the Epi-demic Routing scheme [37] sends identical copies of a mes-sage simultaneously over multiple paths to mitigate the ef-fects of a single path failure; thus, it increases the possibilityof successful message delivery. However, flooding a net-1Note that, unlike previous studies on opportunistic routing, wefocus on data transfer scenarios that assume the source has a com-plete set of messages before initiating data forwarding (rather thangenerating messages on the fly, as in [12] [39]).

work with duplicate data tends to be very costly in terms oftraffic overhead and energy consumption.

To address the problem of excess traffic overhead causedby flooding, Harras et al. proposed a Controlled Floodingscheme to reduce the flooding cost while maintaining re-liable message delivery [18]. In this scheme, flooding iscontrolled by three parameters, namely, willingness prob-ability, Time-to-Live, and Kill Time. Additionally, once amessage has been delivered to the receiver successfully, aPassive Cure is generated to “heal” the nodes in the networkthat have been “infected” by the message. Therefore, by re-moving the excess traffic overhead problem, while providingreliable data delivery, controlled flooding can substantiallyreduce the network overhead.

Node mobility also impacts on the effectiveness of oppor-tunistic routing schemes. Previous studies have shown thatif the network mobility departs from the well-known randomway-point mobility model (e.g., the Pursue Mobility Model[9] or the Reference Point Group Mobility Model [19]), theoverhead carried by epidemic- and/or flooding-based routingschemes can be further reduced by considering node mobil-ity. For instance, the Probabilistic Routing scheme [27] cal-culates the delivery predictability from a node to a particulardestination node based on the observed contact history, andforwards a message to its neighboring node if and only if thatneighboring node has a higher delivery predictability value.The scheme was revised by Leguay et al. [24] by taking themobility pattern into account, i.e., a message is forwarded toa neighbor node if and only if that node has a mobility pat-tern more similar to the destination node. [24][25] show thatthe revised mobility pattern scheme is more effective thanprevious schemes.

Another class of opportunistic network routing schemes isbased on encoding techniques, which transform a messageinto a different format prior to transmission. For instance,an integration of network coding and epidemic routing tech-niques has been proposed to reduce the required number oftransmissions in a network [42], and [39] proposes combin-ing erasure coding and the simple replication-based routingmethod to improve data delivery for the worst delay perfor-mance cases in opportunistic networks.

Although network coding-based routing schemes can re-duce the number of transmissions, and thereby improve therouting efficiency in a network, they may fail to provide ef-fective data delivery when the delivery latency is dominatedby some extremely large inter-contact time (i.e., the time be-tween two communication opportunities). In such extremecases, forwarding schemes based on erasure coding wouldbe more efficient, since the destination is able to reconstructthe message by just receiving a certain number of erasurecoded blocks, instead of all the transmitted data.

Following the concept of erasure coding-based data for-warding [39], an Estimation based Erasure-Coding routingscheme (EBEC) has been proposed to adapt the delivery oferasure coded blocks using the Average Contact Frequency

2

(ACF) estimate [26]. Moreover, [12] proposes a hybrid schemethat combines the strength of erasure coding and the advan-tages of Aggressive Forwarding, so that it not only remainsrobust in worst delay performance cases, but also performsefficiently in very small delay performance cases.

3. H-EC: AN ERASURE CODING-BASEDHYBRID ROUTING APPROACH

3.1 H-EC OverviewIn this subsection, we present an overview of erasure codes

and a forwarding scheme based on the erasure code proposedin [39].

Erasure coding is a scheme that provides better fault-toleranceby adding redundancy without the overhead of strict replica-tion of the original data [41]. Two of the most popular era-sure coding algorithms are Reed-Solomon coding and Low-Density Parity-Check (LDPC) based coding (e.g., Gallagercodes, Tornado codes, and IRA codes) [29] [35]. Thesealgorithms differ in the degree of encoding/decoding effi-ciency, the replication factor, and the minimum number ofcode blocks needed to reconstruct a message. The selectionof an appropriate erasure coding algorithm is not within thescope of this paper, since our work is based on the conceptof generic erasure coding.

In a generic erasure coding scheme, given a message ofsize M bytes, a replication factor of erasure coding r, anda coded message fragmented into several blocks of identicalsize b bytes, one can obtain the number of coded blocks byN = M×r

b . Moreover, this message can be successfully re-constructed as long as 1

r of the coded blocks are received,i.e., the minimal number of coded blocks required to suc-cessfully reconstruct the message is N/r.

In [39], an erasure code-based forwarding algorithm (EC)is proposed, as illustrated in Fig. 1. In this scheme, the era-sure coded blocks are split equally among n relays, whichare only allowed to send messages to the destination directly(this is the well-known “two-hop” scenario used in [10] [17]).Each relay forwards the same number of coded blocks (thereare no duplicates in a relay). The number of blocks for-warded by each relay can be obtained by2

N

n=

Mr

bn(1)

As reported in [39], the EC scheme can provide the bestworst-case delay performance with a fixed amount of over-head. However, the drawback of the EC scheme is that it cannot provide good very small delay performance compared toother popular replication-based approaches. The reason forthis inefficiency lies in its block allocation method. In theEC scheme, the number of transmitting blocks in each con-tact is fixed (i.e., Mr

bn , in accordance with Eq. 1) regardless2For simplicity, following [12] [39], we assume that N=n for allcases.

Figure 1: Illustration of the erasure coding-based dataforwarding algorithm (EC). In this figure, one erasurecoded block (A) is split equally among four relays (n =4).

Figure 2: Illustration of the A-EC scheme, i.e., EC withaggressive forwarding. In this figure, four erasure codedblocks (A,B,C,D) are transmitted, and n = 4.

of the duration of each contact. As a result, the EC schemecan only utilize each network contact effectively when thecontact duration is slightly longer than the time required tosend the relayed data. If most network contacts are muchlonger than the required time, the EC scheme tends to wastethe residual contact period, which results in ineffectiveness,as illustrated in Fig. 1.

To resolve this problem, [12] proposed an enhanced schemecalled A-EC, i.e., EC with an aggressive forwarding feature,as shown in Fig. 2. In this scheme, the source sends asmany coded blocks as possible during each contact (totallyMrbn blocks, i.e., Mr

n bytes). It has been shown that the A-EC scheme is better able to utilize network contacts; thus,it can be expected to outperform the EC scheme for verysmall delay performance cases. However, for worst delayperformance cases, it has been shown that A-EC yields apoor delivery ratio and/or a very large delivery delay whenblack-holes3 are present in the network [12].

Taking advantage of the strengths of the EC and A-ECschemes to achieve better message delivery performance inboth worst delay performance and very small delay perfor-mance cases, Chen et al. proposed a hybrid scheme, calledH-EC, which is illustrated in Fig. 3.

In the H-EC scheme, two copies of EC blocks (constructedbased on the erasure coding and replication techniques de-scribed previously) are transmitted by the sender. The firstcopy of EC blocks is sent in a similar way to how the origi-nal EC scheme sends blocks (shown as white blocks in Fig.3), while the second copy is sent using aggressive forward-ing during the residual contact time after sending the first ECblock (shown as gray blocks in Fig. 3). For general oppor-tunistic network scenarios (i.e., without black-hole nodes),

3A node in a network is called a black-hole if it is either unreli-able (e.g., it has very limited battery power and/or buffer size) or ithardly moves towards the destination [12].

3

Figure 3: Illustration of the H-EC scheme. Underthe scheme, two copies of four erasure coded blocks(A,B,C,D) are transmitted: the first copy of EC blocks(the white blocks) is sent using the EC algorithm, andthe second copy (the gray blocks) is sent using the A-ECalgorithm in the residual contact time. Each coded blockis split into 4 equal-sized sub-blocks (n = 4). This is ac-tually the HEC-SF scheme, which we will elaborate on inthe next subsection.

the H-EC scheme utilizes each contact opportunity betterbecause of the aggressive forwarding feature; however, ifblack-hole nodes are present in the network, the scheme’sperformance is expected to be similar to that of the EC scheme,which achieves better forwarding in worst delay performancecases.

The performance of the H-EC scheme depends to a largeextent on the message scheduling algorithm used in the ag-gressive forwarding phase, which is not discussed in [12].However, we are interested in assessing the impact of dif-ferent message scheduling algorithms on the performanceof H-EC routing. To this end, we propose three messagescheduling algorithms in the following subsection.

3.2 Message Scheduling in H-ECWe propose three message scheduling algorithms, namely

Sequential Forwarding (SF), Full Interleaving (FI), and Block-based Interleaving (BI), for transmitting the second copy oferasure coded blocks (i.e., using aggressive forwarding) inH-EC scheme. For simplicity, we assume the data file tobe transferred in an opportunistic network is L messagesin size. After applying erasure coding (i.e., adding redun-dancy), the size of each message is N erasure code blocks.We denote the n-th block of the l-th message as Ml,n. More-over, we assume that each message can be reconstructedwhen at least B out of N blocks are successfully received bythe destination node. We now describe the three algorithmsin detail.

3.2.1 Sequential Forwarding (SF)In the Sequential Forwarding (SF), the second copy of era-

sure coded blocks is sent sequentially in accordance with theorder of the messages. The main advantages of this schemeare (a) it is intuitive and easy to implement, and (b) it re-quires the minimal amount of buffer on the sender’s side(i.e., it does not need to apply erasure coding to all messagesin advance). The SF algorithm is detailed in Alg. 1.

3.2.2 Full Interleaving (FI)Different to the Sequential Forwarding (SF) scheme, the

Algorithm 1 The H-EC algorithm with Sequential Forward-ing (HEC-SF). The initial values of (l, n) are (0, 0).

Function HEC-SF (l, n)if l = L and n = N then

return EndOfMessageelse if l = 0 and n = 0 then

l ← 1; n ← 1return Ml,n

else if 1 ≤ l ≤ L and 1 ≤ n < N thenn ← n + 1return Ml,n

else if 1 ≤ l < L and n = N thenl ← l + 1; n ← 1return Ml,n

elsereturn Error

end if

Algorithm 2 The H-EC algorithm with Full Interleaving(HEC-FI). The initial values of (l, n) are (0, 0).

Function HEC-FI (l, n)if l = L and n = N then

return EndOfMessageelse if l = 0 and n = 0 then

l ← 1; n ← 1return Ml,n

else if 1 ≤ l < L and 1 ≤ n ≤ N thenl ← l + 1return Ml,n

else if l = L and 1 ≤ n < N thenl ← 1; n ← n + 1return Ml,n

elsereturn Error

end if

proposed Full Interleaving (FI) algorithm interleaves the sec-ond copy of the erasure coded blocks for H-EC. More pre-cisely, while performing aggressive forwarding, the FI algo-rithm transmits the “first” coded block of all the messagesat the outset, then the second block of all the messages, andso forth. The advantages of this scheme are (a) it distributesthe blocks of each message in a more diverse manner, andis thus expected to be more resilient to black-hole scenarios;and (b) since a message can be reconstructed after receivingjust a portion of all the coded blocks, there should be lessoverall delivery latency than in the SF algorithm. The FIalgorithm is detailed in Alg. 2.

3.2.3 Block-based Interleaving (BI)The main drawback of the FI scheme is the very long re-

sponse time needed to reconstruct messages when L and/orB are large (i.e., the time between sending the first block andsuccessfully reconstructing the first message). More specif-

4

Algorithm 3 The H-EC algorithm with Block-based Inter-leaving (HEC-BI). The initial values of (l, n) are (0, 0).

Function HEC-BI (l, n)i ← Int(n/B); j ← n mod Bif l = L and n = N then

return EndOfMessageelse if l = 0 and n = 0 then

l ← 1; n ← 1return Ml,n

else if j! = 0 and n < N thenn ← n + 1return Ml,n

else if (j = 0) or (j 6= 0 and n = N ) thenif l = L then

l ← 1; n ← (i + 1)×B + 1else

l ← l + 1; n ← i×B + 1end ifreturn Ml,n

elsereturn Error

end if

ically, the response time of the FI scheme is much greaterthan the time required to forward the first L × (B − 1) + 1blocks4. A clever solution to this problem is to send Bblocks during each contact period and interleave the send-ing process among L messages, instead of just sending asingle block, as in the FI scheme. This is called the Block-based Interleaving (BI) scheme. Note that, the FI scheme isa specialized case of the BI scheme with B equal to 1. Thealgorithm for the BI scheme is detailed in Alg. 3.

4. LAYERED MULTIPLE DESCRIPTIONCODING WITH UNEQUAL ERASUREPROTECTION

Layered MDC has been proposed as a means of com-bining Multiple Description Coding (MDC) [16] and Lay-ered Coding [30] for emerging multicast and peer-to-peeraudio/video streaming applications. More specifically, mul-tiple descriptions are spread across multiple packets (or paths)via MDC, and transmitted to a collection of clients, therebyreducing packet loss due to network congestion or the fail-ure of unreliable hosts. Applications of MDC include IP-level multicast [13] and application-level multicast [32] [33].Moreover, by using Layered Coding, multimedia data can beencoded into different quality levels, so that clients can playthe best possible video/audio quality level according to theircapabilities, such as screen resolution and link bandwidth.

By combining MDC and Layered Coding, the LayeredMDC scheme spreads the layered data across multiple pack-

4The response time would be even larger if we also consider dataloss, data disorder, and inter-contact time.

Figure 4: Illustration of the Layered MDC scheme withunequal erasure protection. Each video frame is encodedinto k quality levels using layered coding, and the i-thquality level video frame is erasure protected with repli-cation factor i and split equally among N relays (N ≥ k).

ets with multiple descriptions. Then, clients can play thelayered data as long as the required number of descriptionsare received successfully. Of course, the more descriptionsa client receives, the better the reconstructed data qualitywill be. In practice, the Layered MDC scheme is usuallyimplemented in conjunction with Unequal Erasure Protec-tion (UEP) [14], which provides different levels of erasureprotection to the Layered MDC blocks by adding differentamounts of redundancy (i.e., the more essential the codeblocks are, the more protection/redundancy is added). In thisstudy, we propose the use of Layered Multiple DescriptionCoding with unequal erasure protection for video file trans-fer and web surfing applications in opportunistic networks.Hereafter, we call the resulting scheme LMDC.

4.1 LMDC for Video TransferFig. 4 illustrates the LMDC scheme for video transfer

applications. From the figure, we observe that the quality ofa layered video frame improves as the size of the collectedvideo bit stream increases. More specifically, if one of thelayered video frames is Sk bits in size, one can split it into

5

k equal-sized pieces and reconstruct it to Qi quality level byusing any i out of the k pieces (i.e., the required bit streamsize for reconstructing Qi level frame is Si = i× Sk/k).

Each layered video frame is then split among N pack-ets (N ≥ k) with unequal erasure protection on each framepiece. For instance, the i-th piece of the layered video frameis erasure coded with replication factor r equal to N/i andspread among N packets (i.e., the i-th piece of the videoframe can be reconstructed by any i out of the N packets).The size of the i-th coded frame piece, bi, can be obtained byEq. 2, and the size of the resulting N packets, bpacket, canbe obtained by Eq. 3. Moreover, comparing with the Lay-ered Coding scheme, the traffic overhead of LMDC scheme,boverhead, can be obtained by Eq. 4.

bi = (Si − Si−1)× N − (i− 1)N

=Sk

k(1− i− 1

N) (2)

bpacket =k∑

i=1

Sk

k(1− i− 1

N) = Sk(1− k − 1

2N) (3)

boverhead = Nbpacket − Sk = (2N − k − 1

2)Sk (4)

Note that, since N ≥ k and k is a positive integer (i.e.,k ≥ 1), one can conclude that (a) boverhead = 0 when N =k = 1 (i.e., no LMDC); and (b) boverhead > 0 otherwise.Fig. 5 shows an example of an LMDC video frame at fourquality levels, i.e., Q1, Q3, Q6, and Q10, (k=10 and N=10).From this example, it is clear that the perceived video framequality improves significantly as the quality level increases.

(a) Quality Level 1 (b) Quality Level 3

(c) Quality Level 6 (d) Quality Level 10

Figure 5: Various qualities of a layered video frame(k=10).

4.2 LMDC for Web SurfingIn addition to video file transfer applications, we apply

the LMDC scheme to web surfing applications in oppor-tunistic networks. More precisely, the LMDC scheme isapplied to each MHTML document [34] that is a MIME(Multipurpose Internet Mail Extensions [7]) HTML docu-ment enclosing one or more objects, such as text, images,and videos. Unlike video transfer applications, the layeredcoding scheme encodes MHTML documents by looking upa pre-determined codebook, rather than splitting messagesinto equal-sized pieces. For instance, we present a codebookthat is based on the MIME type of each web object, as shownin Table 1.

Table 1: Layered coding for web objectsLayers MIME types

1 text/{html,css,plain,javascript,xml}; application/javascript2 Layer 1 + image/{gif,jpeg,png,bmp,x-icon}3 Layer 2 + application/{pdf,octet-stream,x-shockwave-flash}4 Layer 3 + video5 Layer 4 + others

In the codebook, Layer 1 web objects can be HTML doc-uments (source codes only), cascading style sheets (CSS),java script, or plain text. Layer 2 contains images files inaddition to Layer 1 objects. Layer 3 contains additional ap-plication objects (such as PDF files and flash games). Layer4 contains additional video objects. Layer 5 contains otherobjects (e.g., audio and unknown objects). Note that the de-sign of the codebook can be customized according to an ob-ject’s size, semantics, importance, and other design choices.The approach proposed in this paper is based on the genericLMDC concept.

Similar to video transfer applications, the LMDC schemefirst encodes each MHTML document into k quality levelsusing layered coding. The layered document is then splitamong N packets (N ≥ k) with unequal erasure protectionon each layered piece. More precisely, the i-th layered piececontains the i-th layered document, excluding the (i − 1)thlayered piece. Unequal erasure protection is applied to eachlayered piece, such that the i-th piece is erasure coded with areplication factor equal to i and split among N packets (i.e.,the i-th layered document can be successfully reconstructedfrom any i of the N packets). Note that, the i-th layeredpiece may be null if the original MHTML document doesnot have the corresponding types of objects as specified inthe codebook.

For example, using the codebook shown in Table 1, eachMHTML document is encoded into 5 layers using LMDC(i.e., k = 5). Fig. 6 shows the quality of sample web doc-uments in Layer 1, 2, and 3. As shown in the figure, Layer1 provides basic text-only descriptions at the lowest qualitylevel, Layer 2 adds images of intermediate quality to the de-scriptions, and Layer 3 adds flash animations of the highest

6

(a) Layer 1 (b) Layer 2 (c) Layer 3

Figure 6: Examples of the MLB.com web page after layered coding (Layer 1: the web document contains text only;Layer 2: the web document contains images in addition to Layer 1 objects; Layer 3: the web document contains flashanimations in addition to Layer 2 objects).

quality. The quality of the web document improves incre-mentally as the layer number or the amount of received dataincreases.

5. EVALUATIONIn this section, we evaluate the delay performance under

file transfer workload in opportunistic networks. We imple-mented the EC, HEC-SF, HEC-FI, and HEC-BI schemes andperformed simulations in DTNSIM [2], a java based DTNsimulator. For video transfer applications, we applied Lay-ered Coding to a 2000-frame video clip using JPEG2000 [4]codec, and added unequal erasure protection to each videolayer so that the video was LMDC encoded. For simplicity,the number of video quality levels, k, and the resulting pack-ets for each frame, N , were both set to 10 in the simulation.

In addition, for web surfing applications, we assume thereexists a gateway (say, Performance Enhancing Proxy, PEP)bridging the Internet and the opportunistic network (i.e., sim-ilar to the second HTTP scenario in [31]). The gatewayarchives web documents (in MHTML format) and deliversthem to the opportunistic network after LMDC coding. Tobe realistic, for the simulation, we selected a set of web doc-uments comprised of the top 500 requested documents ac-cording to the hit-count statistics of our campus proxy serverfor the period Apr.’06 to Sept.’06, and applied LMDC to theselected documents in the simulation. Table 2 shows the em-bedded object properties of the selected web documents, andTable 3 shows the distribution of the selected documents thatachieved the best possible quality with limited layers (i.e., aweb document may achieve the best possible quality withfewer layers if it does not contain objects located in higherlayers). The average number of layers required for the bestpossible quality documents in this study was 2.56. More-over, we also assumed that data transmission is wireless ata fixed rate of 1Mbps. The simulation results presented inthis section were all obtained by taking the average perfor-

Table 2: Object properties of the selected web docu-ments.

Object types Request (%) Avg. Size (bytes)

Layer 1 27.81 9,082Layer 2 excluding Layer 1 objects 63.11 6,974Layer 3 excluding Layer 2 objects 8.17 66,725Layer 4 excluding Layer 3 objects 0.15 2,028,783Layer 5 excluding Layer 4 objects 0.76 182,289

Table 3: The distribution of the required layer numbers,which can result in full quality document, of the selectedweb documents.

Layer # Number of documents % of documents

Layer 1 0 0%Layer 2 240 48%Layer 3 240 48%Layer 4 20 4%Layer 5 0 0%

mance of 200 simulation runs. In each simulation run thesource and the destination pair was randomly selected fromall participating nodes.

We evaluate three network scenarios. One is generatedaccording to the power-law distribution by setting both inter-contact time and contact duration of the network with power-law distributed values of coefficient 0.6 (as reported in [10][21]). The scenario consists of 34 participating nodes. Theother two scenarios are based on realistic campus wirelessnetwork traces, i.e., the Dartmouth [1] and UCSD [5] traces,which are publicly available for research references. Table 4outlines the basic properties of the three network scenarios

7

Table 4: The properties of the network scenarios in thisstudy.

Trace Name Power-Law UCSD Dartmouth

Device N/A PDA WiFi Adapter

Network Type N/A WiFi WiFi

Duration (days) 16 77 1,177

Devices participating 34 273 5,148

Number of contacts 25,959 195,364 172,308,320

Avg # Contacts/pair/day 2.89205 0.06834 0.01105

examined5.More specifically, the UCSD trace is a client-based trace

that records the availability of WiFi-based access points (APs)for each participating portable device (e.g., PDAs and lap-tops) on the UCSD campus. The network trace covers a twoand half-month period, and there are 273 participating de-vices. Similar to [10] [11] [21], we assume that two partic-ipating devices in ad hoc mode encounter a communicationopportunity (i.e., a network contact) if and only if they areboth associated with the same AP during the same time pe-riod.

The Dartmouth trace is an interface-based trace that recordedthe APs that associated with a particular wireless interfaceduring a three-year (1,177 day) period. However, it shouldbe noted that, wireless interfaces can be used by differentdevices at different times, and each device may use multi-ple wireless interfaces. For simplicity, we assume that eachnetwork interface represents a single mobile user in the net-work. Moreover, like in the UCSD scenario, a network con-tact is encountered when two mobile users are associatedwith the same access point. Note that, although the Dart-mouth trace is longer than the UCSD trace and has moreparticipating mobile nodes, its network connectivity is actu-ally very poor. This is because network contacts (for eachsource-destination pair) occur much less frequently than inother scenarios (only one-sixth of the UCSD scenario and0.4% of the Power-Law scenario).

5.1 Evaluation: Data File TransferIn the first set of simulations, we evaluate the delay per-

formance of our proposed message scheduling algorithms(i.e., HEC-SF, HEC-FI, and HEC-BI) for data file transfer inopportunistic networks. Three network scenarios are exam-ined via simulations, and a huge data file (i.e., 100MBytes)is selected for the file transfer. For all schemes, the erasurecoding parameters (r, n) are set to (2, 16), which is consis-tent with the settings used in [12] [39]. The simulation isperformed 200 times for each scheme, and Fig. 7 depicts theaverage data latency distribution results in ComplementaryCDF (CCDF) curves.5In the Dartmouth network trace, there were 13,888 devices in thenetwork, but only 5,148 of them had contact with other devices.

(a) Power-Law Scenario

(b) UCSD Scenario

(c) Dartmouth Scenario

Figure 7: Distribution (CCDF) of average latency per-formance of the EC, HEC-SF, HEC-FI, and HEC-BIschemes (N = 16 and r = 2).

8

From Fig. 7, we observe that HEC-based schemes (i.e.,SF, BI, and FI) outperform the EC scheme in almost all testcases, which further affirms the results of the previous stud-ies [12]. Specifically, the three variants of H-EC performnearly identically in the Power-Law scenario. They can suc-cessfully transfer more than 96% of the data file in the sim-ulation, compared to 80% transferred by the EC scheme.

However, we note that the completion ratio degrades asthe network connectivity (i.e., the average number of net-work contacts per node pair, per day) decreases. For in-stance, while HEC-based schemes can achieve about 96%completion ratio in Power-Law scenario, they can only achieveabout 45% in the UCSD scenario and 15% in the Dartmouthscenario. Thus, it is quite difficult to transfer complete datafiles in an opportunistic network, or it requires a much longertransfer time than that employed in the simulation scenarios.Therefore, from an end user’s point of view, transferring datafiles in an opportunistic network may not be feasible, un-less the receiver has the ability to read partial informationfrom an incomplete file. This finding motivates us to inves-tigate transferring video files with the LMDC technique. Wepresent the evaluation of LMDC-based video file transfer inthe next subsection.

Additionally, the HEC-SF and HEC-BI schemes performsimilarly in all scenarios, and perform consistently better atthe beginning of the three examined scenarios. This is incontrast to the HEC-FI scheme, which only performs betterabove a certain latency value. These results seem to con-tradict our initial intuition that HEC-BI should achieve thebest performance at all times. The reason is that the HEC-SFand HEC-BI schemes achieve better contact efficiency, sincethey allow the receiver to reconstruct the original messageas long as a certain number of coded blocks are received.In contrast, the interleaving nature of the HEC-FI schemerequires the receiver to wait for at least a few contacts (de-pending on the erasure coding parameter, r) before it cancollect sufficient coded blocks to reconstruct the message.As a result, as shown in Fig. 7, HEC-SF and HEC-BI usuallyperform better or more aggressively in cases with small la-tency, which represents good network connectivity (i.e., theleft portion of the figure), and HEC-FI usually performs bet-ter or more resiliently in cases with the worst latency, whichrepresents poor network connectivity (i.e., the right portionin the figure.

5.2 Evaluation: Video File TransferIn the second set of evaluations, we investigate the per-

formance of video file transfer (in terms of Peak Signal-to-Noise Ratio, PSNR) in opportunistic networks with andwithout LMDC-based coding schemes. Note that the LMDCscheme (with unequal erasure protection) is similar to the ECscheme, except that LMDC employs multiple redundancylevels (i.e., the r parameter) for each individual video qual-ity layer, instead of using just one redundancy level for thewhole message. Moreover, we apply the concepts of HEC-

(a) Power-Law Scenario

(b) UCSD Scenario

Figure 8: Average quality of video transfer using directcontact and the LMDC-based schemes. (Both N and kparameters are set to 10)

SF/FI algorithms to the LMDC scheme by sending the sec-ond copy of the LMDC blocks during the remaining con-tact time as in the HEC-SF/FI scheme. We call the result-ing schemes LMDC-SF and LMDC-FI respectively. Notethat there is no LMDC-BI scheme in our evaluation, sinceN = k, i.e., the minimal number of blocks required to re-construct an original quality video is exactly the same as thenumber of blocks transferred over the network6.

We only evaluated the performance of video file transferin Power-Law and UCSD scenarios, since the network con-nectivity of Dartmouth scenario is very sporadic and the datadelivery is very poor, as discussed in the previous subsection.For the Power-Law and UCSD scenarios, we took the aver-age performance results of 200 runs and, in each run, we ran-domly selected one node as the video source and one node asthe video destination. The video file was also 2000-framesin length, and all simulation parameters (i.e., transmission6When N > k, one can derive the LMDC-BI scheme by simplyapplying the HEC-BI algorithm with the number of blocks set equalto k.

9

(a) 500,000 seconds

(b) 1,000,000 seconds

(c) 5,000,000 seconds

Figure 9: Comparison of average video quality (i.e.,PSNR for each video frame) for the UCSD scenario af-ter 500,000, 1,000,000, and 5,000,000 seconds.

rate and erasure coding parameters) were the same as theprevious experiments. Fig. 8 shows the average PSNR per-formance of the 2000-frame video file using different codingand forwarding schemes.

Fig. 8 shows that, as expected, the H-EC based schemes

(i.e., LMDC-SF and LMDC-FI) outperform the LMDC schemeand the dir scheme, which employs the direct contact algo-rithm [39] to transfer video file directly without LMDC cod-ing. More precisely, the LMDC-FI scheme would performscomparatively better than the LMDC-SF scheme in the sim-ulation. This is because the FI-based strategy combines bothaggressive forwarding and interleaving techniques that notonly make aggressive use of precious network contacts, butalso do their best to alleviate the negative effects of possibleblack-holes.

The performance of the LMDC scheme is only slightlybetter than that of the dir scheme. This result contradictsour intuition about the LMDC scheme’s added resilienceto error-prone and/or poorly-connected networks [14]. Thesimilarity in performance is due to the amount of overheadcarried by LMDC (i.e., layered coding and erasure protec-tion). The overhead is too high to achieve a performancegain in extremely challenging network scenarios.

Apart from looking at the overall average PSNR of thevideo, it is also important to consider the frame-by-framePSNR performance of video file transfers, since the variancein the frame-by-frame PSNR can also have a strong impacton the perceived video quality for the end user. Fig. 9 showsthe frame-by-frame PSNR performance of three time points(500,000, 1,000,000, and 5,000,000 seconds) in the UCSDscenario.

From Fig. 9, it is clear that the frame-by-frame PSNRquality consistently improves as each encoding/forwardingscheme is given more time (also evident in Fig. 8). More-over, we note that the LMDC-FI scheme consistently outper-forms the other schemes for all three time points; whereasthe LMDC-SF scheme performs similarly to the LMDC andthe dir schemes in the beginning, but clearly outperformsthose two schemes after 1,000,000 seconds. The results con-firm that the aggressive forwarding phase can significantlyenhance the performance of data forwarding in opportunis-tic networks. They also demonstrate that the interleavingtechnique is useful for spreading LMDC blocks over the net-work, which alleviates the influence of potential black-holes.

Another noteworthy point evident in Fig. 9 is that theaverage PSNR value of all LMDC-based schemes degradesslightly as the number of frames increases. This is becausethe schemes all basically send video frames (or coded blocks)in sequence, regardless of whether it is in the first regularEC sending phase or in the second aggressive forwardingphase. This problem can be easily solved by sending videoframes in a uniformly random order; however, the tradeoff inthis instance is the computation overhead and memory spaceoverhead increase. We will consider this issue in our futurework.

It should also be noted that, in Fig. 9, the PSNR valueof the dir scheme oscillates heavily in terms of frequencyand amplitude, whereas the curves of LMDC-based schemesare much smoother (especially the curve for the LMDC-FIscheme). This is because, in the dir scheme, each video

10

frame is either successfully received or completely lost; thereis no intermediate quality video that can be played by theend user. As a result, the per frame PSNR performancesof the dir scheme is subject to large variations. As notedin [45], such drastic PSNR variations are detrimental to theend users’ perceived video quality. Therefore, based on ourobservations above, LMDC-based schemes can indeed yieldhigher frame-by-frame PSNR performance for video trans-fer in opportunistic networks, as well as providing better per-ceived quality video to the end users.

5.3 Evaluation: Web SurfingHere, we evaluate the performance of web surfing appli-

cations in opportunistic networks with and without LMDC-based coding schemes. Similar to the video file transfercases discussed in the previous subsection, the LMDC-basedschemes (i.e., LMDC, LMDC-SF, and LMDC-FI) are in factequivalent to the EC-based schemes (i.e., EC, HEC-SF, andHEC-FI schemes), except that they provide different levelsof erasure protection to each layered piece of a web docu-

(a) Power-Law Scenario

(b) UCSD Scenario

Figure 10: Average quality of web surfing results usingthe dir and LMDC-based schemes. (N = k = 5, andthe results have been normalized to the average qualityof the employed web documents, i.e., 2.56)

ment. Moreover, there is no LMDC-BI scheme, since N =k = 5 in our evaluation (i.e., using the codebook shown inTable 1).

We evaluated the performance of web surfing in the Power-Law and UCSD scenarios, based on the average performanceresults of 200 runs. In each run, we randomly selected onenode as the Internet gateway (which also served as the LMDCencoder) and one node as the web surfer. In the simulation,we used 500 web documents based on the top-500 hit-countstatistics of our campus proxy server for the period Apr.’06to Sept.’06 (the basic properties of the selected web docu-ments are shown in Tables 2 and 3). Fig. 10 shows thenormalized average quality of web surfing using differentcoding and data forwarding schemes.

The results in Fig. 10 clearly show that, regardless of thecoding and forwarding schemes employed, the average surf-ing quality improves over time, and eventually converges af-ter a certain period. The reason for this is very straightfor-ward: over time, a web surfer has more chances of makingmore contacts in the network; thus, he/she is more likelyto receive his/her requested web documents. The resultsalso show that the LMDC-based schemes consistently out-perform the dir scheme in both scenarios; more specifically,the HEC-based schemes (i.e., LMDC-FI and LMDC-SF) per-form comparatively better than the non-HEC scheme (i.e.,LMDC) in all test cases. The reason, not surprisingly, isthe same as for video transfers, i.e., LMDC-based schemesspread web documents more widely over the network, andallow the receiver to preview part of a document, even be-fore it has been transferred completely. Moreover, sincethe HEC-based schemes allow more redundancy in data for-warding, the data forwarding performance should be bet-ter. Finally, it is worth mentioning that LMDC-FI slightlyoutperforms LMDC-SF in the simulations, which confirmsour previous findings in LMDC-based video transfer appli-cations.

In addition to evaluating the average web document qual-ity in web surfing, we compare the surfing quality of eachreceived web document, which is more representative of thesurfing experience of end users. Fig. 11 shows the CDF dis-tribution of the web document quality received by a surfer,based on different coding and data forwarding schemes, atthree time points (500000, 1000000, and 5000000 seconds)in the UCSD scenario7.

In Fig. 11, it is apparent that, in spite of the coding anddata forwarding schemes employed, the CDF curve falls overtime, i.e., the curve for 5,000,000 seconds is lower than thatfor 1,000,000 seconds, and the curve for 1,000,000 secondsis lower than that for 500,000 seconds in all schemes. Theresults are consistent with our intuition that the quality ofweb document transfer should improve as the overall de-livery latency increases. Moreover, for each of the threetime points, the curves of the LMDC-based schemes are

7The evaluation results of the Power-Law scenario are omitted dueto space limitations.

11

(a) dir scheme

(b) LMDC scheme

(c) LMDC-SF scheme

(d) LMDC-FI scheme

Figure 11: Comparison of received web document qual-ity distribution (i.e., CDF) after 500,000, 1,000,000, and5,000,000 seconds using different coding and data for-warding schemes in the UCSD scenario.

consistently lower than the one of dir scheme’s curve, andthe HEC-based schemes have lower curves than the LMDCscheme. The results again confirm our previous evaluationresults that HEC-based schemes outperform all the other schemesconsidered in this study.

Additionally, Fig. 11 shows that, when the dir scheme isemployed, about 70% of web documents are unobtainable(i.e., the quality level is equal to 0), even after 5,000,000seconds. In contrast, only about 35% of documents are un-obtainable at the same time point when the LMDC-SF orLMDC-FI schemes are used. Moreover, since the dir schemeonly provides “one-or-zero” delivery of web documents, itdoes not allow scalable web surfing when only partial datais received; thus, there are no received Layer 1 quality webdocuments in the simulation (using the same set of 500 webdocuments shown in Table 3). However, the LMDC-basedschemes allow web surfers to browse lower quality (Layer1 quality) web documents before they have been transferredcompleted.

To sum up, the evaluation results demonstrate that theLMDC scheme enhances web document transfer in oppor-tunistic networks by enabling surfers to preview lower qual-ity web documents, even though the data has only been par-tially transferred. Moreover, in conjunction with the H-ECdata forwarding strategy, the LMDC-SF/FI schemes are ableto better cope with link outages and provide scalable websurfing in challenging network environments.

6. CONCLUSIONThere is an urgent need for an effective data dissemina-

tion scheme for opportunistic networks, as communicationopportunities in such challenged networks are opportunis-tic in nature and thus very precious. In this paper, usingthree types of content as examples (file, video, and web doc-uments), we have presented three message scheduling algo-rithms that enhance the data delivery capabilities of the H-EC scheme for file transfers. In addition, we have proposeda content-centric framework that combines layered codingand multiple descriptions coding to facilitate data dissem-ination for video transfers and web surfing applications inopportunistic networks. The proposed schemes were eval-uated using simulations as well as realistic network traces.The results show that the schemes can achieve a much betterlatency performance for file transfers. Specifically, the re-sults indicate that HEC-BI and HEC-SF achieve a good per-formance for networks with good connectivity, while HEC-FI yields a more resilient performance in cases of poor net-work connectivity. Moreover, we have shown that our pro-posed techniques enable the end user to “preview” video orweb content at a lower quality, even before the data has beencompletely transferred, thereby improving the overall view-ing experience. The effectiveness and robustness of our mes-sage scheduling algorithms and their corresponding contentdissemination techniques make them ideal solutions that cango a long way toward effective data dissemination in oppor-

12

tunistic networks.

7. REFERENCES[1] Crawdad project. http://crawdad.cs.dartmouth.edu/.[2] Delay tolerant network simulator.

http://www.dtnrg.org/code/dtnsim.tgz.[3] The disruption tolerant networking project at umass.

http://prisms.cs.umass.edu/diesel/.[4] Jpeg 2000. http://www.jpeg.org/jpeg2000/.[5] Ucsd wireless topology discovery project.

http://sysnet.ucsd.edu/wtd/.[6] The zebranet wildlife tracker.

http://www.princeton.edu/ mrm/zebranet.html.[7] N. Borenstein and N. Freed. Mime (multipurpose

internet mail extensions) part one: Mechanisms forspecifying and describing the format of internetmessage bodies. Technical report, IETF RFC 1521,September 1993.

[8] J. Burgess, B. Gallagher, D. Jensen, and B. N. Levine.Maxprop: Routing for vehicle-baseddisruption-tolerant networking. In IEEE Infocom,2006.

[9] T. Camp, J. Boleng, and V. Davies. A survey ofmobility models for ad hoc network research. WirelessCommunication and Mobile Computing Journal,2(5):483–502, 2002.

[10] A. Chaintreau, P. Hui, J. Crowcroft, C. Diot, R. Gass,and J. Scott. Pocket switched networks: Real-worldmobility and its consequences for opportunisticforwarding. Technical Report UCAM-CL-TR-617,University of Cambridge, Computer Laboratory,February 2005.

[11] A. Chaintreau, P. Hui, J. Crowcroft, C. Diot, R. Gass,and J. Scott. Impact of human mobility on the designof opportunistic forwarding algorithms. In IEEEInfocom, 2006.

[12] L.-J. Chen, C.-H. Yu, T. Sun, Y.-C. Chen, and H. huaChu. A hybrid routing approach for opportunisticnetworks. In ACM SIGCOMM Workshop onChallenged Networks, 2006.

[13] P. A. Chou and K. Ramchandran. Clusteringsource/channel rate allocations for receiver-drivenmulticast under a limited number of streams. In IEEEICME, 2000.

[14] P. A. Chou, H. J. Wang, and V. N. Padmanabhan.Layered multiple description coding. In IEEE PacketVideo Workshop, 2003.

[15] J.-H. Cui, J. Kong, M. Gerla, and S. Zhou. Challenges:Building scalable mobile underwater wireless sensornetworks for aquatic applications. IEEE Network,Special Issue on Wireless Sensor Networking, May2006.

[16] V. K. Goyal. Multiple description coding:Compression meets the network. IEEE SignalProcessing Magazine, pages 74–93, September 2001.

[17] M. Grossglauser and D. Tse. Mobility increases thecapacity of ad-hoc wireless networks. In IEEEInfocom, 2001.

[18] K. A. Harras, K. C. Almeroth, and E. M.Belding-Royer. Delay tolerant mobile networks(dtmns): Controlled flooding in sparse mobilenetworks. In IFIP Networking, 2005.

[19] X. Hong, M. Gerla, R. Bagrodia, and G. Pei. A groupmobility model for ad hoc wireless networks. In ACMInternational Workshop on Modeling, Analysis andSimulation of Wireless and Mobile Systems, 1999.

[20] J.-H. Huang, S. Amjad, and S. Mishra. Cenwits: Asensorbased loosely coupled search and rescue systemusing witnesses. In ACM SenSys, 2005.

[21] P. Hui, A. Chaintreau, J. Scott, R. Gass, J. Crowcroft,and C. Diot. Pocket switched networks and humanmobility in conference environments. In ACMSIGCOMM Workshop on DTN, 2005.

[22] S. Jain, K. Fall, and R. Patra. Routing in a delaytolerant network. In ACM SIGCOMM, 2004.

[23] J. LeBrun, C.-N. Chuah, and D. Ghosal. Knowledgebased opportunistic forwarding in vehicular wirelessad hoc networks. In IEEE VTC Spring, 2005.

[24] J. Leguay, T. Friedman, and V. Conan. Dtn routing in amobility pattern space. In ACM SIGCOMM Workshopon Delay Tolerant Networks, 2005.

[25] J. Leguay, T. Friedman, and V. Conan. Evaluatingmobility pattern space routing for dtns. In IEEEInfocom, 2006.

[26] Y. Liao, K. Tan, Z. Zhang, and L. Gao. Estimationbased erasure-coding routing in delay tolerantnetworks. In International Wireless Communicationsand Mobile Computing Conference, 2006.

[27] A. Lindgren and A. Doria. Probabilistic routingprotocol for intermittently connected networks.Technical report, draft-lindgren-dtnrg-prophet-01.txt,IETF Internet draft, July 2005.

[28] A. Lindgren, A. Doria, and O. Schelen. Probabilisticrouting in intermittently connected networks. ACMSIGMOBILE Mobile Computing and CommunicationsReview, 7(3):19–20, July 2003.

[29] M. G. Luby, M. Mitzenmacher, M. A. Shokrollahi,and D. A. Spielman. Efficient erasure correctingcodes. IEEE Transactions on Information Theory,47(2):569–584, February 2001.

[30] S. R. McCanne. Scalable Compression andTransmission of Internet Multicast Video. PhD thesis,University of California, Berkeley, 1996.

[31] J. Ott and D. Kutscher. Bundling the web: Http overdtn. In WNEPT, 2006.

[32] V. N. Padmanabhan, H. J. Wang, and P. A. Chou.Resilient peer-to-peer streaming. In IEEE ICNP, 2003.

[33] V. N. Padmanabhan, H. J. Wang, P. A. Chou, andK. Sripanidkulchai. Distributing streaming mediacontent using cooperative networking. In IEEE

13

NOSSDAV, 2002.[34] J. Palme, A. Hopmann, and N. Shelness. Mime

encapsulation of aggregate documents, such as html(mhtml). Technical report, IETF RFC 2557, March1999.

[35] J. S. Plank and M. G. Thomason. A practical analysisof low-density parity-check erasure codes forwide-area storage applications. In The InternationalConference on Dependable Systems and Networks,2004.

[36] D. Snowdon, N. Glance, and J.-L. Meunier. Pollen:using people as a communication medium. ElsevierComputer Networks, 35(4):429–442, February 2001.

[37] A. Vahdat and D. Becker. Epidemic routing forpartially-connected ad hoc networks. Technical ReportCS-2000-06, Duke University, 2000.

[38] R. Y. Wang, S. Sobti, N. Garg, E. Ziskind, J. Lai, andA. Krishnamurthy. Turning the postal system into ageneric digital communication mechanism. In ACMSIGCOMM, 2004.

[39] Y. Wang, S. Jain, M. Martonosi, and K. Fall. Erasurecoding based routing for opportunistic networks. InACM SIGCOMM Workshop on Delay TolerantNetworks, 2005.

[40] Y. Wang and H. Wu. Dft-msn: The delay fault tolerantmobile sensor network for pervasive informationgathering. In IEEE Infocom, 2006.

[41] H. Weatherspoon and J. D. Kubiatowicz. Erasurecoding vs. replication: A quantitative comparison. InThe 1st International Workshop on Peer-to-PeerSystems, March 2002.

[42] J. Widmer and J.-Y. L. Boudec. Network coding forefficient communication in extreme networks. In ACMSIGCOMM Workshop on Delay Tolerant Networks,2005.

[43] W. Zhao, M. Ammar, and E. Zegura. A messageferrying approach for data delivery in sparse mobile adhoc networks. In ACM MobiHoc, 2004.

[44] W. Zhao, M. Ammar, and E. Zegura. Controlling themobility of multiple data transport ferries in adelay-tolerant network. In IEEE Infocom, 2005.

[45] M. Zink, O. Kunzel, J. Schmitt, and R. Steinmetz.Subjective impression of variations in layer encodedvideos. In IWQoS, 2003.

14