performance comparison of four unicast routing schemes in dtns

of 13

Performance Comparison of Four Unicast Routing Schemes in DTNs Mooi Choo Chuah*,† , Peng Yang, Yong Xi

Department of Computer Science and Engineering, Lehigh University, Bethlehem PA, USA 18015-3084

Summary

Disruption tolerant network(DTN)s have been proposed to provide data communications in challenging network scenarios where an instantaneous end-to-end path between a source and destination may not exist. In this paper, we focus on comparing the performance of four recently proposed unicast routing schemes, namely (i) Multihop Routing with Custody Transfer (MRCT), (ii) Two-Hop Relay with Erasure-Coding (THEC), (iii) MaxProp, and (iv) Prophet. We conduct performance studies for different DTN scenarios, e.g., DTNs with different node densities, DTNs with different mobility models, networks with different percentage of nodes supporting DTN functionality, etc. Our results indicate that (i) the custody transfer concept significantly improves the delivery ratio in a sparsely connected network for all four schemes, (ii) multihop routing approach achieves higher delivery ratio with lower message delivery latency than the two-hop routing approach but incurs higher transmission overhead, (iii) MRCT scheme provides the best delivery performance by achieving high delivery ratio and low average delay, (iv) both MRCT and THEC schemes perform better with Zebranet than with the random waypoint (RWP) mobility model but MaxProp and Prophet performs worse with Zebranet than with the RWP model, (v) among the three multihop schemes, the transmission overhead for MaxProp and Prophet is higher than that incurred using the MRCT scheme. All these four schemes do not perform well when the network is very sparse (when each node has fewer than 1.5 neighbors within its transmission range) and hence an adaptive scheme that allows nodes to act as message ferries when its local neighborhood is very sparse is needed for such scenarios.

KEY WORDS: disruption tolerant networks; routing schemes; performance comparison

*Correspondence to: Mooi Choo Chuah, Department of Computer Science and Engineering, 19 Memorial Drive West, Lehigh University Bethlehem, PA 18015 †E-mail: [email protected] This work is sponsored by Defense Advanced Research Projects Agency (DARPA) under contract W15P7T-06-C-P430

of 13

I. INTRODUCTION Packet-switched network communication has

been studied for decades. Most packet-switched applications are built on top of TCP/IP protocol suite whose service model rely on a few key assumptions e.g. an end-to-end path exists between a source and a destination pair, the end-to-end packet drop probability is small and the maximum round trip time between any node pair is not excessive. However, there are many scenarios in which an end-to-end connection is not guaranteed or even possible, and so an intermediary is needed, perhaps to translate between protocols or to provide temporary storage (e.g., in mail servers). In these cases, without such intermediaries, communication would fail. In other cases, communication may fail not because of a lack of instantaneous connection, but because the connection properties fall beyond the expected bounds (excessive round-trip-time or high packet loss probability).

Solutions have been proposed to deal with some specific situations, e.g., using link layer retransmissions to deal with high packet loss probability in wireless environments [1] or using performance enhancing proxies [8]. Such proxies typically modify the end-to-end data stream, to try to fool TCP/IP based end stations into operating more efficiently over paths with poor or unusual performance. However, these solutions still do not work in situations where there are no end-to-end paths. In contrast, DakNet [2] deploys physical transport devices, e.g., buses and motorcycles, to carry mobile access points between village kiosks and hubs with Internet connectivity so that the data carried by the physical transport devices can be automatically uploaded and/or downloaded when the physical transport devices are in the wireless communication range of a kiosk or a hub. Similar techniques are proposed in [3],[4]. In the past two years, a considerable amount of research focusing on delay/disruption-tolerant networking and communications has been published (e.g., [5],[6],[11],[13],[14],[15]). DieselNet [11],[14] is a disruption tolerant network where connections between nodes are short-lived and occasional. A common approach used to address delays and disruptions is via the use of a store-and-forward mechanism similar to electronic mail [10]. This makes communication possible, even when an instantaneous end-to-end path does not exist. Message ferrying schemes [12],[15] are proposed where special mobile nodes called message ferries are used to facilitate connectivity between nodes. The message ferries visit the nodes in the network and deliver data among them.

Figure 1: Proposed DTN architecture [5]

In [5], Fall describes an architecture for delay tolerant networking. It proposed the idea of topological regions connected by gateways, which were responsible for storing messages in non-volatile storage to provide for reliable delivery. End-point addressing in his scenario consisted of a region name used for inter-region routing and a locally-resolvable name for intra-region delivery. As shown in Figure 1, a sensor or adhoc network can each form a DTN region. DTN gateways interconnect regions running potentially dissimilar protocol stacks. In addition, a custody transfer feature is defined [13],[17] for DTN message forwarding. A node accepting a message with custody transfer means it promises not to delete it until it can reliably deliver the message to another node providing custody transfer or the message arrives at the destination. More recently, we have proposed an enhanced disruption-tolerant network architecture called EDIFY (Enhanced Disruption and Fault Tolerant Bundle Delivery) [6]. Our approach builds on many ideas from Fall [5], but adds support for multiple, overlapping name spaces and node and group mobility..

Several DTN unicast routing schemes have been proposed [19],[20],[21],[22],[23]. These DTN unicast routing schemes can be classified into three categories. In one category, special nodes referred to as message ferries or data mules are used to connect nodes that are far apart geographically. These message ferries visit the nodes using fixed or dynamic ferry routes. The second category of routing schemes keep track of contact histories and may use of such history information to derive deliverable probabilities. Each node will pass the message to another node that it comes across if the other node has a higher deliverable probability to the destination node. In the third category, the nodes use two-hop forwarding schemes [19]. The source node distribute K copies of a message to the first K contacts it comes across. The nodes that receive these messages will store these messages and deliver them only if they happen to come into

of 13

contact with the destination. Erasure coding approach may be used to improve the efficiency of such schemes e.g. a message may be coded into n data blocks and the destination only needs to receive k (where k<n) data blocks for message reconstruction.

In this paper, we focus on comparing the performance of four recently proposed DTN routing schemes, namely (a) multihop routing with custody transfer (MRCT) scheme [9], (b) two hop with erasure coding (THEC) scheme [19], (c) MaxProp [11], and (d) Prophet [22]. Specifically, our contributions in this paper are:

(a) we study the effectiveness of the custody transfer feature with several DTN routing protocols in mediumly sparse adhoc networks. It is important to know when the custody transfer feature needs to be turned on since the deployment of the custody transfer feature incurs extra overhead.

(b) we study the impact of mobility models on the performance of DTN routing protocols in mediumly sparse adhoc networks. We consider both the random waypoint model and the Zebranet-like model. Zebranet-like mobility model is a mobility model derived from the traces obtained from the sensor nodes attached to zebras in the ZebraNet project described in [19].

(c) we compare the performance of four DTN routing schemes, namely (i) the two-hop with erasure coding (THEC) scheme [19], (ii) the multihop routing with custody transfer (MRCT) scheme, (iii) MaxProp, and (iv) Prophet in mediumly sparse DTNs.

Our evaluations indicate that when the nodes move according to the random waypoint (RWP) mobility model, the MRCT scheme achieves the best delivery ratio and lowest average delay in mediumly sparse adhoc networks. This is followed by the Prophet and MaxProp. The two-hop scheme with erasure-coding performs the worst but has the smallest transmission overhead. The Prophet scheme performs better than the MaxProp scheme when the nodes move according to the RWP model but the two schemes have similar performance when the Zebranet mobility model is used. The THEC scheme is optimized for Zebranet mobility model and gives better data delivery performance than MaxProp and Prophet when Zebranet model is used. All schemes start to perform poorer when the network is really sparse e.g. with a node having fewer than one neighbor within its transmission

range. Under that circumstance, an adaptive routing scheme that allows a node to behave as a message ferry will perform better. We have presented such a scheme and evaluate its performance in [29].

The rest of our paper is organized as follows: we first give an overview of the three routing approaches that have been proposed for forwarding unicast messages in DTNs in Section II, and discuss their advantages and disadvantages. In Section III, we describe our simulation model. In Section IV, we study the impact of node density, mobility models, the percentage of DTN nodes on the data delivery performance of four DTN routing schemes. In addition, we also compare the performance of the four schemes. Our results indicate that the MRCT scheme can provide higher delivery ratio and lower packet delivery latency. We give some concluding remarks in Section V.

II. UNICAST ROUTING SCHEMES AND

CUSTODY TRANSFER FEATURE FOR DTNS

In this section, we first describe the various categories of the unicast routing schemes that have been proposed for DTNs. Then, we describe the custody transfer feature proposed for DTNs.

A. Unicast Routing Schemes

Three categories of forwarding schemes have been proposed for DTNs. The first category uses special nodes to connect partitioned networks. The second category uses multihop routing approach where contact history information to determine the next hop node to pass a message. The third category uses two-hop routing approach where the intermediate nodes that receive messages from any source has to store the messages until deliver them when they come into contact with the destination. We elaborate on each category in subsequent paragraphs.

In the first category [12],[15],[20],[28], special nodes called message ferries or data mules are used to collect data from stationary or moving sources and deliver them to their destinations. The message ferries move around the deployment area and are responsible for carrying data between nodes. The design of ferry routes will have significant impact on network performance. However, in these existing researches, regular nodes cannot become message ferries. Some preliminary work on turning moving nodes into message ferries in areas with sparse connectivity has been explored in [29].

of 13

In the second category [22], [24],[11], history-

based routing is proposed in which each node maintains a utility value for every other node in the network, based on a timer indicating the time elapsed since the two nodes last encountered each other. These utility values, which carry indirect information about relative node locations, get diffused through nodes’ mobility. Nodes forward message copies only to those nodes with a higher utility for the message’s destination. For example in [22], the authors propose a probabilistic metric called delivery predictability at every node A for each known destination B. This metric indicates how likely it is that node A will be able to deliver a message to that destination. The delivery predictability ages with time and also has a transitive property, i.e., a node A that encounters node B which encounters node C allows node A to update its delivery predictability to node C based on its (A’s) delivery predictability to node B and node B’s delivery predictability to node C. In [22], a node will forward a message to another node it encounters if that node has higher delivery predictability to the destination than itself. Such a scheme was shown to produce superior performance than epidemic routing [24]. The MaxProp approach in [11] is similar to [22], MaxProp differs from the approach in [22] only in how they maintain the contact probabilities. MaxProp renormalizes the contact probabilities a node maintains for all its neighbors but the probabilistic routing approach in [22] uses different equations to update the contact probability a node maintains for its neighbors. The performance of the history-based routing approach depends on how frequent the various history-based schemes perform neighbor discovery. Doing it more often improves the delivery ratio but incurs more control overhead.

In the third category [7],[19], [23], a two-hop relay forwarding scheme is proposed in which the source sends either single or multiple copies (e.g., different erasure coding blocks) to different relaying nodes and the relaying nodes deliver the copies they have to the destination node when they encounter the destination node. The Binary Spray and Wait scheme proposed in [7] assumes that each source node starts with L copies of the same message, and sends half of its remaining copies to a new contact until it has a single copy left. Then, the source node switches to direct transmission (which means it will only forward to the destination node). The authors in [19], [23] use a

similar approach except that they optimize transmission overhead by using erasure coding. Thus, in erasure-coding based two hop routing approach [19],[23], a message is coded and divided into n blocks and send to n contacts but the destination node only needs k (where k<n) copies to reconstruct the original message. With the same amount of overhead, the erasure-coding based approach allows the source to send copies of the same message to more contacts and hence increase the chances of having the destination node receives enough copies to reconstruct the original message. We anticipate that a two-hop routing strategy can achieve small transmission overhead but may not enjoy high delivery ratio especially if the messages have delivery deadlines.

B. Custody Transfer Feature As indicated earlier, custody transfer feature has

been proposed for DTNs [13],[17]. This custody transfer feature [13],[17] provides reliable communications in an intermittently connected network. When a node accepts a message with custody transfer, it means the node promise not to delete the message until it can be reliably delivered to another node providing custody transfer or the message arrives at the destination. Nodes holding a message with custody are called custodians. Normally, a message has a single custodian (referred to as sole custody) but in some circumstances, more than one custodian owns a message or message fragment (referred to as joint custody). Applications can optionally request the custody transfer feature on a per-message basis and they will receive a custody acknowledgement when their host system can find one or more nodes that are willing to take custody of the message. A node may agree to accept custody for messages initially and refuse to do so when its local node resources, e.g., buffers, become substantially consumed. Potential problems that may occur with custody transfer are discussed in [17]. One of the problems discussed in [17] is the head of line blocking problem. At a node R, messages destined to node A may be sitting behind a message destined to node B and there is no communication link between node R and node B but the communication link between node R and node A is available. Node R needs to be able to transmit those messages destined to node A.

III. SIMULATION MODEL

We use NS-2 simulator [27] for our extensive simulation studies. Below we describe the new modules that we add to the NS2-simulator.

of 13

A. Implementation of the four DTN Routing Schemes We are interested in understanding whether the

multihop routing approach can perform better than the two-hop routing approach that has been proposed for DTNs. Hence, we implement four DTN unicast routing schemes, namely (a) the multihop routing with custody transfer (MRCT) scheme, (b) MaxProp, a multihop approach proposed in [11], (c) Prophet, yet another multihop approach proposed in [22], and (d)the two-hop with erasure-coding (THEC) scheme [19].

For the MRCT scheme, we assume the custodian of each message will try to find a route to the destination node by issuing route request messages as in DSR [18]. The difference between the MRCT scheme and the DSR scheme is that the message custodians instead of just the source node will perform route discovery and route repair. In addition, the custody transfer feature is turned on by nodes that support DTN functionalities.

The two-hop with erasure coding (THEC) scheme [19] is implemented as follows: the nodes exchange hello messages with one another periodically. This allows them to build a neighbor table. Each message is encoded and divided into eight blocks at the source node. Upon meeting a new 1-hop contact, the source node sends a data block to that new neighbor. Thus, all eight blocks will be disseminated (to eight contacts). The contacts carry the block until they meet the destination. If that happens, a contact will deliver the message block. As long as the destination receives four out of the eight blocks, that message is considered successfully delivered to the destination.

In [11], Burgess et al propose MaxProp, a protocol for effective routing of DTN messages. MaxProp uses several mechanisms to define and prioritize the transmission schedules for the packets. A ranked list of the peer’s stored packets is dynamically built based on the delivery likelihood to the destinations of the packets. In MaxProp, the delivery likelihood is used as a cost metric, and this information is exchanged each time two peers meet. Packets with the highest delivery likelihood will be ranked with the highest priority. Packets that are ranked with the highest priority are the first to be transmitted during a transfer opportunity.

In [22], Lindgren etc designed probabilistic routing protocol using history of encounters and

transitivity (PROPHET) for intermittently connected networks. This probabilistic routing scheme establishes a probabilistic metric called delivery predictability at every node a for each known destination b. When two nodes meet, they exchange the delivery predictability information they stored. This information is used to update the estimated delivery predictability to the destination. A message is transferred to the other node if the delivery predictability of the destination of the message is higher at the other node.

B. Implementation of Custody Transfer feature We implement the custody transfer feature in

our simulator as follows: when a DTN node has a message to send for which it holds custodianship, it checks its cache to see if it has a route to the destination node. If it finds more than one routes, it picks the one with the lowest cost (e.g., using hop count, delivery latency etc., as metrics). When a route is selected, it checks the DTN nodes included in this selected route to see which node is the best candidate for custody transfer, e.g., the closest DTN node that has buffer space available. Then, it sends a custodian request to that downstream DTN node. If the DTN node can accept the custodianship, it will respond with a custody acknowledgement. Otherwise, it sends a negative reply.

If the sending DTN node cannot find a route to the destination of the message, it will send a custody request to its 1-hop DTN neighbors to see if any one of them has a route to this destination. If there is a custodian accept reply from any 1-hop DTN neighbors, then, this sending DTN node will send the bundle to that replying node. If there is no reply (after a wait-for-reply timer expires), then this sending DTN node will trigger its underlying ad hoc network layer to look for a route or neighboring nodes that are closer to the destination than itself. At the ad hoc network routing layer, all DTN nodes that receive a route reply message with the DTN option flag set will set a bit in the appropriate position (according to its hop distance from the sending node of the route request) to indicate buffer availability before relaying the route reply message. Thus, our dual-layer (at ad hoc network routing and DTN layers) approach allows a node to identify downstream nodes to which we can forward the messages. Once a custodian node is selected, the sender transmits a message to it and waits for an acknowledgement. If the sender does not receive custodian acknowledgement from the new custodian node, it will retransmit up to a certain maximum number of times. If the sender

of 13

still fails to receive acknowledgement after multiple attempts, the sender can select another node to be the custodian. Our custody transfer implementation avoids the head of line blocking problem described in [17] by allowing the DTN node to search through the queued messages until it finds a message that can be sent to the next hop node.

C. Mobility Models Since we intend to compare the impacts of

different mobility models on the delivery performance of various routing schemes, we implement a scaled Zebranet mobility model. As in [19], we create a semi-synthetic Zebranet mobility model as follows: we synthesize node speed and turn angle distributions from the observed data and create other node movements using the same distribution. We scale the grid size to 1500mx1500m and use a transmission range of 250 m (as compared to a grid size of 6Kmx6Km and a transmission range of 1000 m used in [19]). In the original Zebranet trace [19], the inter-sample interval is 8 minutes but we scale this interval to 2 minutes in the experiment we conducted. This allows us to keep the node speed to be similar to what is reported in the Zebranet trace. D. Traffic Models

For traffic models, we typically use UDP-based Constant Bit Rate traffic model. In addition, we also implement bidirectional flows. Bidirectional flows [26] are implemented as follows: a source sends a message to a destination and the destination will respond with a message of the same size back to the source before the source generates further messages.

IV. PERFORMANCE STUDY OF UNICAST

ROUTING SCHEMES IN DTNS

In this section, we perform extensive simulation studies to understand how different factors e.g. node densities, mobility model, the percentage of DTN nodes affect the performance of all four DTN routing schemes. The performance metrics we use are: (a) transmission overhead [19] which is defined as the number of transmitted bytes over the number of generated bytes. Note that in this case, the transmitted bytes include the routing overhead. Each routing message and each custody

transfer request/acknowledgement message is assumed to be 35 bytes long. (b) the average end-to-end delivery latency (denoted as Avg Delay in the tables or plots), and (c) the packet delivery ratio (PDR)

For each scenario, we conduct 10 simulation runs and report the average of the metrics obtained from these 10 runs. A warm up period of 1000 seconds is used for each simulation run and the simulation is run for 10,000 seconds. In our simulations, we use IEEE 802.11 radio with a transmission range of 250m and a peak data rate of 2 Mbps. Unless otherwise stated, we use a network scenario with 40 nodes distributed randomly in the following areas: (a) 1000x1000m2(4x10-5node/m2) , (b) 1500x1500 m2 (1.8x10-5 node/m2 ), (c) 2000x2000 m2(1.0x10-5 node/m2 ) , and (d) 3000x3000 m2 (4.4x10-6 node/m2 ). A. Impact of Node Densities

In this section, we investigate how the DTN node density impacts the delivery performance of all four schemes. The nodes are moving according to the random waypoint mobility model with a maximum speed of 5 m/s. First, we run some experiments assuming that the nodes do not support custody transfer, i.e., they are just regular adhoc network nodes. Then, we run the same experiments assuming all nodes turn on the custody transfer feature.

10 source/destination pairs are used in this set

of experiments. The source/destination pairs are randomly picked among the 40 nodes. Each source generates one packet every 4 seconds. Thus, the traffic model used is the Constant Bit Rate traffic model. The packet size is 512 bytes.

Figures 2,3, and 4 plot the delivery ratio, the average delay, and the transmission overhead for the case with and without custody transfer feature using either the MRCT or THEC schemes. Figures 5,6, and 7 plot the same metrics using either MaxProp or Prophet with and without custody transfer. Without custody transfer, the packets will be dropped if they are at any intermediate node for more than 60 seconds irrespective of what routing scheme is used. With custody transfer, the packets will be held at any intermediate node forever until there is a buffer overflow.

of 13

Figure 2: Delivery ratio with and without custody transfer using MRCT or THEC scheme.

Figure 3: Average Delay with and without custody transfer using MRCT or THEC scheme.

Figure 4: Transmission Overhead with and without custody transfer using MRCT or THEC scheme.

Figure 5: Delivery Ratio vs Node Density with and without custody transfer using MaxProp or Prophet.

Figure 6: Average delay vs Node Density with and without custody transfer using

MaxProp and Prophet

Figure 7: Transmission Overhead vs Node Density with and without custody transfer

using MaxProp and Prophet.

From Figure 2, we see that the packet delivery

ratio (PDR) using the MRCT without the custody transfer starts to drop significantly (by 19%) when the node density changes from 4x10-5 node/m2 to 1.8x10-5 node/m2. The drop in PDR is 30% as the node density decreases from 1.8x10-5 node/m2 to 1.0x10-5 node/m2 without the custody transfer feature. The custody transfer feature significantly improves the packet delivery ratio for those scenarios where the node density is below 1.8x10-5

node/m2. In the 1x10-5 node/m2 (40 nodes over 2000x2000m2) case, we see that the delivery ratio has dropped to 48% without custody transfer. However, the packet delivery ratio increases to 98.6% when the custody transfer feature is turned on. The additional price to pay for this improvement is an increase of transmission overhead by almost 217% for the case with a node density of 1x10-5 node/m2. We expect the delivery ratio to continue to drop significantly as node density continues to drop below 0.5x10-5 node/m2, and that the custody transfer feature alone will not be enough to allow the sparsely connected nodes to communicate with one another. For such cases, we propose to use message ferries to connect partitioned nodes [29]. Similar drop in the delivery

ratio with increasing network sparseness is seen when the THEC scheme is used. The delivery ratio performance using the THEC scheme is poorer than that for the MRCT scheme. In fact, the delivery ratio achieved using the THEC scheme with custody transfer is similar to the performance achieved by MRCT without the custody transfer. Figure 5 shows that both MaxProp and Prophet also benefit significantly from invoking the custody transfer feature. Comparing Figures 2 and 5, we see that with custody transfer, the multihop approach (MRCT, MaxProp and Prophet) achieves higher delivery ratio than the THEC approach. In addition, we see that the for the RWP model, the MRCT scheme achieves the highest delivery ratio until a node density of 0.5x10-5 when MaxProp and Prophet seems to perform slightly better. This may be due to the fact that the route discovery process used in the MRCT scheme does not perform well in very sparse networks.

From Figure 3, we see that the average delay achieved using the MRCT or THEC scheme with custody transfer is significantly higher than that achieved when custody transfer is not turned on. With custody transfer, extra retransmission delay and waiting time are incurred in delivering packets

of 13

that are thrown away in the scenario where the custody transfer is not turned on. The plot in Figure 3 also indicates that the average delay achieved using the MRCT scheme with custody transfer is significantly lower than that achieved using the THEC scheme with custody transfer. The price to pay for better delivery performance in the MRCT scheme is higher transmission overhead since the packets are delivered via multiple hops. Comparing Figures 3 & 6, we see that the average delay achieved using MRCT, MaxProp, and Prophet with custody transfer are very similar and much lower than that achieved using the THEC scheme.

Figure 4 indicates that the transmission

overhead increases with decreasing node density initially and then it drops. The initial increase in the transmission overhead is caused by the increased retransmission cost. However, when the network becomes very sparse, fewer packets are delivered and such packets are usually delivered using small number of hops, hence the transmission overhead decreases. Comparing Figure 4 & 7, we see that the transmission overhead with MaxProp and Prophet is significantly higher than that for MRCT and THEC schemes. Among the four schemes, the MRCT scheme achieves the best overall delivery performance. B. Impact of different mobility models

Next, we conduct experiments to understand the impact of mobility models on the message delivery performance. We use the same network scenario. The nodes move either according to random waypoint (RWP) model or according to Zebranet movement [19]. For the random waypoint movement, the nodes have a maximum speed of 5m/s. Again, ten CBR flows with the source and destination nodes randomly picked among the 40 nodes are used in this set of experiments. The packet generation rate is 0.25 packet/sec. Figures 8, 9, and 10 plot the delivery ratio, the average delay, and the transmission overhead for both mobility models using the MRCT and THEC schemes with custody transfer. The THEC scheme is marked as the two-hop approach while the MRCT scheme is marked as the multihop approach in the plots. Figures 11-13 show similar plots for the MaxProp and Prophet schemes with custody transfer.

From Figures 8 & 11, we see that with RWP mobility model, MRCT achieves the highest delivery ratio among the four schemes until the node density is 0.5x10-5 node/m2 , after which Prophet performs the best with MaxProp follows closely behind. THEC performs the worst using RWP mobility model. With Zebranet model, both MRCT and THEC scheme achieve similar delivery ratios (>90% even at the node density of 0.5x10-5) while both Prophet/MaxProp performs worse. Both Prophet/MaxProp can achieve only 43% delivery ratio at a node density of 0.5x10-5. MRCT and THEC achieve higher delivery ratios with the Zebranet than with the RWP mobility model but Prophet/MaxProp performs poorer with Zebranet than with RWP model. The difference lies in the fact that both Prophet/MaxProp uses periodic hello messages to exchange neighbor and delivery probabilities information while MRCT uses on-demand route discovery procedure. The periodic exchange of information does not seem to cope well with the more chaotic movement in the Zebranet model. Since THEC is a two-hop approach, it is more adaptive to the chaotic movement and hence THEC performs better than MaxProp/Prophet when Zebranet mobility model is used.

From Figures 9 & 12, we see that with RWP

mobility model, the multihop approaches like MRCT, Prophet and MaxProp achieves lower average delay than the THEC scheme. The MRCT scheme achieves the lowest delay followed by Prophet and MaxProp. The average delay achieved by THEC with RWP mobility model is significantly higher than the other three schemes. Interestingly, with Zebranet mobility model, MRCT and THEC achieve similar average delay. Both MaxProp and Prophet have lower average delay than MRCT and THEC when the node density is 0.5x10-5 with Zebranet model. This can be attributed to the fact that those packets that are not delivered with both the MaxProp and Prophet schemes take longer to be delivered with the MRCT and THEC schemes. Such packets contribute to a larger average delay reported for both the MRCT and THEC schemes.

From Figures 10 & 13, we see that the

transmission overhead for the THEC scheme is the lowest. The multihop approaches incur higher transmission overhead. Both Prophet and MaxProp incur more overhead than the MRCT scheme for both mobility models. Prophet incurs smaller transmission overhead than MaxProp.

of 13

Figure 8: Delivery Ratio with RWP and Zebranet mobility models

Figure 9: Average Delay vs Node Density using THEC and MRCT with RWP and Zebranet mobility models

Figure 10: Transmission Overhead vs Node Density using THEC and MRCT with RWP and Zebranet mobility models.

Figure 11: Delivery Ratio vs Node Density using MaxProp and Prophet Routing Schemes

Figure 12: Average Delay vs Node Density using MaxProp and Prophet schemes.

Figure 13: Transmission Overhead vs Node Density Using MaxProp and Prophet schemes.

It is interesting to observe from Figure 10 that the transmission overhead using the MRCT scheme increases initially with decreasing node density for both mobility models. This can be attributed to the increased retransmissions with increasing network sparseness. Such behavior is also observed when the THEC scheme is used with Zebranet model but its peak occurs at a node density of 1x10-5 node/m2. With the RWP model, we suspect that the peak for the THEC scheme may have occurred at 4x10-5 node/m2.

Similar trend is seen in Figure 13 for the

MaxProp and Prophet schemes with the RWP mobility model. Their peaks occur at a node density of 1x10-5. With Zebranet mobility model, the transmission overheads for both MaxProp and Prophet schemes increase with decreasing node density. We anticipate the transmission overheads for both schemes will drop when the delivery ratio drops further with a node density smaller than 0.5x10-5. Overall, we believe the MRCT scheme is the best scheme for a node density that is higher than 0.5x10-5. However, as we discuss earlier, MRCT scheme is not able to perform well with very sparse network and hence an adaptive scheme as described in [29] should be used for a network

that may be very sparse in some regions and mediumly sparse in other regions. C. Impact of percentage of DTN nodes

In this section, we investigate how the percentage of DTN nodes impacts the delivery performance. In this set of experiments, we simulate 40 nodes randomly distributed in an area of 3000x3000 m2. The nodes move according to the random waypoint or Zebranet mobility model. In addition, we select 10 source/destination pairs that require many hops for packet delivery. A certain percentage of the nodes will turn on DTN custody transfer while the rest of the nodes will not turn on DTN custody transfer. Figures 14,15,16 plot the delivery ratio, the average delay and the transmission overhead when the MRCT and THEC schemes are used. Figures 17,18, 19 plot similar metrics when either MaxProp or Prophet is used.

From Figures 14 and 17, we see that with decreasing percentage of nodes supporting DTN functionality, the delivery ratios for all four schemes drop. The performance degradation with decreasing percentage of DTN nodes is more significant with the MaxProp and Prophet schemes than with the MRCT and THEC schemes.

of 13

Figure 14: Delivery Ratio vs Percentage of DTN nodes using MRCT or THEC scheme.

Figure 15: Average Delay vs Percentage of DTN nodes when either MRCT or THEC scheme is used.

Figure 16: Tran smission Overhead vs Percentage of DTN nodes when either MRCT or THEC scheme is used.

Figure 17: Delivery Ratio vs Percentage of DTN nodes using MaxProp or Prophet

Figure 18: Average Delay vs Percentage of DTN nodes using MaxProp and Prophet.

Figure 19: Transmission Overhead vs Percentage of DTN nodes using Maxprop and Prophet.

With the RWP mobility model, 100% of the nodes need to support DTN functionality to ensure a delivery ratio of at least 58% with multihop routing approaches in a network with a node density of 0.5x10-5. With Zebranet model, both MRCT and THEC can still achieve 80% delivery ratio with only 50% of the nodes supporting DTN functionality.

From Figures 15 and 18, we see that the average delay increases with more nodes supporting DTN functionality. This is because packets incur more retransmissions and waiting when more intermediate nodes support DTN functionality. One interesting observation is the decreasing trend for the average delay that matches closely with what is seen from the delivery ratio plot. The trend is more concave with the MRCT and THEC schemes but more convex with the MaxProp and Prophet schemes.

From Figures 16 and 19, we see that the transmission overhead drops with decreasing percentage of nodes supporting DTN functionality. The decreasing trend is less steep with the MRCT and THEC schemes than with the MaxProp and Prophet schemes. The decreasing trend is steeper with the RWP model than with the Zebranet model for both the MaxProp and Prophet schemes.

D. Impact of Traffic Patterns Next, we investigate the delivery performance

when bidirecitonal traffic model [26] is used. We performed this set of experiment using 10 bidirectional flows with each flow generating 0.25 pkts/sec. The nodes move either according to the random waypoint or Zebranet mobility model. In this set of experiment, we assume all nodes support DTN functionality. We ran 10 runs for each node density value, and compute the average delivery ratio, average delay and transmission overhead. Our results are plotted in Figures 20 to 22 for the MRCT and THEC schemes. Similar results for MaxProp and Prophet are plotted in Figures 23 to 26. From the plots in Figures 20 and 23, we see that the delivery ratios for bidirectional traffic achieved by all four schemes are slightly lower than that for the one-way traffic. From Figures 21 and 24, we observe that the average delay incurred by the bidirectional traffic is only 10-15% more than that for the one-way delay for all the multihop approaches. This can be attributed to the fact that the reverse path is often the reverse of the forwarding path and hence the traffic in the reverse direction can be delivered relatively quickly.

of 13

Figure 20: Delivery Ratio vs Node Density for Bidirectional Traffic using MRCT and THEC schemes.

Figure 21: Average Delay vs Node Density for Bidirectional Traffic using MRCT and THEC schemes.

Figure 22: Transmission Overhead vs Node Density for Bidirectional Traffic using MRCT and THEC schemes.

Figure 23: Delivery Ratio vs Node Density for Bidirectional Traffic using MaxProp, Prophet schemes.

Figure 24: Average Delay vs Node Density for Bidirectional Traffic using MaxProp, Prophet schemes.

Figure 25: Transmission Overhead vs Node Density for Bidirectional Traffic using MaxProp, Prophet schemes.

However, the difference between the average end-to-end delay for bidirectional traffic and the one-way traffic grows (see Figure 21) with decreasing node density when the THEC scheme is used. The

transmission overhead for bidirectional traffic doubles when the MRCT scheme is used but the increase is only about 25% for the MaxProp and Prophet. The increase in transmission overhead for

bidrectional traffic decreases from near 100% to 25% for the THEC scheme as the node density decreases. This may be due to the fact that as the node density decreases, less bidirectional traffic can be delivered and those that can be delivered incur smaller transmission overhead.

Figures 26 & 27 plot the delay distribution

using these 4 schemes and a network with 40 nodes distributed over 2000x2000 m2. From the plot, we see that MRCT has the smallest delay distribution tail while THEC has a long delay distribution tail. Prophet and MaxProp have longer delay distribution tails than the MRCT scheme but shorter tails when compared to the THEC scheme. MaxProp has a slightly longer delay distribution tail than Prophet.

Figure 26: Delay distribution with 100% DTN nodes

using MRCT and THEC schemes.

of 13

Figure 27: Delay distribution with 100% DTN nodes using MaxProp/Prophet schemes

V. CONCLUDING REMARKS

In this paper, we describe four routing schemes for delivering unicast messages in disruption tolerant networks. Via extensive simulation studies, we look into how mobility models, node densities affect the delivery performance of these four routing schemes. Our results show the usefulness of the custody transfer feature in improving delivery ratio of the messages for a network with node density larger than 0.5x10-5 node/m2. With a very sparse network (smaller than 0.5x10-5 node/m2 using a transmission range of 250m), the custody transfer feature alone does not provide sufficiently high delivery ratio for communications. Instead, message ferry needs to be used to achieve high delivery ratio [28], [29].

In addition, our results indicate that all three multihop schemes perform better than the THEC scheme by achieving higher delivery ratios and lower average delay with the RWP mobility model. The price to pay is higher transmission overhead. With the RWP model, the MRCT scheme achieves the highest delivery ratio until a node density of 0.5x10-5 after which Prophet and MaxProp shows better performance. However, with the Zebranet model, the MRCT scheme performs the best among the four schemes. The THEC scheme performs better than both the Prophet and MaxProp scheme with the Zebranet model. In addition, both the Prophet and MaxProp schemes incur higher transmission overhead than the MRCT scheme. We also did some experiments to investigate how the percentage of nodes supporting DTN functionality affects the data delivery performance. Our results show that for the Zebranet model, both MRCT and THEC

schemes can achieve more than 80% delivery ratio even when 50% of the nodes support DTN functionality in a network with a node density of 0.5x10-5 . However, with the RWP model, all nodes need to support DTN functionality to achieve a delivery ratio of at least 58% at the same node density with either one of the three multihop routing schemes. Last but not least, our experiment with bidirectional traffic model indicates that the average end-to-end delay for bidirectional traffic using the multihop approaches is not much higher than that for one-way delay since the backward routes used often are the reverse of the forwarding routes. This is not true with the THEC scheme where the difference between the average delay for bidirectional traffic and one-way traffic grows with decreasing node density. Acknowledgments

This work is sponsored by DARPA under contract W15P7T-06-C-P430. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of DARPA. This document is approved for public release, unlimited distribution. We also thank E. Krohne from Colarado School of Mines for his comments on an earlier version of our paper that help us to improve our presentation. REFERENCES [1] H. Balakrishnan, V. N. Padmanabhan, S.

Seshan, and R. H. Katz, “A comparison of mechanisms for improving TCP performance over wireless links”, IEEE/ACM Transactions on Networking, 5(6):756-769, December 1997.

[2] A. Pentland, R. Fletcher, and A. Hasson, “DakNet: Rethinking connectivity in developing nations”, IEEE Computer, 37(1):78-83, January 2004.

[3] Y. Gitman, “Magic Bike: Wireless Internet”, http://www.magicbike.net/

[4] L. Arent, G4techTV, Tech-Live, “Wi-Fi on Two Wheels”, http://www.g4techtv.com/techtvvault/features/47657/WiFi_on_Two_Wheels.html, May 2004

[5] K. Fall, “A delay-tolerant network architecture for challenged Internets”, Proceedings of SIGCOMM’03, August 2003.

[6] M. Chuah, L. Cheng, and B. D. Davison, “Enhanced Disruption and Fault Tolerant Network Architecture for Bundle Delivery

of 13

(EDIFY)”, in Proceedings of IEEE Globecom, Nov. 2005

[7] T. Spyropoulos, K. Psounis, C. S. Raghavendra, “Spray and Wait : An efficient routing scheme for intermittently connected mobile networks”, Proceedings of ACM Workshop on WDTN, Sept, 2005.

[8] J. Border, M. Kojo, J. Griner, G. Montenegro, Z. Shelby, “Performance enhancing proxies intended to mitigate link-related degradations”, IETF RFC 3135, June 2001.

[9] M. Chuah, P. Yang, B. Davison, L. Cheng, “Store and Forward Performance in a DTN”, poster, Proceedings of VTC Spring, May, 2005.

[10] J. Klensin, Editor, “Simple mail transfer protocol”, IETF RFC 2821, April 2001.

[11] J. Burgess, B. Gallagher, D. Jensen, B. N. Levine, “MaxProp: Routing for Vehicle-Based Disruption-Tolerant Networking”, in Proceedings of IEEE Infocom 2006. April 2006.

[12] W. Zhao and M. H. Ammar, “Message Ferrying: Proactive Routing in Highly Partitioned Wireless Adhoc Networks”, Proceedings of the IEEE Workshop on Future Trends in Distributed Computing Systems, Puerto Rico, May, 2003.

[13] V. Cerf et al, “Delay-Tolerant Network Architecture”, Internet Draft, draft-irtf-dtnrg-arch-06.txt, March 2006.

[14] B. Burns, O. Brock and B. N. Levine, “MV Routing and Capacity Building in Disruption Tolerant Networks”, IEEE Infocom, March 2005.

[15] W. Zhao, M. Ammar, and E. Zegura, “A message ferrying approach for data delivery in sparse mobile ad hoc networks”, ACM Mobihoc, May 2004.

[16] W. Ma and M. Chuah, “Comparisons of Interdomain Routing Schemes for Heterogeneous Adhoc Networks”, Proceedings of WoWMoM, 2005.

[17] K. Fall, W. Hong and S. Madden, “Custody transfer for reliable delivery in delay tolerant networks”, IRB-TR-03-030, July 2003.

[18] J. Broch, D. A. Maltz, D. B. Johnson, Y.-C. Hu, and J. Jetcheva, “A performance comparison of multihop wireless adhoc network routing protocol”, Proceedings of Mobicom, 1998.

[19] Y. Wang, S. Jain, M. Martonosi, and K. Fall, “Erasure-Coding Based Routing for Opportunistic Networks”, Proceedings of Sigcomm WDTN workshop, 2005.

[20] R. Shah, S. Roy, S. Jain, and W. Brunette, “Data mules: Modeling and analysis of a three-tier architecture for sparse sensor networks”, Elsevier Ad Hoc Networks Journal, Vol 1, pp 215-233, Sept. 2003

[21] P. Juang, H. Oki, Y. Wang, M. Martonosi, L.-S. Peh, and D. Rubenstein, “Energy-efficient computing for wildlife tracking: design tradeoffs and early experiences with zebranet”, Proc ASPL0S, Oct 2002.

[22] A. Lindgren, A. Doria, O. Schelen, ”Probabilistic routing in intermittently connected networks”, Sigmobile, Mobile Computing and Communications Review, Vol 7(3), pp 19-20, 2003.

[23] S. Jain, M. Demmer, R. Patra, K. Fall, “Using Redundancy to cope with Failures in a Delay Tolerant Network”, Proceedings of ACM Sigcomm, 2005.

[24] A. Vahdat, D. Becker, “Epidemic Routing for partially connected adhoc networks”, Tehnical Report CS-200006, Duke University, April 2000.

[25] W. W. Brown, T. Krout, “Future Performance Expectations for Mobile Wireless Communication Networks”, preprint, 2005.

[26] J. Hsu, S. Bhatia, M. Takai, R. Bagrodia, M. J. Acriche, ”Performance of Mobile Adhoc Networking Routing Protocols in Realistic Scenarios”, Proceedings of IEEE Milcom, 2003.

[27] UCB/LBNL/VINT, “The Network Simulator ns-2”, Online at http://www.isi.edu/nsnam/ns/

[28] W. Zhao, M. Ammar, E. Zegura, “Controlling the mobility of multiple data transport ferries in a delay-tolerant network”, Proceedings of IEEE Infocom, 2005.

[29] M. Chuah, P. Yang, “Node-density based Adaptive Routing Scheme for Disruption Tolerant Networks”, Proceedings of IEEE Milcom, 2006.

performance comparison of four unicast routing schemes in dtns

Documents