CNN: A Cluster-based Named Data Routing for Vehicular Networks

VANET routing aims to interconnect vehicular nodes via wireless links to transmit network packets. The goal of VANET routing is to reduce the communication cost, decrease the latency and increase the interoperability of the network. In this paper, a cluster-based routing protocol for VANETs called CNN is proposed. It takes advantage of the Hamming distance technique to partition a vehicular network into information-centric clusters based on the mobility of vehicular nodes. However, the proposed approach uses a named data network technique to forward network transmissions according to a hybrid communication model of Dedicated Short Range Communication and Mobile Agent. The former focuses on reactive intra-cluster link establishment, while the latter proactively forwards inter-cluster transmissions. A simulation measures the performance of the proposed approach and compares the results with two well-known VANET routing protocols: AODV and A- STAR. According to the simulation results, CNN outperforms the benchmarks in average end-to-end delay, path length, data delivery ratio, and total transmitted traffic, especially when the network is dense and the nodes are highly mobile.


I. INTRODUCTION
Recent advances in Vehicular Ad-hoc Network (VANET) address several new applications such as autonomous vehicle navigation, traffic monitoring, and fleet management [1]. Autonomous driving vehicles (e.g., driverless cars or AGVs) collect and analyze ambient data such as environmental conditions, traffic status, road safety, and location information for navigation. They may need to communicate and share their data to avoid accidents, obstacles, and traffic, leading to increased public safety.
VANET includes vehicle nodes that are connected via cellular (e.g., 5G), WiFi, Bluetooth, and Zigbee links. The nodes are typically equipped with sensors/cameras to collect environmental data and use a wide range of on-broad transceivers (e.g., simple RFID or satellites) to communicate with Road-Side Unites (RSUs) or with each other [2].
VANET is similar to mobile Ad-hoc networks (MANETs) in terms of routing, as both are self-configuring [3] and [4]. However, additional routing issues need to be addressed for the following reasons [5], [6] and [7]: (1) The speed of VANET nodes is comparatively higher than MANET.
(2) Data transmissions are more time sensitive in VANETs than in MANET. Delayed and stale transmissions in VANETs can lead to transport and security risks. (3) In VANETs, there is a predictable mobility pattern as nodes usually move along predefined roads. However, the mobility of MANET is usually random. (4) VANET nodes usually do not suffer from energy constraints as they are continuously powered, whereas MANET nodes suffer from high energy constraints. (5) VANET changes network topology comparatively faster than MANET due to fast mobility. (6) VANET increases the transmitted network traffic compared to MANET due to frequent topology changes. (7) VANETs are large, dense and unbounded while MANETs are usually limited and sparsely distributed.
Tables (PIT) at each receiving (intermediate) node if they are new and do not correspond to any PIT record. They are forwarded through the Forwarding Interest Base (FIB) until the content store (i.e., the destination node) is reached [11]. However, a DI packet is discarded if PIT matches a link record. Although NDN routing can provide benefitsespecially a loop-free route-it increases network traffic and resource consumption when the network topology changes rapidly [12], [13]. To solve this problem, [14] proposes a distributed adaptive caching strategy technique over a hierarchical NDN backbone to reduce the amount of transmitted traffic for large and congested vehicular networks.

FIGURE 1. NDN diagram
This article proposes a routing algorithm called Clusterbased Named data routing for vehicular networks (CNN) to address the existing drawbacks of VANET routing. VANET routing suffers from frequent changes in network topology. This leads to increased network resource consumption, message collisions, communication delays, and increased network traffic, especially when the network is large and dense and the vehicular nodes are moving fast. To address these drawbacks, CNN combines Dedicated Short Range Communication (DSRC) and Mobile Agent (MAs) routing models and takes advantage of the NDN technique to establish IC links over which network traffic is routed. CNN uses a lightweight and decentralized clustering technique, Hamming distance [15], to support on-demand, context-aware ad hoc network partitioning according to node mobility. It reduces the network transmission since the network clustering technique restricts the network communication according to the convergence principle. However, the cluster heads perform in-network data aggregation to avoid transmission of redundant packets. The DSRC communication establishes connections within the cluster, while the MAs are responsible for the routes between the clusters. The contributions of CNN are described below: • The proposed approach uses a lightweight, decentralized, and self-configuring clustering technique to partition the network and support in-network aggregation according to IC. The result is a reduction in transmitted traffic and network resource consumption.
• CNN takes advantage of the IC connection setup, which results in a higher data delivery ratio. • CNN uses the NDN technique to reduce the data traffic transmitted over the network, resulting in a lower message transmission error rate and greater conservation of network resources, especially bandwidth. • The proposed approach takes advantage of (NDNenabled) MAs to minimize path length, avoid looping, and reduce communication delay. This paper is organized as follows: Section II reviews the routing techniques of VANETs and highlights the existing drawbacks and advances in this research area. Section III presents the remarks, features and techniques of CNN and explains how CNN addresses the existing drawbacks of VANET routing. Section IV presents the experimental research plan and describes a simulation model to evaluate CNN performance. Section V discusses the evaluation results of the proposed protocol using four key metrics: average end-to-end delay, path length, data delivery ratio, and total transmitted traffic. The performance of CNN is compared with two well-known benchmarks: A-STAR [16] and AODV [17]. Section VI summarizes the contributions of CNN and outlines the main points of this research that should be addressed in further work.

II. RELATED WORKS
This section briefly discusses the VANET routing categories and highlights their design principles, superiorities, and functionalities. According to Figure 2, VANET routing protocols are classified based on schemes, network architectures, transmission strategies, and routing information features [18], [19].

A. ROUTING SCHEME
There are two VANET routing schemes: Client/server and mobile agent (MA) [20]. Client/server forms single-hop or multi-hop DSRC ties to forward network traffic through [4]. Optimized Link State Routing (OLSR) [21] is a client/server routing protocol that supports DSRC and forwards network traffic through multi-hop links using a multipoint relaying 1 MA can be either in the form of software (e.g. code migration) or hardware (e.g. drones). In this research, software MAs. flooding) increases the number of transmissions because it forwards network traffic without restriction. Multicast routing avoids flooding but forwards network transmissions to specific destination nodes. Compared to unrestricted broadcast routing, it reduces the traffic transmitted over the network. Location-Aided Routing (LAR) [27] is a limited broadcast routing that forwards network packets to the nodes that are located in a specific area. Unicast routing avoids scheduling multiple paths. It aims to establish a single connection to each target node and forward network traffic through.
VANETs usually use broadcast communication because they have no energy constraints. However, packet broadcast is easy to implement and can cover both unicast (e.g., a single target ) and multicast (e.g., several targets ) communications. However, it increases overhearing as every node has to repeatedly listen to the wireless channel and check the received network packets.
CNN uses the NDN technique to establish context-aware communication, which reduces the overhead of overhearing, limits broadcasting, and reduces the number of transmissions.

D. TOPOLOGY-BASED ROUTING
Topology-based routing forwards network traffic according to two models: proactive and reactive. Proactive routing (or table-driven routing) aims to collect routing information and store it in tables. It reduces routing delay compared to reactive path planning [20]. However, updating routing tables increases routing cost, network resource consumption and transmitted traffic when nodes are moving fast. Destinationsequenced Distance-Vector (DSDV) [28] is a well-known proactive routing protocol that keeps the path length (hop count) and updates the routing tables in two forms: incremental and full-dump. In the former, only the changed records in the table are changed, while in the latter, the entire table is updated.
Reactive routing builds a path on demand. In fact, in reactive routing, a node forwards a route request when it needs to establish a communication link. This reduces the routing cost and the consumption of network resources. However, reactive routing increases the delay compared to proactive routing. Ad-hoc On-demand Distance Vector (AODV) [17] is a reactive routing protocol that forwards route requests either unicast or multicast as needed. Route requests are recorded and forwarded accordingly through intermediate nodes until destinations are reached.
Hybrid routing combines the key aspects of proactive and reactive routing to reduce routing costs and delays in VANETs. There are two techniques for hybrid routing: proactive zones and network backbone [20]. In the first case, network zones (e.g., clusters) support network communication, while in the second case, a stable network backbone forwards network traffic. CNN takes advantage of a network backbone to support hybrid routing. VOLUME 4, 2016

E. GEOGRAPHIC-BASED ROUTING
Geographic-based (location-based) routing uses location information to forward the network traffic. This results in a reduction in routing traffic transmission and delay. The geographical information is usually obtained from the Global Position System (GPS). As a result, location-based routing usually suffers from GPS availability and accessibility.
Anchor-based Street and Traffic-Aware Routing (A-STAR) [16] establishes the shortest path with maximum connectivity in VANETs. It finds the shortest link between the source node and the destination node using Dijkstra route selection algorithm. For this, A-STAR uses a GPS, to calculate and evaluate the path length of all available routes, and then selects the shortest one if it has the maximum connectivity.
CNN avoids the continuous GPS information collection due to the cost and constraints. It uses RSUs to collect mobility data (e.g., roads, lanes, and directions) on demand and at minimal cost.

III. THE PROPOSED PROTOCOL FIGURE 3. CNN's Methodology Diagram
In this section, we propose a cluster-based routing protocol called CNN that uses NDN to forward network traffic in VANET. It supports IC communication to reduce network traffic, communication delay and network resource consumption. For this purpose, CNN partitions a vehicular network according to the mobility of nodes (e.g., region and direction) using Hamming distance technique. DSRC communication supports intra-cluster connections and in-network packet aggregation, while MAs forward inter-cluster packets over an RSU backbone. Figure 3 shows a diagram summarizing the CNN methodology including steps, objectives, and techniques.

A. THE NETWORK MODEL
CNN's network model consists of vehicular nodes and Road-Side Units (RSUs) interconnected by wireless links. The RSUs are typically equipped with roadside sensors/cameras for data collection and wireless transceiver modules for communication. They provide a robust communication infras- tructure capable of supporting node connections. Nevertheless, RSUs can provide location information, navigate vehicle nodes, transmit network traffic (e.g., MAs), and report ambient data. Figure 4 shows a CNN application in which vehicle nodes move through dedicated (urban) streets. They are grouped according to their mobility (e.g., location). Each group is managed by a cluster head (CH) whose task is to perform innetwork data aggregation, intra-cluster communication, and inter-cluster data dissemination (e.g., MA ).

B. NETWORK CLUSTERING
CNN uses a distributed clustering algorithm to partition the network according to the mobility of the nodes (i.e., movement region, road ID, and direction). Clustering aims to reduce network traffic as it supports packet aggregation at each CH. However, it reduces packet congestion and collisions, especially when the wireless communication channel is restricted [29]. Network clustering decreases transmitted traffic and consequently minimizes the consumption of network resources, especially bandwidth, power, and memory. Moreover, clustering supports contention-free MAC (Medium Access Control) protocols and provides benefits such as reduced overhearing and idle listening [30].

1) Network Clustering
CNN forms the clusters in a self-configuring manner. According to Algorithm 2, nodes are clustered if they are in the same region and move through the same road ID and lane /*Calculate Hamming distance between node 1 and node 2. */ Hamming function HAM(LC1 ∧ LC2) /*starts from the first bit until end.
Algorithm 1: Hamming Distance Algorithm number. For this, Hamming distance technique [15] is used to find similarities in mobility. Here, the number of flipped bits is counted to understand the distance (dissimilarity) of node mobilities. As Algorithm 1 shows, Hamming distance counts the number of different (flipped) bits in the location codes (LCs) containing the address and road/lane number of the node's moving region. The vehicular nodes receive/update LCs as soon as they find a connection to RSUs. For example, the LC of node B contains {road 1 & lane 1} according to Hello (H) messages are simple and lightweight packets used to form clusters. They are commonly used to set up ad hoc networks, join a communication infrastructure, and/or reconnect [20], [31], [32]. In this research, H messages consist of two fields: Data and LC. The former consists of data parameter naming (e.g., data interest), routing information, and clustering data (e.g., data type), while the latter contains information about node location and mobility, especially region ID and road/lane number. Each H message is assigned with a timestamp (tsmp) to avoid looping and support data freshness. Thus, an H message is immediately discarded when tsmp has expired. CHs and the nodes with false cluster flags (CFs) are the only nodes allowed to send an H message. CHs forward H messages to invite next-hop nodes to join their clusters, while false CF nodes use them to join a cluster. They broadcast H messages as soon as an RSU link is available. CF flag is true if the node belongs to a cluster, while it is false if there is no cluster association (e.g., a new node). A node whose CF is false presents itself as CH if no H response is received until the timer Timer(t) expires. Timer(t) is used to avoid starvation and a long wait for joining a cluster. [33] recommends a two-second timer on a singlehop path until the response to an H message is received. The topology of the VANET changes rapidly due to the mobility of the nodes. This means that VANET clusters are frequently updated as vehicular nodes lose links within clusters due to changes in direction or region. Therefore, nodes set CF to false before leaving a cluster. The leaving node is detected as disconnected until it joins a new cluster or forms a new cluster. As Algorithm 3 shows, leaving nodes may need to take different actions depending on their role (CH or CM): • A CH should find a replacement to keep the cluster VOLUME 4, 2016 interconnected as much as possible. To this end, CH orders the routing table entries according to the degree of connectivity. Therefore, a node with maximum connectivity is selected as the next CH candidate if the ranked list is not empty. The candidate CH rebuilds the cluster to cover the maximum number of CMs in a selfconfiguring and decentralized manner. This prevents a cluster from breaking into isolated nodes when a CH is eliminated. However, small clusters are formed if the candidate CH has a low connectivity degree. Section III-B3 addresses how CNN combines small clusters into larger clusters. • A simple CM leaves a cluster by broadcasting a leave message to inform the CH and neighboring CMs to update their routing tables. The leaving CM resets its CH's ID to null and CF to false and then leaves the cluster.

3) Cluster Combination
Frequent cluster updates have to do with the instability of the network topology and bring disadvantages -mainly the consumption of network resources. To solve this problem, CNN combines clusters with the goal of reducing the number of (isolated) nodes/clusters. As Algorithm 4 shows, neighboring clusters are combined into one large cluster if they share the same mobility (e.g., the same LC). This leads to a reduced number of clusters and consequently to increased in-network data aggregation and reduced network transmissions. To this end, CHs that are directly linked and have a Hamming distance of zero participate in a cluster combination. In this way, each CH advertises its connectivity degree using an H message as soon as an RSU link is available. The receiving node (i.e., CHs or nodes with false CF) joins the advertising cluster and updates its neighbors (e.g., CMs) if it has a lower connectivity degree, while it replies to the message and asks the sender node to join if it has a higher connectivity degree. In this section, we present a CNN routing approach in which network packets are classified into two classes: Data Interest Query (DIQ) and Data Interest Reply (DIR). DIQs are forwarded from source nodes to request VANET routing information such as road safety, traffic conditions, and/or parking availability, while DIRs are routed back from targets (i.e., CSs) to source nodes to provide the requested data.
DIQ and DIR are structured in four fields: Source Mac Address, Data Interest/Response Content, Relay Mac Address, and Packet Timer. MAC addresses are unique IDs to identify the nodes. Data interest/reply allows nodes to check or update the message content according to IC routing and FIB strategy. Relays are the last interconnected nodes on each path. They are used for immediate response and establish a return path. This allows nodes to maintain a return link with minimal cost.
A timer is assigned to each network packet to provide data freshness and avoid looping. CNN assumes that the nodes are synchronized and the source nodes reset the timer before each transmission. According to [33], the allowable interval for lost messages is one second for IEEE 802.11. Thus, the timer is set to one second to support data and avoid loss of data packets. This means that a packet is still fresh and should be forwarded for further processing if it is delivered within one second. However, in real applications, data packets may have different priorities that provide a different interval for packet loss. Some messages (e.g., accident alerts) have a high priority and should be delivered with minimal latency, while others have a low priority (e.g., parking availability). CNN assumes that all network packets have the same priority and uses a packet loss time (one second). However, Section VI addresses a future work to extend CNN where routing messages have different priorities.

1) Intra-cluster Routing (NDN)
As Algorithm 5 shows, source nodes take advantage of NDN routing to forward DIQs. To this end, each CH checks PIT, to find a record that matches the DIQ. It updates PIT, discards DIQ, and forwards back a DIR if the DIQ record matches PIT. Otherwise, it sends a MA, which moves through the RSU infrastructure to find the requested CS in another cluster. Expired DIQs (or DIRs) are immediately discarded to support data freshness and to avoid a loop.
DIRs are forwarded from CSs (e.g., target nodes) to source nodes, using Algorithm 6 takes two approaches: DSRC and MA. DSRC is used for intra-cluster forwarding DIR, where a node checks PIT for a matching record when it receives DIR. The node drops the DIR packet if there is a PIT record. Otherwise, DIR is forwarded and PIT is updated accordingly. MA is used for inter-cluster DIR routing. With this, CH dispatches and forwards a MA to the RSU infrastructure and forwards it to catch the source node. of four main components: identity, data, code, and path. Identity contains the basic information of MA such as MA, ID, dispatcher and time. MA, data addresses the DIQ or DIR parameters, while code contains MA migration algorithm and routing information. Path records the MA's traversed path list of visited nodes throughout a journey. As Algorithm 7 shows, CHs dispatch MAs when inter-cluster communication is required to forward DIQ/DIR. A CH receives and aggregates data requests from CMs and then sends MAs throughout the network to reach the target nodes.
CHs forward MAs to the available RSUs according to an NDN routing. The RSU checks its routing table and forwards the MA if it matches the MA's DI and the target cluster is still interconnected. Otherwise, the RSU finds out the cluster's movement pattern (i.e., direction), computes the LC of target CH, and forwards the MA. However, the RSU clones the MA into multiple copies and forwards it to the next RSUs if finds no DI. Although cloning MA is avoided in resource-constrained networks (e.g., wireless sensor networks) [6], it has the potential to increase the parallelism of data dissemination when nodes are equipped with continuous power resources (e.g., VANET). The same algorithm forwards back the DIR MA to the source cluster once the target is captured. CNN monitors MA itineraries to avoid a loop if a particular RSU is visited more than once. MA loops can cause network congestion, packet collisions, and consume network resources.

IV. EXPERIMENT METHODOLOGY
Simulations are used to test and evaluate the performance of CNN. The reason is that real vehicular networks are very expensive in terms of communication infrastructure, navigation platform and route planning. Moreover, real VANETs are capable of handling public safety risks such as accidents and road congestion. Accordingly, simulations are often used to test VANET routing protocols. OMNET++ [34] is a discrete event simulator that has an open source framework called INET [35] to support node mobility, MAC protocol and communication model. It provides the implementation of well-known VANET routing protocols (e.g., AODV). Therefore, OMNET++ was chosen to implement, test and evaluate CNN. Nevertheless, a Metropolitan Grid (M-Grid) Manhattan model [36] is used to simulate a grid-based road infrastructure through which vehicular nodes move.
As Figure 5 shows, three scenarios are studied in the simulation experiments: sparse, medium, and dense networks with 50, 250, and 500 nodes, respectively. In this way, the scalability and performance of CNN are studied as a function of the change in network density. Each vehicle is assigned a unique ID and is randomly localized in a (12.5km × 12.5km) field. It moves through the predefined roads provided by the M-Grid mobility pattern. According to Figure 6, the grid infrastructure consists of 500m squares (e.g., city buildings) and 100m two-lane roads. The vehicle nodes are assigned a constant speed for each simulation  A Carrier Sense Multiple Access (CSMA/CA) MAC protocol is used in the simulation experiments to enable wireless channel sharing and avoid collisions. In addition, a lineof-sight (LOS) wireless signal propagation model is used for all simulation experiments. This assumes no wireless interference from physical obstacles (e.g., trees and buildings) and ambient factors (e.g., noise and weather). For each experiment, source and target nodes are randomly selected to simulate data packet dissemination.   Four metrics are measured to evaluate CNN performance: average end-to-end delay, route length, data delivery ratio, and total traffic transmitted. These metrics are commonly used to evaluate the performance of VANET routing protocols in the literature [9], [38], [39], [13].
1) Average End-to-End Delay (ETE): it measures the delay since a packet leaves a node until it reaches the target. ETE is affected by the number of intermediate hop count, node mobility, packet transmission latency, bandwidth availability, and network traffic. Data accuracy and freshness are achieved when ETE is minimized.
2) Route Length (RL): this is the number of intermediate hop count from a source to a target node (e.g., CS). RL is collected to study the routing protocol's ability to avoid random/blind path selection and establish paths with minimum hop count paths. A reduced RL will result in a lower ETE. 3) Data Delivery Ratio (DDR): counts the number of successful replied DIQs. A data query is replied if the routing approach successfully establishes a path between a source and target. The DDR can be affected by the MAC protocol, network traffic, available bandwidth, and routing algorithm design. 4) Total transmitted Traffic (TtT): it represents the total number of network messages (e.g., control and data packets) transmitted during the routing process. An increase in TtT results in increased consumption of network resources -mainly bandwidth, increased packet congestion and reduced DDR.

V. RESULTS AND DISCUSSIONS
The performance of the CNN is measured and compared with two benchmarks (A-STAR [16] and AODV [17]) using four metrics: ETE, RL, DDR, and TtT. A-STAR was developed specifically for VANETs in urban areas with M-grids, while AODV was implemented in the real world and has a verified implementation model on OMNET++ (INET) [40].

A. AVERAGE END-TO-END DELAY
VANET routing aims to reduce average End-to-End Delay (ETE) by reducing the traffic transmitted over the network and the number of path hop count. It provides data freshness benefits to increase public safety in VANETs. According to the Figure 7, CNN outperforms AODV especially when the speed of the node increases. This is because a hybrid routing scheme is used to forward network traffic. CNN forwards network transmissions (e.g., MAs) over the RSU infrastructure according to a proactive routing. This results in ETE reduction compared to reactive routing (e.g., AODV), which needs to collect routing information on demand for each transmission. As the Figure shows, the delay in AODV depends on the number and speed of intermediate nodes involved in routing. This is due to the routing delays caused by frequent changes in the network topology. AODV has to repeatedly collect routing information that is constantly changing due to the mobility of the nodes.
CNN outperforms A-STAR especially when the network is dense and the topology changes frequently. This is because CNN uses the clustering technique. As mentioned earlier, clustering has the capacity to offer benefits -mainly innetwork packet aggregation. It results in reduced network transmission which reduces wireless channel access and packet transmission delays. Therefore, the ETE in clusterbased VANETs (e.g., CNN) is lower compared to flat A-STAR.
CNN underperforms A-STAR, when nodes move slowly. A-STAR establishes the shortest path for forwarding network messages between source and target nodes. It reduces the path length and hence reduces the ETE. The path stays interconnected longer time if the nodes move slowly and the network topology changes infrequently. In fact, slow changes in network topology lead to minimal route computation and hence lower ETE. However, A-STAR needs to update and rebuild the shortest paths frequently when the nodes move fast and the network topology changes rapidly. This increases the ETE of A-STAR compared to CNN.

B. ROUTE LENGTH
According to Figure 8, CNN minimizes RL compared to benchmarks when the deployed network is dense and the topology changes rapidly. The proposed protocol uses clustering to support in-network aggregation, reducing the number of route hops. In this way, source nodes avoid establishing flat routes to forward network transmissions to the targets. However, they take advantage of cluster-based routing to forward messages such as DIQ hierarchically. In this way, CHs collect and aggregate network messages (e.g., DIQ) to reduce the number of hops traversed between source and target nodes. Moreover, CNN reduces RL since it uses MA to support loop-free path planning. However, A-STAR uses Dijikstra algorithm to select the shortest path. This reduces RL compared to CNN and AODV when the network topology changes slowly and route re-establishment is minimized.
As the results show, AODV increases the RL compared to A-STAR and CNN. This is because AODV builds the network paths heuristically. Therefore, path hop-count is greatly increased in dense networks, especially when the topology changes rapidly.  According to Figure 9, CNN increases the DDR compared to the benchmarks when the deployed network is dense and/or the node speed increases. This is due to the benefits of using RSU infrastructure to transmit network traffic in the proposed protocol. CNN reduces the number of message failures as it forwards messages over stable links. It also uses clustering to limit concurrent access to wireless channels. As a result, the CMs do not have to forward the data packets directly to target nodes; instead, they transmit the data packets to the CHs for in-network aggregation. This minimizes network transmission, wireless channel access failure, and network packet congestion in the network compared to the benchmarks. DDR is reduced in AODV when the network topology changes rapidly. This is due to the fast mobility of nodes and frequent disconnections. In fact, frequent communication outages reduce packet delivery in AODV.

D. TOTAL TRANSMITTED TRAFFIC
As Figure 10 shows, CNN outperforms AODV and A-STAR, when the speed of mobility increases. CNN uses RSU infrastructure for inter-cluster communication. It is a stable infrastructure that does not require control message to be transmitted to re-establish/update links. Also, CNN uses clustering to reduce the number of messages within a cluster. Clustering supports in-network aggregation and avoids pointless/redundant data transmissions. In addition, CNN forwards the network packets using NDN routing. Thus, intermediate nodes avoid retransmission of redundant and meaningless data interest packets. This results in lower TtT. However, AODV frequently forwards control messages to A-STAR has better performance compared to CNN when nodes are moving slowly. This is because A-STAR avoids clustering. This leads to reduction in network transmission as compared to CNN which transmits multiple control messages to form clusters. However, the TtT increases in A-STAR, when the speed of the nodes increases. This is due to the increased number of network messages that are transmitted frequently to update the failed routes.

VI. CONCLUSION AND FURTHER WORK
CNN proposes a (hybrid) cluster-based routing approach for VANETs. It partitions the network according to the movement region of the vehicular nodes, road ID, and lane number (direction) using Hamming distance technique. However, CNN uses NDN routing to reduce network transmissions (TtT) and save network resources, mainly bandwidth. It uses DSRC and MA communication techniques to forward network traffic throughout the network. MAs are used for inter-cluster links over an RSU infrastructure, while DSRC provides intra-cluster communications. CNN outperforms the benchmarks (i.e., AODV and A-STAR) for ETE, RL, DDR, and TtT when the deployed network is dense and nodes are moving fast.
The performance of CNN still needs to be improved by multi-hop clustering. CNN partitions the network using SAEID POURROOSTAEI ARDAKANI received the Ph.D. degree in computer science from the University of Bath, U.K., in 2015. He currently works with the University of Nottingham Ningbo China (UNNC), as an Assistant Professor in computer science. His research interests include design and implementation of various aspects of networking, including simulation, resource management, synchronization, and routing. He is also interested in big data analysis, blockchain, and the Internet of Things (IoT) applications.