Interest-Centric Mobile Ad Hoc Networks - IEEE Xplore

0 downloads 0 Views 610KB Size Report
Abstract—Mobile ad hoc Networks (MANETs) pose signifi- cant shared communication medium constraints such as finite memory, number of access channels, ...
2012 IEEE 11th International Symposium on Network Computing and Applications

Interest-centric Mobile Ad hoc Networks Renato C. Dutra and Heberte F. Moraes and Claudio L. Amorim Laboratorio de Computacao Paralela e Sistemas Moveis - COPPE Universidade Federal do Rio de Janeiro, Rio de Janeiro, Brazil Email: rcdutra, brebete, [email protected] Abstract—Mobile ad hoc Networks (MANETs) pose significant shared communication medium constraints such as finite memory, number of access channels, and bandwidth to the development of effective communication protocols. Furthermore, multihop message forwarding multiplies the amount of simultaneous transmissions, which augment both channel contention and network congestion, increasing interference and reducing protocol performance. With these issues in mind, we propose a variation of MANET which we called inteRestcentric mobile ad-hoc network or simply Radnet in which every participant node implements in the network layer an Active Prefix (AP) composed of a prefix and an application interest, which the Radnet protocol uses for node identification, message addressing, probabilistic message forwarding, and name search in a distributed way. To evaluate the effectiveness of Radnets for generating lower disturbance in the shared communication medium, thus enabling resource savings, and reducing message overheads, we compared the simulated performances of Radnet protocol (RP) against those of AODV and AODV + Gossip3 (G3AODV) protocols. The results showed that RP achieved a 16 percent higher message delivery rate and one order of magnitude lower latency and message overhead for short 0.5 kB messages in scenarios with 150 mobile nodes. For 5 kB messages, however, AODV and G3AODV rarely ran until the end, due to their overwhelming message overheads, while Radnet protocol showed performances similar to that of the short message case. To evaluate the use of Radnets in practice, we ran experiments in the laboratory, including a chat application on a 20-node Radnet.

of MANETs which we called inteRest-centric mobile adhoc networks or simply Radnets in which every participant node makes use of a novel mechanism namely Active Prefix (AP) to compensate MANETs for missing basic network characteristics while diminishing message overheads. An AP is a simple data structure implemented in node’s network layer and composed of a node prefix and an application interest to be used in a distributed way by an AP-based routing protocol. Specifically, the node prefix is used for node identification, message addressing, and probabilistic message forwarding, whereas the application interest is used for name searching. We developed a Radnet protocol (RP) to evaluate the effectiveness of a routing protocol using node prefixes for generating low disturbance in the shared communication medium, thus enabling resource savings, and reducing message overheads. Our results showed that RP reduced transmission interference significantly, and consequently both latencies and message overheads up to one and two orders of magnitude, respectively, and thus, they can run applications more effectively. The rationale behind our approach is as follows. The intrinsic characteristics of MANETs, notably, selforganization, mobility, multihop message forwarding, and the anonymity of nodes, lead to high overhead in the execution of an IP-based communication protocol because the protocol implementation has to compensate for several functionalities on which such a protocol is based and that are missing on MANETs, such as unique node identifiers, establishment of paths between nodes, route maintenance, and name servers. Thus, we introduce the AP mechanism aiming to reduce this functional gap and enabling Radnets to be built as a simple variation of MANETs while offering a great potential for running practical applications in a simple and effective manner. We used the NS-3.8 Network Simulator [5] to compare the performance of Radnet protocol against two representative IP-based protocols (AODV [6] and the AODV+Gossip3 (G3AODV) [7]) for MANETs in a scenario with 150 mobile nodes uniformly distributed in a rectangular area of 750 m x 300 m. The simulation results showed that RP achieved, on average, 16 percent higher delivery rates as well as latency and message overhead one order of magnitude lower than either AODV or G3AODV. Moreover, we developed and

I. I NTRODUCTION The widespread use of mobile devices with wireless communication interfaces has made applications for mobile networks, particularly mobile ad hoc networks (MANETs), increasingly attractive for physical environments with defective or infrastructure-less communication. However, despite many years of research efforts, routing protocols for MANETs [1], [2] still present limited message delivery, high latency, and large message overhead, which strongly inhibit the applications of MANETs in practice [3]. Recently, it has been noted that the main reason for the limited performance of routing protocols for MANETs is the fact that they assumed node addressing and protocol stack models that were originally designed for IP-centric wired networks, whereas MANETs have characteristics and dynamics radically distinct from their wired network counterpart [4]. Owing to these major differences, we propose a variation 978-0-7695-4773-2/12 $26.00 © 2012 IEEE DOI 10.1109/NCA.2012.32

130

Bit 0

ran a chat application on a 20-node Radnet where each node consisted of a host coupled with a Tmote. These practical experiments validated the simulation results and confirmed the use of APs to easily support the development of applications for Radnets [8]. In summary, the main contributions of this paper are the following: • Introducing the active prefix, a novel mechanism for building new communication protocols for ad hoc networks, particularly efficient cross-layer protocols for Radnets with lower message overheads and latencies between the Application and the Network layers; • Using active prefixes to implement conventional sourceto-destination addressing plus interest group addressing and name searching in Radnets; • Showing that active prefixes enable the development of P2P and client-server applications of Radnets and, also, that they can support traditional IP - based applications; • Presenting results in detail from a significant number of simulations and a practical experiment of a chat application to demonstrate the feasibility of Radnets and the improved performance of Radnet protocol in comparison to two representative IP - based protocols for MANETs. This paper is organized as follows. In Section II, we describe the design of Radnets in more detail and show examples of applications. In Section III, we present simulation results of a comparative evaluation of the performance of RP against those of AODV and G3AODV along with a preliminary experimental evaluation of Radnets. In Section IV we discuss the main results achieved in this work. Section V presents the related work, and Section VI concludes the paper.

24

(a)

N ode P ref ix (Pi ) App Interest (I)

Bit 0

8

(b)

V ersion 48 72 96

16 HopLimit

24 HeaderLength

M essage ID Destination P ref ix Source P ref ix App Interest

Figure 1.

Radnet elements: (a) Active Prefix and (b) AP message header

uses to set its current interest. Figure 1 (b) shows the header of an AP message, which contains a Radnet protocol version, a hop limit, a header length, a message ID, two node prefixes that identify the associated source and destination nodes, and an application interest. A node generates its Pi with n fields, each of which has m bits, in such a way that the nxm bits provide node identification for addressing purposes and a matching filter for message forwarding. Specifically, a node generates a sequence of n fields, where to each field the node assigns a value with m bits using a random variable with a probability distribution. Next, the resulting set of field values, in order, constitutes the node’s Pi while its identification is obtained by concatenating the field values in the same order. For the sake of simplicity, we used the node’s Pi and the node’s identification interchangeably. The communication protocol running in a node implements message forwarding by using the node’s Pi fields as a matching filter with the corresponding Pi fields of messages that the node receives, wherein a particular node compares every pair of Pi fields and forwards the received message if any pair matches; otherwise, the node discards the message. In addition, the message ID and the source prefix are stored in the nodes and used to compare with the corresponding fields of the next received messages so that duplicate messages are detected and not forwarded. For illustration purposes, consider a Radnet with multiple nodes of which two nodes A and B run a Gnutellalike P2P file sharing application and generate 24-bit (8 fields x 3 bits) prefixes PA = (7, 3, 6, 2, 0, 1, 0, 0) and PB = (0, 5, 3, 2, 6, 0, 1, 1) which correspond to identifier A = 73620100 and B = 05326011. In this case, both A and B have the same interest I(A) = I(B) =< Gnutella >. Next, assume that the user at node A asks for f ilex. The Gnutella application will set its payload with a f ilex query, which the Radnet protocol uses to broadcast the message < null >< 73620100 >< Gnutella >< queryf ilex >. On receiving the message, node B will transfer the message to the Gnutella application because I(A) = I(B). Also, node B will forward the message to its neighbors because

II. D ESIGN OF R ADNETS This section presents the design of Radnets and uses two representative examples of applications, namely a file sharing and a web server, to illustrate how Radnets support effective implementations of the P2P and the client-server communication models, and thus Radnets can also be used for other applications of MANETs in general. First, we describe the organization of APs, which features nodes built in a distributed manner and used to communicate, and which implements message forwarding. Afterward, we show how nodes use APs for source-to-destination and interest group addressing modes, and compare the design of Radnets with other alternative solutions. A. Active Prefixes and Message Forwarding Figure 1 presents the components of APs and AP message headers in a typical Radnet. As can be seen in Figure 1(a), an AP is a simple data structure implemented in the node’s network layer and composed of a node’s prefix (P i) and an application (App) interest (I) that each running application

131

PA matches PB on the fourth field value (< 2 >). In the case of node B having f ilex it can send that file to A either by broadcasting the message < null >< 05326011 >< Gnutella >< pushf ilex > or directly to A using the message < 73620100 >< 05326011 >< Gnutella >< pushf ilex > with PA instead of < null >. Note that the use of interest (e.g., < Gnutella >) allows an application running in one node to identify directly the instances of the same application in the other nodes whereas the payload (f ilex) serves to share information among the application instances. Other classes of applications that can benefit from this model include social utilities, games, and emergency applications. On building Pi as a set of probabilistic values we accomplished two main objectives. First, this structure provides node identification in a distributed way with a probability of duplicity as low as (1/2)m.n , for example, in a 100-node Radnet with m = 8 and n = 3, the probability of two nodes having the same prefix pattern is 10−8 . Second, previous works [9], [10] demonstrated that gossip-based protocols exhibit bimodal behavior which explained their high message delivery rates while avoiding flooding. In [11] we showed that Radnet protocol obtained the equivalent behavior and performance results. The benefits of using prefixes instead of interests for message forwarding are threefold. First, we avoid the problems associated with naming that arise by using interest as a string of characters with variable length for a matching filter. Second, popular interests can cause flooding in the network whereas unpopular ones can prevent messages from reaching their destinations. Third, the use of prefixes can reduce the node isolation caused by unmatched interests, as the Pi fields enable neighbor nodes to forward messages collaboratively even if they do not have any common interest. Once a node configures its Pi fields, it functions as a relay because it automatically starts message forwarding independently of whether its I fields have been set. Another advantage is that it is possible to choose the length of Pi to better fit a given Radnet size by defining n and m accordingly. The field values of a node prefix may not match any of its neighbors and thus discard all of the messages it sends and vice versa. Upon detecting this shortcoming, the node either regenerates or adjusts its Pi fields, for example, by copying the value of a Pi field that it received from one of its neighbors.

modes, namely, conventional source-to-destination (S2D) and the Interest-Group (IG) addressing as follows. In the S2D mode, the source node requires the Pi descriptor of the destination node to send its messages. In contrast, in IG addressing mode, the source node only indicates the interest that the I field of the group of destination nodes should match, and sets the destination’s Pi as null to indicate message broadcast. To find other nodes with a common interest, the AP message header is used as a matching filter with the APs of other intermediate nodes, as the message is forwarded from node to node over the network until it finds its group of destination nodes or is discarded. If an interest matching occurs at an intermediate node, the message is copied and delivered to the associate AP application locally. The corresponding source and group of destination nodes can continue to communicate in IG mode, though any node of the group can change to S2D mode, because the source Pi is now known. Note that the intermediate nodes can store the APs of the messages that they receive to allow RPs to use message footprints for the optimization of Radnet addressing; however, this issue is beyond the scope of the present work. An interest in Radnets is primarily used by a distributed application to define interests only as matching filters to find and establish preliminary communication between Radnet nodes running the same application previously downloaded. In the general case, however, the use of interest defined as a string of characters, leads to the problem of identification and finding the information correctly. Current applications solve this problem by coupling a search engine to the communication protocols that present alternatives to the information being searched. Radnets can also adopt the same solution. Node A

Node B

Node C

Client AppSearch PA = 73620100 IA = M ybookstore

Server AppM ybookstore PB = 05326011 IB = M ybookstore

Client AppSearch PC = 3421015 IC = M ybookstore

(1) < null >< P xA >< IA >< null >

(2) < P xA >< P xB >< IB >< U RL >

dropped

B. Message addressing modes In Radnets, each running application on a specific node sets its I field with the appropriate keywords and terms to identify its current interests. For the sake of simplicity, in this work, we assume one single interest field per AP, but a different Radnet implementation can support multiple interest fields. Radnets use APs to implement two addressing

(3) < P xB >< P xA >< IA >< GET >

Figure 2.

132

A basic example of client/server application in Radnets

C. An example of client-server application

the application’s interest matches the one in the message header. The use of interests not only supports the IG addressing mode but it also allows reducing the overheads of protocols for MANETs. In fact, while message discarding at the network layer avoids the overheads of unnecessary message processing by the Radnet network layer and above, the message delivery directly to the associate application enables the Radnet protocol to eliminate the processing overheads between layers an reducing the latency. Furthermore, an interest cannot be simply substituted by a number of a communication port like the ones provided by IANA (Internet Assigned Numbers Authority) for use by application servers because such a number has no meaning for an application. In contrast, in Radnets a running application in a node defines an interest to carry information that is used as a matching filter with the interest of each received AP message by the cross-layer mechanism implemented in the node’s network interface. Finally, the security of data in AP messages can be attained by either cryptographic signatures or passwords using another I field.

Radnets also support TCP/IP applications as Figure 2 illustrates using a basic Client-Server (C/S) application example. In the figure, nodes A and C are clients that wish to connect with Mybookstore web site, and node B is the Mybookstore server, running the application (AppM ybookstore) with the interest set as Mybookstore. The Radnet protocol in nodes A, B, and C generates the corresponding prefixes PA , PB , and PC , respectively. Suppose that a user at node A runs a search application (AppSearch) and asks for Mybookstore, which leads to the interest (IA = M ybookstore) being set. The Radnet protocol at node A builds the message header with PA as source prefix, a destination prefix as null, Interest = M ybookstore, and Payload null. In the next step, the protocol sends the message (1) as shown in the figure. Node C received and dropped the message as PC unmatched the message’s source prefix(PA ). The Radnet application at node C receives the message (IA = IC ) and discards it because the message is from another client application. When node B receives this message, it verifies the interest field and replies to the network with the message (2) and the interest field (IB = M ybookstore) being set. In addition, the message payload will contain a URL that node A will use to establish a connection to the appropriate IP-based protocol, e.g., the HTTP protocol, in node B. Node A will transfer the message received from node B to the running application AppSearch, which stores locally PB and sends the URL to the upper layer using the HTTP protocol. Afterward, the next message of the HTTP protocol will contain the GET method and the URL address of the requested page to the Mybookstore server, e.g., GET /mybookstore/index.html. The Radnet protocol will build the message (3) with < PB >< PA >< IA >< GET /mybookstore/index.html > that is sent to node B using its PB . Nodes B and A will continue to communicate using the HTTP protocol, which steps being performed without modifying its methods, and thus, the node A can download the Mybookstore server’s page through the HTTP protocol superimposed over the Radnet protocol. An alternative solution that combines a gossip-based protocol with random IPs assigned to nodes could obtain some of functionalities that prefixes provide for Radnet protocol, namely node identification and message forwarding. However, an active prefix also defines an interest field, which allows the Radnet protocol running on each node to perform an unique function that has no equivalent in the combined solution, namely the implementation of a crosslayer mechanism between the application and the network layers. Specifically, the interest field of an active prefix in the message header is effectively used on the reception of each message by the Radnet protocol to either discard or delivery it to the associate application depending whether

III. E VALUATION OF R ADNET P ROTOCOL This section presents the simulation scenarios and parameters. We use AODV and G3AODV (AODV+Gossip3 algorithm) protocols as the reference IP-based communication protocols for performance comparison purposes. The main difference between AODV and G3AODV protocols, or simply AODV and G3AODV, is that AODV uses flooding to find routes and maintain the route tables whereas G3AODV instead uses gossiping to considerably reduce the message overheads of the base AODV protocol. We use two simulation scenarios for testing and evaluating short and long message communications, using source-todestination addressing and interest group addressing. In the later, only the Radnet protocol (RP) is evaluated because only it have such an addressing mode. We used 512 Bytes of payload for short message communication, and ten short messages in sequence, comprising a total of 5k Bytes, for long message communication. In addition, we report results from an experimental evaluation using a chat application that we developed and tested using a Radnet built with 20 Tmotes in the laboratory. A. Simulation Testbed and Parameter Configuration We implemented Gossip3 (with p=0.65, m=1, and k=4), G3AODV, and RP in the NS-3.8 network simulator, and we used the native version of AODV. We configured the simulator with 150 nodes uniformly distributed in a rectangular area of 750 m x 300 m; node mobility using Randomwaypoint(RWP) standard with [2,8] m/s and pause times of 1 s, 100 s, and 200 s since RWP has been widely used as the mobility model in preliminary protocol

133

Figure 3.

Delivery rate (5 pairs)

Figure 4.

the number of communication node pairs increased to 30, the delivery rate of RP decreased to 68 percent, whereas the delivery rates of G3AODV and AODV decreased to 55 percent and 50 percent, respectively. Figures 5 and 6 present the message overhead for each protocol according to the number of pairs and the pause time. As can be seen in Figure 5, the message overhead achieved 9, 000 in RP regardless of the pause time, whereas in G3AODV and AODV, the message overhead varied from 8, 000 to 80, 000 for 5 pairs, depending on the pause time.

Delivery rate (30 pairs)

evaluations, AdhocWifiMac, Constant Rate WiFi manager, Wifib-2mbs, 512-Byte message size; transmission range of 100 m; with transmitter and receiver gains of −4 dBm. In NS-3.8 simulations of the RWP model, the pause time is a random time in which a node becomes motionless when it arrives at a randomly chosen destination waypoint (DW) after which it moves again towards another DW and so forth. We simulated an ad hoc network (IEEE Standard 802.11b) with a real link layer for protocol evaluations. We present the average values for 50 simulation runs with a 95 percent confidence interval. The simulations of RP used AP s with (8-fields x 3-bit each) Pi and one single (24-bit) interest field. To each of the eight fields of a prefix was assigned a random value using the binomial distribution function. The nodes stored the APs of received messages, which were forwarded only once, although destinations could receive replicated messages because of multiple path effects. We simulated 5 and 30 pairs of source-to-destination nodes that were randomly selected, and each selected source node sent one or ten AP messages to its destination node in a random interval of 580 ms to 780 ms. We measured, the message delivery rate, the message overhead, and the latency, which are defined as follows: • Message delivery rate - the average of the unique messages received (i.e., duplicate messages dropped) by the destination nodes divided by the number of messages sent by the source nodes; • Message overhead - the total number of messages received by all 150 nodes for a given number of messages sent by the source nodes; • Latency - the average elapsed time between an application sending a message and a remote application receiving the message in the destination node.

Figure 5. pairs)

Message overhead (5

Figure 6. pairs)

Message overhead (30

Figure 6 shows that the message overhead was equal to 12, 000 in RP, whereas the message overhead varied from 80, 000 to 140, 000 in AODV and was 120, 000 in G3AODV for 30 node pairs, and pause times of 1s, 100s and 200s. These results reveal that G3AODV and AODV generate a large number of maintenance messages that increase the transmission interference and explain their one order of magnitude of higher message overhead. In Figure 7, the latency in RP achieved 90ms, whereas in both G3AODV and AODV it was much higher at 1, 200 ms for 5 pairs. In Figure 8, as the number of node pairs increased to 30, the latency in RP increased to 120 ms, whereas the latency in G3AODV and AODV reached 1, 500 ms. Therefore, the latency in RP was between 12 and 13 times lower than that of either G3AODV or AODV. The reason for the extraordinary lower latency is that RP does not attempt to find routes as AODV and G3AODV do; instead, RP uses several alternative paths for message forwarding and an efficient cross-layer mechanism for reducing protocol overheads thereby making message delivery faster. For long messages, G3AODV simulations were unable to complete all of the 50 simulation runs, as Table I indicates

B. Source-to-Destination addressing Overheads To evaluate overheads using source-to-destination addressing we used 5 and 30 source-destination pairs of nodes that were chosen randomly, with each pair having a different interest from the other pairs such that a destination node only accepts a message sent by its corresponding source node. As can be observed in Figure 3, message delivery rates achieved nearly 100 percent in RP and at most 80 percent in G3AODV and AODV for 5 pairs. Figure 4 shows that as

Figure 7.

134

Latency (5 pairs)

Figure 8.

Latency (30 pairs)

(AODV simulations were even worse and they are not shown in the table). Specifically, G3AODV failed for all runs for 15 node pairs, and any chosen pause time (PT); therefore, 15 node pairs sets a limit for G3AODV, in which the received message overhead is the maximum that G3AODV can address. From that point onward, G3AODV begins to fail completely to maintain paths and table routes, which confirms previous results in the literature [4]. Table I G3AODV - SIMULATION RUN FAILURES (RF) FOR LONG MESSAGES (%)

Pairs 5 5 5

PT(s) 1 100 200

RF (%) 62 46 38

Pairs 15 15 15

PT(s) 1 100 200

Figure 11.

RF (%) 98 100 100

Latency in Radnet for 10 messages

varied from 5 to 30. This result explains why the message delivery rate of RP was reduced as the number of pair nodes increased. As expected, the message delivery rate was negatively affected by the number of messages that share the communication medium. Figure 11 presents the latency in RP. As can be seen in the figure, the latency is low at 180 ms and 320 ms, for node pairs equal to 5 and 30, respectively, which can be explained by the low message overhead of RP. C. Evaluation of Interest Group Addressing

Figure 9.

Figure 10.

We evaluated the overheads of IG addressing for group communications using short and long messages. We used two groups of 5 and 30 nodes, in which each group had a single common interest, and the AP destination and I were set with Pi = null and the same interest attribute, respectively; therefore, RP had to find the other group nodes with the same interest to deliver the messages sent by the node that was randomly chosen. Figure 12 presents the results of the message delivery rates. For short messages, the message delivery rate was practically 100 percent for the group with a size of 5 nodes, and it was reduced to 68 percent when the group size increased to 30 nodes. For long messages, the message delivery rate achieved almost 100 percent for the group size of 5 nodes and was reduced to 50 percent as the group size increased to 30. Figure 13 shows that the short message overhead achieved 5, 000 and 18, 000 messages for group sizes of 5 and 30 nodes, respectively, regardless of the pause time. However, the long message overhead increased to 7, 000 and 24, 000 for group sizes of 5 and 30 nodes, respectively, because the noise augmented in the communication medium. Figure 14 presents the latency results. This figure shows that short message latencies achieved 120 ms and 240 ms for communication group sizes of 5 and 30 nodes, respectively. The short and long message latencies were the same for group size of 5 nodes whereas the long message latency increased for group size of 30 nodes.

Delivery rate in Radnet for 10 messages

Message overhead in Radnet for 10 messages

In contrast, RP completed all the executions for long messages and the results are presented in Figures 9, 10, and 11. Figure 9 shows that the message delivery decreased from 98 to 51 percent as the number of pair of nodes varied from 5 to 30. This shows that an increase in the number of message transmissions augments the contention for the communication medium, and, therefore, reduces the number of messages being transmitted successfully and thus, the message delivery rate. The message overhead curve shown in Figure 10, increased from 8, 000 to 23, 000, as the number of pair nodes

D. Experimental evaluation We also built a basic 20 - node Radnet to validate the simulation results, where each node consisted of a

135

experimental results were similar to those of the simulation results, although they used a different transmission standard (Zigbee); specifically, they supported the simulation results and confirmed the significant impact of the transmission delay on the protocol performance as well as indicating the feasibility of using RP in practice. IV. D ISCUSSION

Figure 12.

Figure 13.

The simulation and experimental results of previous sections reveal that the use of Radnet protocol for Interestcentric Mobile Ad hoc Networks is notably promising, due to their significant improvements to message overhead, message delivery rates, and latency in comparison with two representative IP-based protocols, thanks to Radnet protocol lower interference in the shared communication medium. The low message overhead reduced the network contention as shown in Figures 15, and 16 that point out the failure message overhead for short messages. This overhead indicates the total number of received message with error by all the 150 nodes, or the message delivery failures using the shared channel (failure messages caused by the MAC layer were nearly zero). In the figures, the failure message overhead was one order of magnitude lower in RP than either G3AODV or AODV, for 5 pairs; and two orders of magnitude lower in RP for 30 pairs. Therefore, this largely favorable performance results for RP can be explained by the large transmission interference caused by the large number - over one million for 30 node pairs - of maintenance messages that were required by either G3AODV or AODV. The high delivery rates of RP are due to the probabilistic message forwarding provided by the active prefix, which enables almost all of the network nodes to be reached. Moreover, the low latency achieved in RP is result of the active prefix to support a cross-layer mechanism, since it allows the network and application layers to communicate directly, using the interest field. In this manner, a message can be checked quickly using the interest as a matching filter at the network layer to determine whether the message contains information of interest to a local application, and whether it should be discarded. Another important factor is Radnet make no use of routes. Overall, the higher performance of RP in comparison with the performance of either the G3AODV or AODV reveals that it is less important which node forwards a message. For Radnet, it suffices that there exists a node in the path to delivery the message with a given rate of success. Adding to this property is the fact that RP does not cause contention, which allows IP-based applications to work better with them than either of the other two protocols.

Delivery rate in interest-group addressing

Message overhead in interest-group addressing

Figure 14.

Latency in interest-group addressing

host coupled with a Tmote (Zigbee) that allowed hosts to communicate either directly or by multihop. Recall that the node mobility had little impact on the Radnet performance; therefore our testbed suffices. Specifically, we compared the message delivery rate of RP 24-bit (8 fields x 3 bits) prefixes with the delivery rate of the Gossip algorithm (p = 85 percent). Moreover, we developed a chat (Java) application that our lab team used to test the Radnet over one week and a network monitoring tool to track the Radnet’s overheads. Specifically, we evaluated RP for the transmission of 100 messages between each of 2 pairs of S2D nodes with a transmission delay varying from 5 s to 60 s. Overall, the results indicated on average that the message delivery rate achieved 84 percent for RP and the Gossip algorithm; in addition, RP delivered messages with 2.4 hops, a 19 ms latency between two neighbor hosts, and 360 ms for the application-toapplication communication which includes 200 ms of Java’s overhead, and 1, 450 messages for message overhead. These

V. R ELATED W ORK There is a vast literature on routing protocols for MANETs with specific methods for route discovery and

136

Figure 15.

to reach destinations. Direct Diffusion generates a new gradient for each new piece of information, which is stored in network nodes even if a network node has no interest in the message. They use in-networking processing, a list of attribute-value-operation tuples, and cache data. By contrast, Radnets neither cache any data nor perform in-networking processing, but they use active prefixes for addressing and forwarding and use AP interest attributes for communication between nodes sharing the same application. Also, their work does not support node addressing, and thus no end-toend addressing is possible; in contrast, Radnets use APs as node addresses and provide source-to-destination addressing.

Failure messages (5 pairs)

VI. C ONCLUSION

Figure 16.

We addressed the problem of high overheads that MANETs cause to communication protocols. We showed that an alternative solution was to add new functionalities to the MANET nodes in a distributed way to reduce interference in shared communication medium. For this purpose, we proposed a variation of MANETs which we called interest-centric mobile ad hoc networks or simply Radnets in which every participant node makes use of a novel mechanism we introduced namely Active Prefix (AP) to compensate for MANET infrastructure-less while diminishing message overheads. In Radnets, APs are used for probabilistic message forwarding and addressing, and also as matching filters for searching the Radnets for nodes with the same application interest. A distinguishing property of APs is that they enable nodes and application interests to be uniquely identified in a distributed way. In addition, APs support interest group communication and source-todestination addressing; moreover, they allow nodes to implement cross-layer communication. Furthermore, Radnets do not rely on IP addressing or identification, decentralized IP access control, or routing tables, but instead support communication that is focused on users and applications and can be adapted to run IP-based applications. Simulation results indicated that Radnet protocol is significantly more effective than either AODV or G3AODV, two representative IP-based protocols for MANETs, in a scenario with 150 mobile nodes. Furthermore, the simulation results were supported by experimental results based on the performance of a 20-host Radnet and a chat application. Our experimental and simulated results showed that Radnet protocol have low overheads and that they can potentially be used for applications of MANETs in general. These promising results strongly support our approach to the problem of building effective communication protocols for MANETs; moreover, the performance of Radnet protocol indicates that it is excellent candidate for applications based on group communication, which is the case with many applications of MANETs such as service-discovery, social utilities, games, and emergency applications. As one step in that direction, we made our Radnet protocol (Android 2.1 upward), the chat application, and

Failure messages (30 pairs)

maintenance messages for node identification and naming [3]. They mostly inherited IP-based addressing and topological node identification [3], [4], [12]. To improve performance of routing protocols, Gossip-based algorithms [9], [13] have been proposed such as the representative G3AODV protocol [7], which implements the Gossip3 algorithm for probabilistic forwarding in the AODV protocol. Gossipbased algorithms exhibit bimodal behavior, which can guarantee high message delivery rates, yet with lower message overheads than those of flooding methods. Although AP networks also use probabilistic forwarding and do exhibit bimodal behavior they do not rely on the topological identification of nodes or routing tables. Geographical routing protocols represent another alternative solution to the identification problem, but they depend on nodes using GPS whereas Radnet protocols do not. Content-based networks [14] support Publish / Subscribe overlays on MANETs, using IP-based addressing, decentralized brokers and identification control, and also exhibit high maintenance overheads. Moreover, the use of content for node identification is challenging because of classical database problems [12], [15]. Radnet protocols can also have the same problem on using a generic name content as interest, but an interest in Radnets is defined by a distributed application for communication between nodes running the same Radnet application. Recent works on naming [4], [16] are the closest to ours as they are not IP-based but data-centric. In one study [16], nodes communicate by sending messages based on user’s interests and utilize the Direct Diffusion method

137

the API for developing applications of MANET available for download at http://www.lcp.coppe.ufrj.br/radnet.

[9] R. Friedman, D. Gavidia, L. Rodrigues, A. C. Viana, and S. Voulgaris, “Gossiping on manets: the beauty and the beast,” SIGOPS Oper. Syst. Rev., vol. 41, pp. 67–74, October 2007. [Online]. Available: http://doi.acm.org/10.1145/1317379.1317390

R EFERENCES [1] I. Akyildiz, X. Wang, and W. Wang, “Wireless mesh networks: A survey,” Computer Networks, vol. 47, pp. 445–487, 2005. [Online]. Available: citeseer.ist.psu.edu/akyildiz05wireless.html

[10] Z. J. Haas, J. Y. Halpern, and L. Li, “Gossip-based ad hoc routing,” IEEE/ACM Trans. Netw., vol. 14, no. 3, pp. 479– 491, 2006.

[2] J. Broch, D. A. Maltz, D. B. Johnson, Y.-C. Hu, and J. Jetcheva, “A performance comparison of multi-hop wireless ad hoc network routing protocols,” in MobiCom ’98: Proceedings of the 4th annual ACM/IEEE international conference on Mobile computing and networking. New York, NY, USA: ACM, 1998, pp. 85–97.

[11] R. C. Dutra, H. F. Moraes, and C. L. Amorim, “Active prefixes for mobile ad-hoc networks,” Technical Report ES 739 / PESC - COPPE - UFRJ, Tech. Rep., 2011. [12] A. Demers, D. Greene, C. Hauser, W. Irish, J. Larson, S. Shenker, H. Sturgis, D. Swinehart, and D. Terry, “Epidemic algorithms for replicated database maintenance,” in Proceedings of the sixth annual ACM Symposium on Principles of distributed computing, ser. PODC ’87. New York, NY, USA: ACM, 1987, pp. 1–12. [Online]. Available: http://doi.acm.org/10.1145/41840.41841

[3] M. Conti and S. Giordano, “Multihop ad hoc networking: The theory,” Communications Magazine, IEEE, vol. 45, no. 4, pp. 78 –86, april 2007. [4] M. Meisel, V. Pappas, and L. Zhang, “Ad hoc networking via named data,” in Proceedings of the fifth ACM international workshop on Mobility in the evolving internet architecture, ser. MobiArch ’10. New York, NY, USA: ACM, 2010, pp. 3–8. [Online]. Available: http://doi.acm.org/10.1145/1859983.1859986 [5] “Ns-3 network simulator,” http://www.nsnam.org, 2011.

offical

[13] A. V. Kini, V. Veeraraghavan, N. Singhal, and S. Weber, “Smartgossip: an improved randomized broadcast protocol for sensor networks,” in IPSN ’06: Proceedings of the fifth international conference on Information processing in sensor networks. New York, NY, USA: ACM, 2006, pp. 210–217.

site:

[14] A. Carzaniga and C. P. Hall, “Content-based communication: a research agenda,” in SEM ’06: Proceedings of the 6th international workshop on Software engineering and middleware. New York, NY, USA: ACM, 2006, pp. 2–8.

[6] C. Perkins, E. Belding-Royer, and S. Das, “Ad hoc ondemand distance vector (aodv) routing,” RFC 3561, 2003. [Online]. Available: http://www.ietf.org/rfc/rfc3561.txt

[15] G. V. Popescu and Z. Liu, “Network overlays for efficient control of large scale dynamic groups,” in DS-RT ’06: Proceedings of the 10th IEEE international symposium on Distributed Simulation and Real-Time Applications. Washington, DC, USA: IEEE Computer Society, 2006, pp. 135–142.

[7] Z. Haas, J. Halpern, and L. Li, “Gossip-based ad hoc routing,” in INFOCOM 2002. Twenty-First Annual Joint Conference of the IEEE Computer and Communications Societies. Proceedings. IEEE, vol. 3, October 2002, pp. 1707 – 1716 vol.3.

[16] J. Heidemann, F. Silva, C. Intanagonwiwat, R. Govindan, D. Estrin, and D. Ganesan, “Building efficient wireless sensor networks with low-level naming,” SIGOPS Oper. Syst. Rev., vol. 35, pp. 146–159, October 2001. [Online]. Available: http://doi.acm.org/10.1145/502059.502049

[8] H. F. Moraes, R. C. Dutra, and C. L. Amorim, “On developing interest-centric applications for ad hoc networks,” in Proceedings of the 13th IEEE International Symposium on a World of Wireless, Mobile and Multimedia Networks - WoWMoM 2012 - To appear. Los Alamitos, CA, USA: IEEE Computer Society, 2012.

138