Performance Analysis of Typical Routing Protocols for Flying Ad Hoc Networks Under Different Mobility Models

Highlights

What are the main findings?

• Multiple mobility models with different degrees of randomness are employed to compare and analyze the performance of multiple typical ad hoc routing protocols in FANET scenarios.
• Simulation results show that the choice of mobility model not only affects the performance of a routing protocol but also affects the judgment of the relative performance of routing protocols.

What are the implications of the main findings?

• The conclusion of which routing protocol is better or worse obtained under a specific mobility model is usually not universal and only holds for the specific mobility model used.
• It is necessary to select appropriate routing protocols to adapt to the complex and ever-changing FANET scenarios.

Abstract

Performance of flying ad hoc networks (FANETs) largely depends on the routing protocol used. Applying conventional traditional mobile ad hoc networks (MANETs) routing frameworks to FANETs requires a careful assessment of their compatibility. Crucially, these protocols must be robust enough to handle the volatile link states and rapid topological shifts inherent in high-mobility UAV clusters. Although there have been many works that evaluated and compared the performance of different MANET routing protocols in FANET scenarios through simulation, they ignored the comparative evaluation of various pathfinding schemes across diverse movement patterns. This research addresses this limitation by examining the efficiency of three representative protocols under distinct mobility scenarios using extensive simulations. The findings demonstrate that the selected mobility model influences not only the protocol’s efficiency but also the comparative ranking of different routing protocols. The conclusion of which routing protocol is better or worse obtained under a specific mobility model is usually not universal and only holds for the specific mobility model used. These conclusions will be more helpful for selecting appropriate routing protocols to adapt to the complex and ever-changing UAV network application scenarios.

1. Introduction

With the increasing maturity of unmanned aerial vehicles (UAVs) and the significant reduction in their costs, UAV swarms have been suggested to serve in diverse sectors including military, transportation, and agriculture [1,2]. UAV swarms accomplish complex tasks through the collaboration of individual UAVs, typically outperforming a single UAV in both coverage range and work efficiency. The collaboration of UAVs relies on a flexible, reliable, and efficient communication network to cope with the drastic changes in link quality and network topology caused by the high mobility of UAVs. Recently, the utilization of mobile ad hoc network (MANET) technology in UAV swarms has captured widespread attention from academia and industry, forming a special type of ad hoc network technology called flying ad hoc network (FANET) [3].

Like traditional MANET, the performance of FANET largely depends on the routing protocol used [4]. The routing protocols designed for MANET can be divided into two categories: proactive and reactive. Representatives of proactive routing protocols include the optimized link-state routing (OLSR) protocol and the destination sequenced distance vector (DSDV) protocol. The ad hoc on-demand distance vector (AODV) protocol is a representative of reactive routing protocols. The simplest way to design the FANET routing protocol is to directly port routing protocols designed for traditional MANET to FANET and make targeted improvements [5,6]. The premise of doing so is to ensure that the selected routing protocol is suitable for FANET application scenarios, especially to adapt to the drastic changes in link quality and network topology brought by high-speed UAVs. Therefore, there have been a large number of works that evaluated and compared the performance of different MANET routing protocols in FANET scenarios through simulation.

Existing works for routing protocol evaluation and comparison in FANET scenarios can be divided into three categories: One is to evaluate the performance of a certain routing protocol under a specific mobility model [5,6,7,8,9,10,11,12,13,14,15], another is to evaluate and compare the performance of multiple routing protocols under a specific mobility model [16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35], and the third is to evaluate and analyze the impact of various mobility models on the performance of a specific routing protocol [36,37,38,39,40,41,42,43,44]. Obviously, these studies neglected to compare the effectiveness of various routing schemes under different mobility models. Therefore, it is difficult to determine whether the performance difference between routing protocols under a specific mobility model depends on the mobility model used, resulting in a lack of deeper perspective into how these routing algorithms behave within the complex and ever-changing UAV network application scenarios.

To fill this gap, this paper compares and analyzes how three representative ad hoc routing algorithms behave across three distinct mobility models through comprehensive simulation. The selected or designed mobility models have different degrees of randomness and can represent different UAV network application scenarios. The simulation results indicate that the selection of mobility model impacts the efficiency of routing algorithms. Under diverse mobility models, a given routing protocol will exhibit different packet delivery rate (PDR), throughput, end-to-end delay (E2ED), and overhead. Meanwhile, the choice of mobility model will also affect the judgment of the relative performance of routing protocols. Performance differences between different routing protocols are different under different mobility models. The conclusion of which routing protocol is better or worse obtained under a specific mobility model is usually not universal and only holds for the specific mobility model used.

The original contributions presented in this paper are as follows: 1. Multiple mobility models with different degrees of randomness were selected or designed, with which comparative analysis of typical ad hoc routing protocols was conducted through simulation. 2. Not only the impact of mobility model on the efficiency of a single routing protocol but also their influence on the comparative behavior of multiple routing protocols was analyzed. 3. The conclusions are more helpful for selecting appropriate routing protocols to adapt to the complex and ever-changing UAV network application scenarios. To the authors’ best knowledge, this is the pioneering study that assesses and contrasts the efficiency of standard ad hoc routing protocols in FANET scenarios under multiple mobility models with different randomness.

The remainder of this paper is organized as follows. Section 2 reviews the related literature, succeeded by Section 3 which briefly introduces the OLSR, AODV, and DSDV routing protocols. Section 4 describes the mobility models chosen or designed for this work and the setup of simulation parameters. Section 5 provides the simulation results in detail and analyzes the performance of OLSR, AODV, and DSDV under different mobility models carefully. Section 6 concludes this paper and points out future research directions.

2. Related Works

2.1. Comprehensive Survey of Ad Hoc Routing Protocols

Given the importance of routing protocols to UAV networks, a significant amount of research has emerged over the past decade, including performance evaluation of routing protocols in FANET scenarios and design and optimization of routing protocols tailored for UAV networks. In recent years, a significant number of review papers have emerged to provide a more in-depth analysis of the technical challenges, research advancements, and development trends in FANET routing protocols, aiming to offer a better summary of research in this field [4,45,46,47,48,49,50,51,52]. For example, Arafat et al. provided a comprehensive review of communication strategies for UAV networks, categorizing and surveying them into topology-based, geographical, hierarchical, deterministic, and stochastic frameworks. Subsequently, a qualitative assessment was conducted to compare these protocols based on their core characteristics and efficiency. Then, the routing protocols were compared qualitatively considering their major features and performance. They also discussed the open problems and research directions from the perspective of design and development [45].

Oubbati et al. provided an extensive overview that examines the structural architecture and design constraints of UAV networks. Their work also details diverse mobility models and routing protocols, as well as specialized simulation platforms. Existing important routing protocols dedicated to UAV networks were classified, described, and compared in detail. Moreover, they depicted future research challenges for UAV networks [46]. After presenting typical UAV communication network architectures, Lakew et al. provided an extensive overview of current pathfinding schemes tailored for FANETs. Moreover, they highlighted the key characteristics, advantages and disadvantages, and various mobility models employed for performance evaluation of existing MANET routing protocols. Finally, they highlighted existing open problems and research challenges in UAV networks [4].

Nazib et al. delivered a meticulous review of current routing mechanisms for UAV-assisted vehicular networks. They categorized these protocols into seven distinct groups based on their foundational principles, followed by a qualitative comparison of their design characteristics, operational parameters, and efficiency indicators. Furthermore, the study delved into optimization strategies to offer valuable insights for both the scientific and engineering communities [47]. Abdulhae et al. presented an extensive overview regarding the cluster-based routing schemes, focusing on their advantages, disadvantages, application domains, supported number of nodes, and possible improvements for serving UAV networks. In addition, several UAV networks using cluster-based routing protocols were analyzed in terms of their topology structure, clustering method, scalability, cluster head selection scheme, reliability, performance metrics, and measurements [48].

Cao et al. provided a comprehensive overview of computational intelligence (CL)-based networking and collaboration algorithms from six typical aspects including channel access, network routing, cooperative task assignment, cooperative path planning, cooperative search, and cooperative jamming. To help researchers choose appropriate algorithms to satisfy the requirements of different missions, they classified CI-based algorithms into four categories, namely, heuristic behavior search-based algorithms, policy design-based algorithms, policy learning-based algorithms, and hybrid algorithms [49]. Oubbati et al. discussed the main features of Network Function Virtualization (NFV) and Software-Defined Network (SDN) technologies in FANETS. Moreover, they provided a thorough discussion of the different types, application scenarios, and difficulties associated with UAV-assisted systems. Their study then explored SDN/NFV-enabled aerial platforms, along with several case studies and issues, such as the involvement of UAVs in cellular communications, monitoring, and routing, to name a few [50].

Mansoor et al. examined several state-of-the-art routing protocols specially designed for UAV networks, including their design principles. Although these protocols address various considerations regarding the development of routing protocols for UAV networks, certain obstacles persist. Therefore, they provided a detailed survey of current routing protocols in UAV networks, as well as a discussion of some open issues [51]. Al-Emadi et al. meticulously reviewed the latest network architectures and various routing protocols designed for ad hoc networks. These categories comprise the proactive, reactive, and hybrid routing protocols and their enhanced variants deployed in UAV networks. Moreover, they developed a customized network architecture based on clustering with the purpose of dealing with the problems in existing works [52].

Although these surveys provide a detailed exposition on the technical challenges, research advancements, and development trends of FANET routing protocols, particularly offering in-depth comparisons and analyses of the applicability of various ad hoc routing protocols in FANETs, these comparative analyses are primarily qualitative rather than quantitative. By synthesizing available research focused on evaluating the performance of routing protocols in UAV networks, these surveys analyze the potential and problems of using traditional ad hoc routing protocols in FANET scenarios, starting from the principles of these routing protocols. However, they have not compared the performance of typical routing protocols horizontally through numerical simulation or actual testing, especially the relative performance under different mobility models. Therefore, they cannot provide quantitative guidance for selecting a routing protocol that is suitable for the target scenario.

2.2. Evaluation of Single Routing Protocol Under Single Mobility Model

To directly port traditional ad hoc routing protocols to FANETs, it is necessary to evaluate whether the selected protocol is suitable for FANET application scenarios, as FANETs have significantly different node speeds, movement trajectories, and so on compared to traditional networks. Considering the complexity and cost of actual UAV network testing, existing studies mostly use network simulators such as NS [53], OMNET++ [54], etc., for the performance assessment of routing protocols. These studies simulate and analyze the performance of the target protocol in specific FANET scenarios by selecting or designing appropriate mobility models while considering the typical speed of UAVs. For example, Huang et al. examined how OLSR performs in NS3 simulations by incorporating the Gauss Markov (GM) mobility model. They also used the position of nodes in 3D space to predict the link duration to improve the construction of the routes between nodes. Experimental outcomes demonstrated that the introduced approach enhances the PDR while minimizing the E2ED [5].

Gangopadhyay et al. evaluated the performance of OLSR using NS3 under the GM mobility model. They also proposed an enhanced OLSR protocol which uses the position of nodes for efficient MPR selection. The enhanced OLSR protocol addresses the high mobility of nodes and rapid changes in topology. This is accomplished by considering the residual energy and node degree of UAVs and combining the position information of the UAV, which serves to calculate the link expiration time [6]. Leonov et al. evaluated the packet delivery performance of OLSR in FANET scenarios using the NS2 simulator and the Random Waypoint (RWP) mobility model. By configuring different parameters of OLSR, the change in packet delivery performance and the main factors causing packet loss were analyzed [7].

Cett et al. evaluated OLSR’s performance across varying scenarios of movement velocity and node concentration using NS2. The metrics for evaluation included mean data throughput, PDR, and average packet latency. They found that the execution of the MPR technique effectively curtails unnecessary data relaying throughout the normal message broadcast. Therefore, the number of unnecessary broadcasts can be decreased [8]. Wheeb et al. examined the performance of the original and modified OLSR protocols in UAV ad hoc networks for three search and rescue scenarios. The performance metrics of these routing protocols including PDR, latency, energy consumption, and throughput were simulated using NS3. According to the simulation data, it was observed that the modified OLSR outperforms the original one in the considered measures [9].

Arya et al. developed a complete simulation framework to simulate different network circumstances and situations. A number of performance metrics, such as total packets received, average E2ED, and throughput, served as benchmarks for gauging the efficiency of the AODV routing protocol under the RWP mobility model. The simulation scenarios taken into account different node densities and traffic loads in order to give a comprehensive evaluation of the behavior of the protocol [10]. Ching et al. evaluated the effectiveness of AODV and analyzed the impacts of mobility speed and node density using NS3. Performance metrics used were PDR, throughput, and average E2ED. Numerical analysis demonstrated that mobility speed and node density affect the operational efficiency of AODV in ad hoc networks [11].

Prasad et al. evaluated the performance of the AODV protocol under the free space, two-ray ground reflection, and shadowing fading channel models and also under different scenarios. NS3 was used and the mobility model is RWP. It showed that the AODV protocol could construct the routes in the given scenarios with severe fading [12]. Fiade et al. compared the consequences of blackhole versus flooding intrusions targeting AODV on vehicle networks using several simulators such as SUMO, NS2, and NAM. To measure the quality of service (QoS), throughput, packet loss, and E2ED were monitored. The simulation results demonstrated that peak throughput is achieved in a healthy environment, while a blackhole attack triggers the most severe packet loss. Flooding attacks, conversely, lead to the highest E2ED and residual energy levels [13].

Menaka et al. analyzed the performance of DSR in FANET scenarios using NS2 under the RWP mobility model. Performance metrics, such as route discovery time, route requests sent, and throughout were employed [14]. Gumaste et al. analyzed the DSR protocol by extensive simulations in NS2 simulator under the RWP mobility model with various performance metrics such as PDR, E2ED, routing overhead, and throughput under various scenarios. They found that it is possible to observe the operational behavior of routing protocols within network nodes through digital simulation environments [15].

Table 1. The characteristics of the aforementioned works evaluating the single routing protocol under a single mobility model.

Reference	Protocol	Simulator	Mobility Model
[5]	OLSR	NS3	GM
[6]	OLSR	NS3	GM
[7]	OLSR	NS2	RWP
[8]	OLSR	NS2	RWP
[9]	OLSR	NS3	Real Trajectories
[10]	AODV	N/A	RWP
[11]	AODV	NS3	RWP
[12]	AODV	NS3	RWP
[13]	AODV	SUMO, NS2, NAM	Real Trajectories
[14]	DSR	NS2	RWP
[15]	DSR	NS2	RWP

summarizes the protocols as well as the simulators and mobility models used in the aforementioned works. It can be clearly seen that these works evaluated the performance of a specific ad hoc routing protocol in UAV network scenarios with a specific mobility model using simulators. The main objective of these works is to evaluate the availability of a specific routing protocol in UAV network scenarios, without involving mutual comparison between different routing protocols nor considering the way routing protocols are influenced by varying mobility models.

2.3. Evaluation of Multiple Routing Protocols Under Single Mobility Model

Depending on the neighboring node discovery mechanism, path selection algorithm, etc., used, different ad hoc routing protocols are generally optimal for different network application scenarios. To assess the comparative performance of ad hoc routing protocols in UAV-based environments, numerous investigations have evaluated and compared different routing protocols through simulation. For example, Ferronato et al. investigated how the OLSR, AODV, and Zone Routing Protocol (ZRP) perform using NS2 under real urban vehicular scenarios. The results showed that the OLSR protocol obtains the greatest reduction in overhead and the AODV protocol obtains better performance for the transfer rate. Meanwhile, the hybrid protocol ZRP obtains the best results in terms of the reduction in E2ED and increase in PDR [16]. Tan et al. developed a more realistic simulation environment using OPNET and evaluated and compared the performance of four typical routing protocols, AODV, DSR, OLSR, and GRP. The performance metrics used were network delay, packet received, packet lost, and throughput. The adaptability of routing protocols to specific UAV network settings was confirmed through simulation findings [17].

Khanh et al. considered in detail the requirements of modern 6G UAV communications and then constructed a highly authentic simulation setup via NS3. This allowed for testing the effectiveness of standard AODV, DSR, and OLSR protocols across latency, PDR, and routing overhead dimensions. Such quantitative findings provide key insights for selecting suitable routing protocols in different 6G UAV scenarios [18]. Tuli et al. evaluated four specific table-driven routing solutions (DSR, AODV, GRP, and OLSR) using the OPNET simulator. The assessment parameters primarily consisted of throughput, delay, and packet loss rate [19]. Rani et al. produced a more lifelike simulation context in NS3 to benchmark four classic routing protocols (AODV, DSDV, DSR, and OLSR) through PDR and throughput analysis. The results demonstrate that alternative routing strategies may function effectively in several UAV communication network circumstances [20].

Hussen et al. analyzed the performance of several ad hoc routing protocols in UAV networks using the OPNET 18.6 software. Four routing protocols were chosen for comparison, including AODV, DSR, GRP, and OLSR [21]. Singh et al. conducted simulation analysis on AODV, DSDV, and OLSR routing protocols in UAV networks with NS2, in which simulated trajectories of real UAV scenarios were used [22]. Jorge et al. analyzed the performance of various routing protocols OLSR, DSDV, AODV, DSR, AOMDV, and HWMP through simulations in NS3 based on throughput, PDR, and average delay under different speeds with tests conducted using 25 and 50 UAVs. The results provide insights into the efficiency of these protocols in static and dynamic environments, offering valuable information for the design and optimization of UAV communication networks [23].

Leonov et al. evaluated the performance of AODV alongside OLSR in ad hoc clusters of mini drones using NS2. The protocols were evaluated with several variables such as the density of UAVs, the mobility of nodes, and the communication range. It was demonstrated that these parameters substantially influence how routing protocols perform. The performance metrics considered in this work were lost packets and PDR [24]. Zheng et al. evaluated the performance of OLSR, AODV, and BATMAN-ADV using OPNET under real flying trajectories. The simulation results showed that, compared with OLSR and AODV, BATMAN-ADV is more suitable for application in large-scale, high-dynamic, and high-load UAV network scenes [25]. Diniz et al. carried out experiments in real environments using Raspberry Pi devices to benchmark the operational capabilities of the OLSR, BATMAN-ADV, and Babel protocols. Experimental results demonstrate that OLSR exhibits superior robustness and effectiveness under fluctuating traffic volumes, stemming from its fast route reconfiguration. In high-mobility contexts, Babel outshines others by offering reduced latency, thanks to its nimble topology refresh mechanism. Conversely, despite its efficiency in specific settings, BATMAN-ADV suffers from increased overhead and volatility when subjected to intensive data traffic [26].

Gupta et al. investigated the AODV, DSDV, and OLSR routing schemes within the NS3 simulation environment, utilizing multiple metrics such as PDR, packet drop rate, throughput, latency, and jitter. The effectiveness of these protocols was evaluated in the presence of varying UAV speeds. It carefully explored how different routing protocols function under varied UAV intensities and speeds [27]. Gupta et al. examined the performance of AODV, DSDV, OLSR, and DSR through comprehensive simulation for their suitability in FANETs. By fluctuating the velocity and concentration of nodes, the packet reception rate was documented. The NS3-based analysis revealed that AODV demonstrates superior efficiency compared to its counterparts in the aforementioned contexts [28]. Choudhary et al. evaluated and compared the performance of AODV, OLSR, DSDV, and DSR suitable for UAV-assisted communication networks. Through extensive simulation experiments utilizing NS3, the performance of these protocols was assessed based on network parameters such as delay, throughput, jitter, and packet loss ratio. The results demonstrated that OLSR exhibits optimal routing performance for smaller areas and less nodes. However, DSR performs better when both the network area and the number of nodes increased [29].

Basu et al. employed different wireless channel models to conduct a comparative study of routing schemes in UAV networks. Specifically, their simulation outputs for Lutz and ITU models were used to benchmark AODV against DSDV, both of which are widely used MANET protocols. The simulator was developed using Matlab (R2022a) and the mobility model used was GM [30]. Feng et al. investigated the operational efficiency of AODV, DSDV, and OLSR by analyzing metrics such as packet delivery rate, E2ED, and throughput. It was found that all three protocols perform similarly and effectively in low-speed scenarios. In the medium-speed scenario, OLSR performs better than both AODV and DSDV routing protocols. In the high-speed scenario, without considering data delay, AODV performs better than both DSDV and OLSR routing protocols [31]. Pandey et al. compared the performance of DSR, AODV, OLSR, and ZRP for UAV-enabled telecommunications networks. Actual-world datasets of UAV trajectories were employed in combination with the COMSNET simulator. The data revealed that DSR surpasses AODV and OLSR in terms of latency and packet shipping ratio. However, further evaluation discovered that AODV and OLSR had superior overall performance beneath varying mobility styles [32].

Ahmed et al. investigated how the AODV, OLSR, and DSDV protocols behave using OMNET++. They found that in a high percentage of tests, AODV surpasses both OLSR and DSDV. The overall throughput, PDR, and E2ED are heavily influenced by the network’s node density. OLSR is well-suited for scenarios necessitating low delay, while DSDV delivered the poorest PDR and throughput [33]. Kamble et al. investigated OLSR, AODV, and DSDV in terms of PDR, average throughput, E2ED, and delay jitter using NS3. The 2D random walk mobility model was used. From the simulation results, they found that AODV outperforms OLSR and DSDV in FANETs [34]. Huang et al. evaluated the routing protocols AODV, DSDV, and OLSR using the NS3 simulator and real UAV trajectories. The results showed that DSDV is suitable for simple communication between UAV groups, while AODV is suitable for complex communication between UAV groups [35].

Table 2. The characteristics of the aforementioned works evaluating multiple routing protocols under a single mobility model.

Reference	Protocol	Simulator	Mobility Model
[16]	OLSR, AODV, ZRP	NS2	Real Trajectories
[17]	AODV, DSR, OLSR, GRP	OPNET	RWP
[18]	AODV, DSR, OLSR	NS3	RWP
[19]	AODV, OLSR, DSR, GRP	OPNET	RWP
[20]	AODV, OLSR, DSR, DSDV	NS3	RWP
[21]	AODV, DSR, GRP, OLSR	OPNET	Real Trajectories
[22]	AODV, DSDV, OLSR	NS2	Real Trajectories
[23]	OLSR, DSDV, AODV, DSR, AOMDV, HWMP	NS3	3D GM
[24]	AODV, OLSR	NS2	RWP
[25]	AODV, OLSR, BATMAN	OPNET	Real Trajectories
[26]	OLSR, BATMAN, Babel	N/A	Real Trajectories
[27]	AODV, DSDV, OLSR	NS3	Real Trajectories
[28]	OLSR, DSDV, AODV, DSR	NS3	RWP
[29]	AODV, OLSR, DSR, DSDV	NS3	RWP
[30]	AODV, DSDV	Matlab	GM
[31]	AODV, OLSR, DSDV	NS3	RWP
[32]	DSR, AODV, OLSR, ZRP	COMSNET	Real Trajectories
[33]	AODV, OLSR, DSDV	OMNET++	RWP
[34]	AODV, OLSR, DSDV	NS3	Random Walk
[35]	AODV, OLSR, DSDV	NS3	Real Trajectories

summarizes the protocols as well as the simulators and mobility models used in the aforementioned works. These works clearly demonstrate how diverse routing protocols perform across UAV network settings through the application of various simulators. These evaluations and comparisons used a specific mobility model and did not take into account the impacts of different mobility models on the relative performance of routing protocols, that is, whether the use of different mobility models would affect the judgment of performance differences in routing protocols.

2.4. Evaluation of Single Routing Protocol Under Multiple Mobility Models

Different mobility models use different methods to describe the behavior of mobile nodes. In order to evaluate the impact of mobility models on routing protocol performance, some studies have conducted simulation analysis on the performance of a specific routing protocol under various mobility models. For example, Hameurlaine et al. evaluated the performance of OLSR utilizing NS2 by analyzing how various factors, such as operational area size, mobility models, node speed, and traffic rate, impact the PDR. Several mobility models, including time and random and group-based movements, were employed to simulate realistic UAV behaviors. The findings emphasized the significant impact of area size on OLSR’s performance, providing valuable insights for optimizing FANET deployments and enhancing the reliability and efficiency of these systems in critical applications [36].

Singh et al. thought that the mobility model plays a very essential role in optimizing the performance of routing protocols in UAV networks. Therefore, they applied the OLSR protocol in FANETs and evaluated it under different mobility models to optimize the performance of OLSR in FANETs. The mobility models used were RWP, Reference Point Group Mobility, Manhattan Grid, and Pursue models [37]. Bezziane et al. evaluated the performance of GPSR under various mobility models, including GM, Mass Mobility, and RWP mobility models, to uncover their strengths and limitations. The simulation results revealed notable variations in performance indicators, including E2ED, PDR, and loss rate across various mobility models. These findings offer valuable insights for selecting the most suitable mobility model for specific applications, and improving routing protocols, thereby advancing the reliability and effectiveness of FANETs [38].

By analyzing data packet transmission and node movement, Mowla et al. found that the mobility model plays an important role to enhance network performances. Therefore, five mobility models, including Manhattan Grid, Boundless, Column, Nomadic, and RWP were employed for comparative analysis in FANETs. The performance analysis of different performance metrics (PDR, throughput, packet loss, E2ED, delay jitter, and control overhead) were analyzed using the AODV routing protocol for those models [39]. Wheeb et al. assessed how OLSR behaves across a variety of stochastic and collective mobility models. The study employed two simulation configurations to examine different motion models. NS3 was used to develop the simulation environments. Findings from the simulation indicate that while RWP yields optimal OLSR efficiency in low-velocity settings, the Nomadic model is more compatible with the OLSR protocol when node speeds increase [40].

Nurwarsito et al. analyzed the impact of mobility on the performance of OLSR using NS3 with PDR, E2ED, and routing overhead. PDR showed the highest value on GM with 80 nodes. The best E2ED was seen in GM with 20 nodes. The routing overhead was the best under the random walk model with 20 nodes [41]. Patel et al. characterized the performance of AODV for different mobility scenarios including random movement scenarios, determined movement scenarios, and realistic mobility models. Several parameters such as E2ED, number of received packets, control overhead, PDR, and throughput were evaluated using NS2 [42].

Essa et al. investigated the performance of the AODV protocol with different transmit powers under several mobility models including RWP, Mass Mobility and Linear mobility. They also examined the sensitivity of network metrics to changes in node count, travel speed, and simulation boundaries. Utilizing the OMNeT++ and INET framework, the researchers analyzed AODV’s PDR, end-to-end delay, and throughput across various MANET scenarios [43]. Wahanani et al. integrated the Manhattan mobility framework into DSR to compare it with the existing RWP in NS2. The evaluation spanned multiple network scales, varying node speeds, and distinct simulation terrains. Performance metrics used included PDR and average E2ED. The results showed that the performance of DSR will be affected by different mobility models [44].

Table 3. The characteristics of the aforementioned works evaluating a single routing protocol under multiple mobility models.

Reference	Protocol	Simulator	Mobility Model
[36]	OLSR	NS2	Reference Point Group Mobility, RWP, Random Walk, GM, Pursue
[37]	OLSR	NS2	Reference Point Group Mobility, RWP, Manhattan Grid, Pursue
[38]	GPSR	OMNeT++	RWP, Mass Mobility, GM
[39]	AODV	NS2	Column, Manhattan Grid, Nomadic, RWP, Boundless
[40]	OLSR	NS3	Reference Point Group Mobility, RWP, Random Direction, Nomadic
[41]	OLSR	NS3	GM, Random Walk
[42]	AODV	NS2	Manhattan, RWP, Real Trajectories
[43]	AODV	OMNET++	Manhattan, RWP, Real Trajectories
[44]	DSR	NS2	RWP, Manhattan Grid

summarizes the protocols as well as the simulators and mobility models used in the aforementioned works. It can be clearly seen that these works evaluated and compared the performance of a specific routing protocol under several mobility models using simulators. However, they did not take into account the relative performance of different routing protocols under different mobility models. Therefore, it cannot be determined whether the performance differences between different routing protocols under a specific mobility model depend on the mobility model used.

2.5. Evaluation of Multiple Routing Protocol Under Multiple Mobility Models

Zucchi et al. analyzed the performance of AODV, DSDV, and OLSR using NS3 under the reconnaissance scenario with multiple UAVs. Several mobility models including RWP, GM, Distributed Pheromone Repel (DPR), and Random Destination with Partitioned Zone were used. Only two node speeds were considered including 5 m/s and 10 m/s. From the simulation results, they found that AODV achieves the highest PDR performance with RWP in the reconnaissance scenario with multiple UAVs [55]. Kim et al. provided an analysis of the performance of AODV, OLSR, and DSDV in the paradigm of a FANET using NS3. Two mobility models, RWP and GM, were used to simulate the node trajectory. Meanwhile, five node speeds were considered, with the maximum one as 50 m/s. The routing protocols were analyzed in several different settings, characterized by the adjustment of various critical variables. Based on the experimental findings, it was observed that proactive schemes exhibit superior efficacy in environments with rapidly changing topologies [56].

Table 4. The characteristics of the aforementioned works evaluating multiple routing protocols under multiple mobility models.

Reference	Protocol	Simulator	Mobility Model
[55]	OLSR, AODV, DSDV	NS3	RWP, GM, DPR, RDPZ
[56]	OLSR, AODV, DSDV	NS3	RWP, GM

summarizes the protocols as well as the simulators and mobility models used in the aforementioned works. It can be seen that only a few works have evaluated the performance of multiple protocols under different mobility models. Although these works are similar to this work, there are still significant differences between them. From one perspective, these works are limited to random mobility models and do not take into account more real UAV trajectories. In addition, most prior efforts are confined to lower node speeds and cannot cover the typical speed range of various UAVs. This work considers both random mobility models and more real UAV trajectories, as well as a wider range of speeds that can represent more UAV network application scenarios.

2.6. Differences and Contributions of This Work

From Table 1, Table 2, Table 3 and Table 4, it can be seen that although many works have studied the performance of different routing protocols in FANET scenarios through simulation, only a few of them involve the performance evaluation and comparison of multiple routing protocols under different mobility models. Neither a comparative study of protocols under a fixed mobility model nor an analysis of a single protocol across varying mobility models can isolate the dependency of inter-protocol performance variations on the specific mobility environment employed. Therefore, it is difficult to have a deeper perspective on the performance of various routing protocols in complex and ever-changing UAV network application scenarios.

The main difference between this paper and existing works is that it compares and analyzes the performance of multiple typical routing protocols under various mobility models with different degrees of randomness, thus making up for the shortcomings of existing works. The conclusions drawn can not only be used to evaluate how mobility model selection influences an individual routing protocol’s performance but also to assess its impact on the comparative effectiveness among various protocols. Therefore, it has better guiding significance for selecting appropriate routing protocols to adapt to the complex and ever-changing UAV network application scenarios.

3. Typical Ad Hoc Routing Protocols

There are a lot of routing protocols designed for MANETs. It is impossible to evaluate and compare all these protocols, especially considering that some of them are not frequently used. Figure 1 shows the frequency of occurrence of routing protocols mentioned in Section 2. Protocols that only appear once are not included. It can be seen that the three most frequently used routing protocols are OLSR, AODV, and DSDV. Therefore, these three protocols were chosen for comparison in this paper. Next, brief descriptions of these three protocols will be provided separately. It should be pointed out that although there have been many enhancements and improvements based on OLSR, AODV, and DSDV for FANET in recent years, the main objective of this paper is to analyze the performance differences in a single routing protocol under different mobility models and multiple routing protocols under a single mobility model, in order to provide reference for choosing routing protocols in FANET scenarios. Therefore, we did not consider the enhanced and improved versions of AODV, OLSR, and DSDV but instead used the original ones.

3.1. Brief Description of OLSR Protocol

OLSR is an enhancement of the classical link-state routing protocol, specifically designed for MANETs. The core of this protocol is to reduce the flooding overhead of control messages in traditional routing protocols by introducing the Multipoint Relay (MPR) mechanism. The MPR effectively improves the efficiency of route discovery and maintenance processes and reduces the bandwidth consumption and latency of MANETs. In OLSR, nodes periodically broadcast HELLO and Topology Control (TC) messages for neighbor awareness and route discovery and use this information for route computation and maintenance. Specifically, nodes establish their own local connection library and neighbor information library by exchanging HELLO messages. The format of a HELLO message is shown in Figure 2. The Htime field contains the sending interval for the next HELLO message, the Willingness field specifies a node’s willingness to forward traffic for other nodes, and the Neighbor interface address field contains the interface address of neighbors of the source node that sent this HELLO message. The nodes receiving the TC message use the data within it to establish a local topology information database and then calculate the route from the current node to each destination node in the network based on this information. The format of the TC message is shown in Figure 3. The ANSN field is the sequence number of this TC message; the advertised neighbor main address field contains the main address of the advertised neighbor.

The main workflow of OLSR includes four steps: (1) Neighbor Discovery: Nodes periodically broadcast HELLO messages to perceive the one-hop neighbors within their communication range as well as the one-hop neighbors of their one-hop neighbors (i.e., two-hop neighbors). Meanwhile, whether the state of links is symmetric (i.e., bidirectional communication) or asymmetric (i.e., unidirectional communication) can also be perceived. This information can be used to construct and maintain the neighbor list. (2) MPR Selection: OLSR reduces the number of nodes forwarding link-state information through the MPR mechanism to achieve the goal of reducing control overhead in the network. Each node selects a set of nodes from its one-hop neighbors that can cover all two-hop neighbors as its MPR and forwards the link-state information through the MPR nodes. (3) Topology Announcement: Nodes selected as the MPR by other nodes periodically broadcast TC messages. The TC message contains a list of its MPR selectors, that is, which nodes have selected it as the MPR. (4) Route Computation: Nodes in the network collect TC messages propagated by MPR nodes and use this information to construct a network topology table, which describes the network connectivity. Based on this topology table and local neighbor information, each node uses the shortest path algorithm (Dijkstra’s algorithm) to construct the routes to all reachable destinations in the network and maintain a routing table.

In the traditional OLSR protocol, route computation follows the principle of shortest hop count, which selects the path from source node to destination node that passes through the least number of relay nodes. Due to the dynamic nature of wireless network environments, this strategy that does not consider link quality may result in the selected path containing poor links, thereby affecting the stability of data transmission and overall network performance. To solve this problem, the OLSRd protocol, an open-source implementation of OLSR, introduces the Expected Transmission Count (ETX) as the link evaluation metric, allowing for routing decisions to consider not only topology but also link quality. The calculation of ETX is as follows:

$$ETX(R)={\displaystyle \sum _{\eta \in R}ETX(\eta )}={\displaystyle \frac{1}{\varphi (\eta )\rho (\eta )}}$$

(1)

where R is a route between two nodes, η is a link on this route, ϕ(η) is the forward reception rate of η, and ρ(η) is the backward reception rate of link η. By prioritizing paths with lower ETX values, OLSRd can effectively improve the success rate of data transmission, reduce end-to-end latency, and enhance the overall stability of network. After introducing ETX, the formats of the HELLO and TC message also need to be modified, as shown in Figure 4 and Figure 5.

3.2. Brief Description of AODV Protocol

AODV is an important on-demand routing protocol for MANET. The core idea of this protocol is to establish and maintain the route only when there is a need to transmit data in the network. When the source node needs to send data, if there is no valid route to the destination node now, it will initiate the route discovery process and begin data transmission only after the path is established. If there is an available route in the routing table, the data will be transmitted directly. This approach effectively reduces network overhead and exhibits good adaptability in highly dynamic network environments.

There are three main types of messages used for route establishment in AODV: (1) Route Request (RREQ) Message. This message is only generated when the sending node needs to send data to the destination node and there is no valid path in the routing table. The format of the RREQ message is shown in Figure 6. The Type field has a fixed value of 1, indicating that this is an RREQ message; the J, R, G, D, and U fields provide special routing strategies, such as multicast support, bidirectional routing establishment, etc.; the RREQ ID field is used to identify the RREQ information together with the source node address; the Destination IP Address and Destination Sequence Number fields are used to indicate the destination node and its routing freshness requirements, respectively, to ensure that the selected path is the freshest; and the Originator IP Address and Originator Sequence Number fields are used to, respectively, identify the IP address of the source node initiating the routing request and the latest sequence number of the source node routing, facilitating the establishment of a reverse path in the future. (2) Route Reply (RREP) Message. When the intermediate node receiving the RREQ message discovers that there is routing information to the destination node locally, it generates an RREP response message and sends it to the source node that needs to transmit data through the reverse path. The format of the RREP message is shown in Figure 7. The Type field has a fixed value of 2, indicating that this is an RREP message; the R and A fields are used for local repair of multicast routing and reachability detection of the unicast link, respectively; the Prefix Size field is mainly used for multicast extension in IPv6; the Destination IP Address and Destination Sequence Number fields, respectively, list the IP address of the destination node and the sequence number of the route; and the Originator IP Address field represents the IP address of the source node. (3) Route Error (RERR) Message. When a route failure is detected, the node broadcasts this message to notify other nodes of the route failure. The format of the RERR message is shown in Figure 8. The Type field has a fixed value of 3, indicating that this is an RERR message; the N field indicates whether to immediately delete invalid routes; the DestCount field identifies the number of failed destination nodes included in the RERR message, facilitating the parsing of subsequent fields; and the Unreachable Destination IP Address and Unreachable Destination Sequence Number fields, respectively, represent the IP address and the latest sequence number of the failed node, to ensure consistency in the routing status identification of all nodes.

The main workflow of AODV is as follows: (1) Route Discovery. When the source node needs to send data, it first retrieves the local routing table. If there are no available valid paths in the routing table, it will broadcast an RREQ message, which includes the IP addresses and serial numbers of the source and destination nodes. When the neighboring node receives the RREQ message from the source node, it first checks whether its RREQ ID is duplicated. If it is, it does not process it. If not, it continues to broadcast the RREQ message. When the intermediate node or destination node discovers that it can provide a valid route, it will reply with an RREP message in reverse to the source node to notify it that a valid path has been found. (2) Route Establishment. Intermediate nodes that can provide effective routes or destination nodes will transmit RREP messages in reverse to the source node. Each node passing through will record the forward path (the path to the destination node) after receiving this RREP message. After receiving this RREP message, the source node can obtain the effective path to the destination node. Afterwards, the source node can transmit data packets along this effective route, while the intermediate nodes forward data packets according to the recorded forward path. As long as the route is within its valid period, data packets can be transmitted continuously. (3) Route Maintenance. Due to the highly dynamic topology of MANET, nodes need to broadcast HELLO messages periodically to detect the status of neighbors. When a node does not receive HELLO messages from a neighbor node within a certain period, the neighbor node is considered as failed, that is, the link is disconnected. When a node detects link failure, it will generate an RERR message, which will list the IP addresses and corresponding serial numbers of all unreachable nodes. This RERR message will be sent to all nodes that use this failed route.

3.3. Brief Description of DSDV Protocol

DSDV is a distance vector-based table-driven routing protocol. The core idea of this protocol is to use a sequence number mechanism for nodes to determine the validity of routing information, thereby avoiding route loops. The main workflow of DSDV is as follows: (1) Routing Table Initialization. All nodes will initialize a routing table, where each record includes the destination node address, the routing sequence number, the address of the next hop node, the hop count, and the lifetime. During the initialization phase of the routing table, nodes can only access themselves, and the routing information of other destination nodes is marked as empty or unreachable. (2) Route Update. After the initialization of the routing table is completed, the node will periodically broadcast its routing table to neighboring nodes, as shown in Figure 9. There are two ways to update the routing table. One is a full update with a large fixed cycle, which sends the complete routing table. The other is an incremental update with a shorter fixed cycle, which only includes the changed entries of the routing table since the last routing table update. After receiving the routing table, neighboring nodes continue to broadcast it to other nodes, so that all network nodes can establish complete routes by exchanging the routing table. (3) Route Maintenance. Each entry in the routing table has its corresponding lifetime. When this lifetime is exceeded and no routing update information is received for this entry, it is considered that the routing information has expired. The destination node of the corresponding entry will be marked as unreachable, and an incremental update will be immediately triggered after increasing the sequence number to broadcast this failed routing information to other nodes in the network. In this way, nodes in the network can always maintain an effective routing table, ensuring the reliability of data transmission. (4) Data Forwarding. When nodes need to transmit data, they directly query the local routing table, find the next hop node of the corresponding destination node, and start forwarding the data. The next hop node continues the forwarding operation until the data reaches the destination node. If unreachable nodes are found when querying the routing table, it is necessary to wait for the next scheduled or incremental routing update.

4. Mobility Models and Simulation Setup

4.1. Mobility Models

RWP is one of the earliest and most classic random mobility models. Each node following this model randomly selects a target point within the simulation area and moves along a straight path at a specified or random speed to the target point. After pausing for a random or specified time at the target point (1 s was used in this work), the node selects a new target point again and repeats the above process. During the process of moving from one point to the next, the node speed remains constant. Figure 10 shows a snapshot of node trajectories under the RWP model, indicating that the network topology exhibits strong dynamism. The same height was assumed for all the nodes. To avoid display confusion, each node randomly selected its trajectory color.

GM introduces the idea of Markov chains, where the velocity and direction of nodes exhibit continuous changes over time. The motion of nodes in the GM model has temporal autocorrelation, which is determined by model parameters such as the range of speed and direction variation. Compared with the random walk in RWP model, the GM model can more accurately reflect the movement trajectory with inertia or trend, such as that of a UAV, a vehicle, and other scenarios. In this work, the range of speed variation was set to plus or minus 20% of the target speed, while the range of direction variation was set to plus or minus 45 degrees. Figure 11 shows a snapshot of node trajectories under the GM model. The same height was assumed for all the nodes. To avoid display confusion, each node randomly selected its trajectory color. It can be seen that the node trajectory is smoother compared to the RWP model and the network structure is more stable compared to the RWP model, which makes it more in line with real UAV movement scenarios.

This paper designed a specific mobility model that is closer to real UAV trajectories (hereafter referred to as the HOVER model) for real-world scenarios, with node trajectories shown in Figure 12. Each node moves along a circle around a fixed center with a radius of 60 m. The distance between adjacent centers is 260 m, and the initial phase of each node on its circle is randomly distributed. The same height was assumed for all the nodes. This model not only combines the actual background of UAV application but also effectively simulates the topology connection interruption problem caused by frequent UAV movements, which can more realistically reflect the challenges faced by UAV networks in practical scenarios. However, compared to RWP and GM models, its network structure is more stable. To confirm this, the neighbor change rates (NCRs) of these mobility models were evaluated, as shown in Figure 13. It is obvious that the NCR of the HOVER model is much smaller than those of RWP and GM models.

4.2. Simulation Setup

NS3 was used to simulate the selected routing protocols under different mobility models.

Table 5. The main simulation parameters.

Parameter	Value
Simulator	NS3 (Version 3.33)
Channel Model	Ideal Channel Model
Simulation Area	2000 m × 2000 m
Simulation Time	150 s
Communication Range	300 m
Number of Nodes	49
PHY/MAC Protocol	IEEE 802.11b
Node Speed	20, 40, 60, 80, 100, 120 m/s
Traffic Type	CBR
Data Packet Length	64 Byte
Data Packet Rate	2048 bps
Communication Pairs	9

shows the main simulation parameters. The simulation time was 150 s. The ideal channel model was used, and the communication range was set to 300 m. Six node speeds were considered including 20 m/s, 40 m/s, 60 m/s, 80 m/s, 100 m/s, and 120 m/s, which basically covers the speed of various UAVs from multi-rotor to fixed wing. For each speed, 10 simulations were performed, and the results were averaged. The communication pairs were randomly selected before each simulation and then kept fixed during the simulation. This paper mainly focuses on the mutual influence between routing protocols and mobility models in the context of FANETs, rather than state-of-the-art UAV communications. Therefore, the IEEE 802.11b used by many existing works were employed. The main parameters of OLSR, AODV, and DSDV adopted in NS3 are shown in

Table 6. The main parameters of OLSR.

Parameter	Value
Hello Interval	2 s
Neighbor Hold Time	3 × Hello Interval
TC Interval	5 s
TC Hold Time	3 × TC Interval
MID Interval	5 s
MID Hold Time	3 × Hello Interval

,

Table 7. The main parameters of AODV.

Parameter	Value
RREQ TTLStart	1
RREQ TTLIncrement	2
RREQ TTLThreshold	7
RREQ Retries	10
ActiveRouteTimeout	3 s
MyRouteTimeout	11.2 s
DeletePeriod	15 s
PathDiscoveryTime	5.6 s
RERRRateLimit	10
TimeoutBuffer	2 s
NodeTraversalTime	40 ms
RERRRateLimit	10

and

Table 8. The main parameters of DSDV.

Parameter	Value
PeriodicUpdateInterval	15 s
SettlingTime	15 s
Holdtimes	3
MaxQueueLen	500
MaxQueueTime	30 s

, respectively.

5. Performance Evaluation and Analysis

5.1. Performance Metrics

Multiple metrics can be used to evaluate the performance of network protocols, with the most commonly used and widely recognized being end-to-end PDR, E2ED, throughput, and routing overhead. PDR is mainly used to evaluate the transmission efficiency of data packets between source and destination nodes. It is defined as the ratio of the number of successfully arrived data packets at the destination node to the total number of data packets sent by the source node, as follows:

$$PDR={\displaystyle \frac{Nu{m}_{receiv}}{Nu{m}_{send}}}$$

(2)

where Num_receiv denotes the number of successfully arrived data packets at the destination node; Num_send denotes the total number of data packets sent by the source node. PDR could reflect the reliability and stability of the network. High PDR presents good data transmission performance, while low PDR indicates the existence of some network problems, such as queue overflow, packet collision, route failure or timeout, etc.

E2ED represents the time it takes for data packets sent by the source node to be successfully received by the destination node. This metric is influenced by various factors, including the transmission distance of data packets in the network, the degree of network congestion, and the processing delay of routers. E2ED is usually measured in milliseconds, and its calculation is as follows:

$$Delay=D_{Proc}+D_{Queu}+D_{Trans}$$

(3)

where D_Proc represents the processing time of data packets at various nodes, D_Queu represents the time it takes for data packets to queue and wait for forwarding at intermediate nodes, and D_Trans represents the time it takes for data packets to be sent from nodes to the link.

Throughput represents the amount of data successfully transmitted per unit of time. It is defined as the ratio of the number of bits of the successfully arrived data packets at destination nodes and the simulation time. It not only depends on the packet size and rate of each source node but also on the number of source nodes. Routing overhead includes the control packets that are needed to construct and maintain the routes. It is defined as the number of the control packets during the simulation.

5.2. Single Protocol Under Multiple Mobility Models

Figure 14a shows the PDR of the OLSR protocol under three mobility models. As the node speed increases, the PDR decreases continuously. The PDR under the HOVER model is the highest, because this model has the most stable network topology. Meanwhile, the performance is similar under both the GM model and the RWP model. Under the RWP model, the random movement of nodes improves network connectivity, resulting in a slightly higher PDR overall compared to the GM model. Figure 14b shows the E2ED of the OLSR protocol under three mobility models. As the node speed increases, the E2ED increases continuously. Due to the unstable network structure and frequent link interruptions under the RWP model, OLSR requires more time to update routing information, resulting in the highest E2ED under the RWP model. Thanks to the more stable network topology, network dynamics of GM and HOVER models are not as high as the RWP model, and the links are more stable. OLSR does not require frequent routing updates, resulting in lower E2EDs. Figure 14c shows the throughput of the OLSR protocol under three mobility models. The trend of throughput changing with node speed is basically consistent with PDR. Figure 14d shows the overhead of the OLSR protocol under three mobility models. The overhead under the HOVER model is basically unchanged, while the overhead under the other mobility models slightly changes with node speed. This is mainly due to changes in the number of TC messages caused by the more random topology changes in the RWP and GM models.

Figure 15a shows the PDR of the AODV protocol under three mobility models. As the node speed increases, the PDR also decreases continuously. PDR under the HOVER model is still the highest, thanks to its most stable network topology. The random movement of nodes improves network connectivity under the RWP model, so the path is easier to establish when data transmission is required. Therefore, a higher PDR is achieved than that under the GM model. Figure 15b shows the E2ED of the AODV protocol under three mobility models. E2ED under the HOVER model is the lowest, also thanks to its most stable network topology. As the node speed increases, the E2ED first decreases and then increases. As the network topology is unstable under the RWP model, more time is required to find available routes when data transmission is needed, so its E2ED generally increases with the node speed. Figure 15c shows the throughput of the AODV protocol under three mobility models. As expected, the trend of throughput changing with node speed is basically consistent with PDR. Figure 15d shows the overhead of the AODV protocol under three mobility models. The overhead under the HOVER model is the lowest, followed by that under the GM model, while the overhead under the RWP model is the highest.

Figure 16a shows the PDR of the DSDV protocol under three mobility models. It can be seen that as the node speed increases, the PDR also decreases continuously. The PDR under the HOVER model is still the highest, also thanks to its most stable network topology. As the nodes move randomly under the RWP model, when the node speed is slow, the link is relatively stable, and its routing availability increases, so the PDR is also high. However, as the node speed increases, DSDV cannot quickly establish available paths, resulting in a rapid decline in its PDR. Under the GM model, nodes move smoothly until they reach the boundary of the simulation area before bouncing back. This results in DSDV being unable to establish effective routes; hence, its PDR is the lowest. Figure 16b shows the E2ED of the DSDV protocol under three mobility models. As the node speed increases, the E2ED shows an increasing trend. Under the HOVER model, due to the stable network structure and high routing effectiveness, E2ED is the lowest. Under the RWP model, the high randomness of node movement leads to unstable routes. As the node speed increases, the insufficient perception of link changes leads to many packet retransmissions, path reselection, and message queuing, resulting in rapid growth of E2ED. Figure 16c shows the throughput of the DSDV protocol under three mobility models. Like OLSR and AODV, the trend of throughput changing with node speed is basically consistent with PDR. Figure 16d shows the overhead of the DSDV protocol under three mobility models. Like AODV, the overhead under the HOVER model is the lowest, followed by that under the GM model, while the overhead under the RWP model is still the highest.

5.3. Multiple Protocols Under Single Mobility Model

Figure 17a,b show the PDR and E2ED of three routing protocols under the RWP model. As the RWP model has the highest randomness, the changes in links and topology are the most drastic. At this point, thanks to the on-demand establishment of routes, AODV has the highest PDR, but also the highest E2ED, as it requires waiting for the completion of route establishment. OLSR considers path quality (in the form of ETX), resulting in a higher PDR. However, considering path quality makes it easier to choose longer paths, resulting in a higher E2ED. DSDV is based on a distance vector and will select paths that are shorter than OLSR, resulting in a lower E2ED. However, it does not consider path quality and is more prone to packet loss. Figure 17c,d show the throughput and overhead of three routing protocols under the RWP model. Like the results of the single protocol under three mobility models, the trend of throughput changing with node speed is basically consistent with PDR for all three routing protocols. In terms of overhead, DSDV is the worst as expected, because flooding is used when constructing the routes. The reason that AODV is worse than OLSR is because the frequent topology changes under the RWP model generates a large number of RERR packets for AODV.

Figure 18a shows the PDR of three routing protocols under the GM model. Under the GM model, the trajectory of node movement is relatively smooth, and the node will not bounce back until it hits the boundary of the simulation area. Thanks to regular updates of routing, OLSR can directly utilize stored routes for data transmission when needed, resulting in a higher PDR. AODV may experience path failure due to node motion during data transmission after route establishment, resulting in a higher packet loss rate. Therefore, its PDR is lower compared to OLSR. Figure 18b shows the E2ED of three routing protocols under the GM model. AODV has the highest E2ED due to the need to wait for the route establishment process. Thanks to regular updates of routing, OLSR can utilize stored routes for data transmission when needed, resulting in a lower E2ED. The slow convergence speed of DSDV makes it easier for the path of data packets to fail during transmission, resulting in E2ED being higher than OLSR. Figure 18c,d show the throughput and overhead of three routing protocols under the GM model. As expected, the trend of throughput changing with node speed is basically consistent with a PDR for all three routing protocols. Like that under the RWP model, the overhead of DSDV is the highest, while OLSR is the best. The reason that AODV has higher overhead than OLSR also because the inherent randomness of the GM model generates a large number of RERR packets for AODV.

Figure 19a shows the PDR of three routing protocols under the HOVER model. Since AODV establishes routes on demand and only starts route establishment when data transmission is required, its PDR is the highest. When establishing routes, OLSR considers path quality, while DSDV only considers the shortest path, so OLSR has a higher PDR than DSDV. Figure 19b shows the E2ED of three routing protocols under the HOVER model. As the node speed increases, the network topology changes dramatically. The E2ED of AODV first decreases and then increases. On the contrary, the E2ED of the two proactive protocols, OLSR and DSDV, increase with the increase in node speed. OLSR considers link quality and chooses routes with larger hop counts, resulting in higher latency than DSDV. Figure 19c,d show the throughput and overhead of three routing protocols under the HOVER model. Like those under the RWP and GM models, the trend of throughput changing with node speed is basically consistent with the PDR for all three routing protocols. However, the obvious difference appears in terms of overhead, for which AODV defeats OLSR. This is consistent with the performance of these two protocols in traditional MANETs. This is because the topology under the HOVER model is much more stable than those under the RWP and GM models. Therefore, it is closer to the degree of topology change in traditional MANETs. As the node speed increases, the topology changes more dramatically, and the overhead of AODV begins to slowly increase until it exceeds that of OLSR.

6. Real-World Flight Experiments

To evaluate the performance differences in these routing protocols in real-world scenarios, real-world flight experiments were designed using one ground station and three UAV nodes. The open-source F450 UAV was employed, as shown in Figure 20a. One UAV was the source node and the other two were the relay nodes. Three UAV nodes were deployed at the same height, with the source node being stationary. Two relay nodes used a mobility model similar to HOVER. That is to say, they moved in a circular motion around their respective centers. One of them is clockwise, and the other is counterclockwise, as shown in Figure 20b. Considering the limitations of the F450 UAV platform, the flight speed of the relay nodes was set to 10 m/s. Two relay nodes were simultaneously within the communication range of the ground station and the source node, but direct communication between the ground station and the source node is not possible. Therefore, data packets generated by the source node can only be forwarded through relay nodes, and the specific relay node used depends on its location. The generation of data packets uses the same settings as simulation. For each routing protocol, 10 tests were performed, with each test lasting for 150 s.

Figure 21 shows the PDR, E2ED, and throughput of real-world flight experiments. As can be seen, OLSR achieved the highest PDR and throughput, followed by AODV, while DSDV had the lowest PDR and throughput, as shown in Figure 21a,c. This is because the topology structure of real-world flight experiments is relatively stable, so the routes can be maintained for a long time. As expected, AODV has the highest E2ED, as shown in Figure 21b, which is the inherent flaw of on-demand routing protocols. Compared with the simulation results under the HOVER model, it can be found that although the network scale of real-world flight experiments is much smaller, its delivery performance is actually worse. This is because the real channel model used in the simulation overestimates the performance of routing protocols. However, the relative performance relationship between protocols is consistent.

7. Conclusions and Future Works

The collaboration of UAVs in swarms relies on a flexible, reliable, and efficient communication network to cope with the drastic changes in link quality and network topology brought by their high mobility. As the network performance largely depends on the routing protocol used, designing feasible routing protocols is essential for UAV networks. The simplest way is to directly port the routing protocols designed for traditional MANETs, as long as the selected routing protocol is suitable for FANET application scenarios. Although there have been a large number of works that evaluated and compared the performance of different MANET routing protocols in the FANET scenarios through simulation, these works overlooked the performance evaluation and comparison of multiple routing protocols under different mobility models. Neither a comparative study of protocols under a fixed mobility model nor an analysis of a single protocol across varying mobility models can isolate the dependency of inter-protocol performance variations on the specific scenario employed. Therefore, it is difficult to have a deeper perspective on the performance of various routing protocols in complex and ever-changing UAV network application scenarios.

This paper fills this gap by analyzing the performance of three typical routing protocols under three different mobility models through comprehensive simulation. In addition to the commonly used RWP and GM models, a specific mobility model that is closer to real UAV trajectories is designed. The selected or designed mobility models have different degrees of randomness and can represent rather different UAV network application scenarios. The simulation results indicate that the choice of mobility model will affect the performance of the routing protocol. Under different mobility models, the same routing protocol will exhibit a different PDR, throughput, E2ED, and overhead. Meanwhile, the choice of mobility model will also affect the judgment of the relative performance of routing protocols. The performance differences between different routing protocols are different under different mobility models. The conclusion of which routing protocol is better or worse obtained under a specific mobility model is usually not universal and only holds for the specific mobility model used.

The conclusions of this work can not only be used to evaluate how mobility model selection influences the performance of individual routing protocols but also to assess its impact on the ranking of various routing protocols. Therefore, it has better guiding significance for selecting appropriate routing protocols to adapt to the complex and ever-changing UAV network application scenarios. For example, simulation results under non-real mobility models cannot be used as performance evaluations under real trajectories. However, it can be seen from the simulation results, the mobility model and node speed are not the only two factors that affect the performance of routing protocols. Other factors such as coverage area, number of nodes, communication distance, and transmission requirements can also affect the performance of routing protocols. How to evaluate the comprehensive impact of so many influencing factors on the performance of routing protocols and find the most important influencing factors in specific UAV network scenarios is one of the important research directions in the future.