Dynamic Cluster Size Optimization in Hybrid Cellular-Vehicular Networks Julian Garbiso *† , Ada Diaconescu * , Marceau Coupechoux * , Bertrand Leroy † * Telecom ParisTech, CNRS, LTCI - Paris, France {name . surname}@telecom-paristech.fr † Institut Vedecom - Versailles, France {name . surname}@vedecom.fr Abstract—Ground-breaking innovations in transport, such as autonomous vehicles, the European Local Dynamic Map (LDM) and related on-line services heavily depend on reliable vehicular connectivity. In the most likely scenario, hybrid vehicular networks will use the IEEE 802.11p protocol for vehicle-to- vehicle (V2V) communication, and the cellular network (e.g. LTE or 5G) as a gateway to remote servers. Both technologies have their own flaws as IEEE 802.11p available bandwidth drops quickly with the accretion of vehicles in the vicinity, and intensive cellular usage can be costly. In this context we intend to minimize this cost while ensuring system reliability. Clustering vehicles can significantly reduce cellular network usage when the Cluster Head (CH) is the only node that communicates with the cellular Base Station (BS) and also performs data aggregation. Other nodes communicate with the CH using multi-hop forwarding over IEEE 802.11p. There is however a tradeoff in cluster design. On the positive side, large clusters lead to high aggregation levels and thus low usage of the uplink cellular resources. On the negative side, large clusters also increase packet losses in the IEEE 802.11p network affecting communication reliability. In this paper, we study the impact of the number of communication hops on the average size of formed clusters, on data compression, and on IEEE 802.11p packet losses in various vehicle densities. We present a new clustering algorithm which delegates the CH election to the cellular BS, significantly improving the cluster formation compared to CH self-election algorithms. We propose a dynamic clustering approach that adapts the cluster size to the vehicle density and optimize data compression under the constraint of an acceptable IEEE 802.11p packet loss. I. I NTRODUCTION Vehicular connectivity is the cornerstone of future Intel- ligent Transportation Systems (ITS), providing a substantial leap forward in road security, and paving the way for inno- vative services, from driving assistance to information and entertainment. However, the network protocol designed for connecting vehicles, IEEE 802.11p easily reaches congestion points [1], while the volume of data to be transmitted is important, even for basic services. The European standard ETSI ITS-G5, built over this protocol, establishes that every vehicle has to broadcast a Cooperative Awareness Message (CAM) at intervals as short as 100 ms, and Decentral- ized Environmental Notification Messages (DENM), which are sent upon a particular incident and usually require re- broadcasting. To this day, the initiative of some major players in the automotive industry is to equip their cars with cellular network interfaces, while keeping the use of this resource as low as possible for high costs reasons. The most likely scenario seems to be that vehicles will have two network interfaces: cellular network (e.g. LTE or 5G) and IEEE 802.11p [2][3]. Messages for certain services, such as security, will be managed locally, while others will require access to remote servers. As the usage of the cellular network resources has a significant cost [4], it is essential that the information be aggregated before being uploaded whenever this is possible. The most widely adopted approach for data collection and aggregation in vehicular networks (as well as in other environments such as wireless sensor networks) is clustering. This technique consists in creating groups of communicating devices in a geographical vicinity, where a member of the group is designated as a Cluster Head (CH) whose functions may vary depending on the application. In our case, this entity is responsible for gathering and compressing data received through the vehicle-to-vehicle (V2V) radio interface, namely IEEE 802.11p, and upload the aggregated informa- tion through the cellular network. In multi-hop clustering algorithms, the cluster size can be increased, extending the CH’s coverage area by allowing packet forwarding between cluster members. However, a trade-off must be made when designing a clustering algorithm with respect to its average size. On one hand, large clusters significantly reduce the cellular network access (thus, costs are minimized), yet packet loss in the V2V interface increases dramatically. On the other hand, small clusters ensure low packet losses in the IEEE 802.11p network, yet offer little or no compression, and hence raise the costs associated to the usage of the cellular network. Our main contributions in this paper are: • We show the impact of the maximum number of hops between a specific vehicle and its CH on the average cluster size, the data aggregation performance and the packet loss in the IEEE 802.11p network. At low vehicle densities, the number of hops should be maximized in order to increase data compression. At high vehicle densities, we observe that increasing the number of hops leads to unacceptable packet losses. Because of such unreliable communications, cluster sizes also decrease and affect data compression.