2025 IEEE/CIC International Conference on Communications in China (ICCC), Shanghai, China, 2025, DOI: 10.1109/ICCC65529.2025.11149335.
Deterministic Task Scheduling in In-Vehicle Networks for Software-Defined Vehicles Keyvan Aghababaiyan, Baldomero Coll-Perales, Luca Lusvarghi, Javier Gozalvez Uwicore laboratory, Universidad Miguel Hernandez de Elche, Elche (Alicante), Spain [email protected], [email protected], [email protected], [email protected] Abstract— Modern vehicles are embedding increasing levels of automation, connectivity, and intelligence, which require advanced in-vehicle networks and computational platforms to support the dependability and deterministic requirements of critical in-vehicle functions. To this end, the automotive industry is shifting towards software-defined vehicles (SDVs) and zonal E/E architectures with centralized computing nodes. Realizing the full potential of these new architectures requires an efficient management of the in-vehicle’s computational workload. In this context, this paper introduces a deterministic task scheduling approach for in-vehicle networks (IVN), and demonstrates that it can better guarantee deterministic service levels than alternative approaches based on the shortest path or the objective to minimize task execution time. Our evaluation also demonstrates that a deterministic task scheduling can satisfactorily support increasing in-vehicle computational workloads and tasks, and achieve a more balanced workload and resource utilization across the IVN. These gains are validated across a variety of IVN topologies, and in hybrid wireless-wired IVN implementations, where a gradual introduction of wireless offers increased in-vehicle connectivity diversity. Keywords—Task scheduling, deterministic, SDV, softwaredefined vehicle, in-vehicle networks, IVN, zonal architecture.
I.
INTRODUCTION
The rapid automotive evolution has led to an increasing demand for complex and diverse in-vehicle functions, driven by industry trends such as connectivity, electrification, automated driving, and smart mobility. This evolution needs increasingly sophisticated in-vehicle computing capabilities to support features and services with stringent reliability and deterministic service level requirements. Additionally, the computational and operational demands of next-generation automotive systems require evolving from traditional invehicle electrical/electronic (E/E) architectures with distributed processing and domain-specific controllers [1]. To address these growing demands, the automotive industry has been shifting towards software-defined vehicles (SDVs), which enable more flexible, configurable, scalable, and upgradable in-vehicle functionalities [2]. The transition to SDVs necessitates significant changes in the in-vehicle network (IVN) and E/E architecture. Traditional architectures, which rely on numerous independent electronic control units (ECUs) to manage sensors and actuators for specific vehicle subsystems, are giving way to zonal IVN architectures with centralized computing [3]. In this new architecture, the vehicle’s This work has been partially funded by the European Commission Horizon Europe SNS JU 6G-SHINE (GA 101095738) project, and by MCIN/AEI/10.13039/501100011033 (PID2020-115576RB-I00, PID2023150308OB-I00).
THIS IS AN AUTHOR-CREATED POSTPRINT VERSION
computational workload is handled by high-performance central compute platforms and zonal controllers, enabling improved coordination between vehicle functions and data fusion, thereby facilitating higher levels of automation and intelligence in decision-making. The zonal IVN architecture with centralized computing also incorporates redundant, diverse, and fault-tolerant elements and links, which are essential for achieving the functional safety levels required in critical systems. However, the efficiency of this IVN architecture depends on the effective management and scheduling of computational tasks, making task scheduling a critical requirement. As vehicle functionalities expand, optimizing workload distribution across processing units is essential to maintain deterministic responsiveness and dependability of critical functions. Traditional static scheduling approaches, where tasks are allocated to predefined computing units or ECUs, may face challenges in this dynamic and demanding environment. The automotive industry advocates for the use of global scheduling with adaptive task allocation approaches to effectively balance workloads across available communication links and computational resources without compromising the execution of critical vehicle functions [1]. In this context, this paper advances the state of the art with the proposal of a deterministic task scheduling approach for zonal in-vehicle E/E architectures with centralized computing. The study demonstrates that a deterministic task scheduling can better guarantee the deterministic service levels of critical in-vehicle functions than alternative state-ofthe-art approaches that schedule tasks based on the shortest path [4], [5] or the objective to minimize task execution time [6]. Our evaluation also demonstrates that a deterministic task scheduling can satisfactorily support increasing in-vehicle computational workloads and tasks, and achieve a more balanced workload and resource utilization across the zonal in-vehicle network. We demonstrate that the benefits achieved with a deterministic task scheduling approach are valid across a variety of IVN topologies, ranging from traditional treebased topologies to mesh topologies with centralized computing based on realistic case studies [4][7]. These benefits are also maintained considering hybrid wirelesswired IVN implementations, where a gradual introduction of wireless offers increased connectivity diversity for linking sensors and actuators to computing units. The results demonstrate that the deterministic task scheduling approach can better adapt to varying operating conditions while enabling efficient resource utilization, thereby preventing resource saturation and enhancing scalability. The remainder of this paper is organized as follows. Section II presents a comprehensive review of state-of-the-art in-vehicle networks and topologies. Section III introduces the
2025 IEEE/CIC International Conference on Communications in China (ICCC), Shanghai, China, 2025, DOI: 10.1109/ICCC65529.2025.11149335. system model, describing the overall architecture and the key HPCU Sensors/Actuators ECU operating conditions for task scheduling within the IVN. Section IV outlines the proposed deterministic task scheduling strategy and state-of-the-art approaches, such as minimum execution time and shortest path task scheduling. Section V details the simulation setup, including the evaluation parameters and scenarios. Section VI evaluates the proposed deterministic task scheduling strategy and compares it with state-of-the-art mechanisms. Finally, Section VII a) Tree b) Basic mesh concludes the paper. II.
STATE-OF-THE-ART IN-VEHICLE NETWORKS AND TOPOLOGIES
Traditional IVN architectures incorporate one ECU for each in-vehicle electronic function with a very specific control task, and a direct interconnection among them. This approach requires new ECUs and interconnections when new sensors or actuators are required. The significant increase in electronic functions has triggered an evolution of IVN architectures for better scalability. A first evolution has been domain-based IVN architectures with several functional domains (infotainment, powertrain, assisted driving, etc.) managed by domain-specific networking technologies and controllers. Despite its benefits, domain-based IVN architectures experience challenges for developing automotive applications that require cross-domain functionality, a need that is growing with vehicle softwarization and the gradual introduction of autonomous driving functions. Zonal IVN architectures have emerged as an alternative to enhance efficiency as vehicle complexity and functionality increase. Zonal IVN architectures group embedded devices and electronics based on physical location rather than logically or per domain. The zonal IVN architecture locally connects sensors and actuators to zonal controllers or ECUs that are physically and strategically distributed through the vehicle. These zonal controllers rely on a high-speed backbone network to connect to each other and to the vehicle’s high-performance central computing platform with advanced processing capabilities. A trend in the evolution of zonal IVN architectures is vehiclecentralized computing [3], and the possibility that sensors/actuators may bypass the zonal ECUs and connect directly to the vehicle’s central computing platform. In line with the transition to zonal IVN architectures with centralized computing, this study analyzes four in-vehicle network topologies, depicted in Fig. 1, which are based on realistic case studies from [4][7]. The topologies share a common structure, defining four in-vehicle zones that represent the front-left, front-right, rear-left and rear-right areas of the vehicle. Each zone includes a zone ECU in addition to the sensors and actuators located within that area. The topologies also incorporate a central High-Performance Computing Unit (HPCU). However, they differ in their degree of connectivity. Fig. 1.a represents a conventional tree IVN topology, where sensors and actuators are connected to their respective zonal ECUs, and the zonal ECUs are connected to the central HPCU. Fig. 1.b represents a basic mesh IVN topology which introduces a connectivity backbone between zone ECUs. The ECUs positioned at opposite ends along the diameter are interconnected via the HPCU. Fig. 1.c follows the IVN topology of the Orion Crew Exploration Vehicle (CEV) as utilized in [4], and we refer to it as cross-zone mesh.
THIS IS AN AUTHOR-CREATED POSTPRINT VERSION
c) Cross-zone mesh
d) Centralized mesh
Fig. 1. In-vehicle network topologies.
This topology adds cross-zone connections, providing redundant links between sensors/actuators and nearby zone ECUs. These connections serve as alternative paths for accessing nearby ECUs and the HPCU. Without loss of generality, we consider these cross-zone connections link front and rear sensors/actuators to the zone ECUs located in the opposite area (i.e., left to right and right to left). Finally, Fig. 1.d depicts a centralized mesh topology, which introduces direct links between the sensors/actuators and the HPCU to the cross-zone mesh topology. This provides an alternative path to access the HPCU. III.
SYSTEM MODEL
The system consists of automotive in-vehicle functions that generate tasks 𝑓𝑛 , where 𝑛𝜖{1, … , 𝑁}). Each task 𝑓𝑛 is characterized by the tuple (𝑜𝑛 , 𝑐𝑛 , 𝑠𝑛 , 𝑠′𝑛 , 𝑡𝑛 , 𝑇𝑛𝑚𝑎𝑥 ) where: 𝑜𝑛 denotes the originating point where tasks can be originated from sensors and actuators (𝑆𝑁𝐴𝑠𝑚 ), zone ECUs (𝑧𝐸𝐶𝑈𝑚 ), and the HPCU (𝐻) , where 𝑚𝜖{1, … 𝑀} represents the invehicle area or zone (M=4) [1], and 𝑠𝑚 𝜖{1, … 𝑆𝑚 } is the number of sensors/actuators in the zone 𝑚. 𝑐𝑛 denotes the computing demand of the task, 𝑠𝑛 represents the task’s size, 𝑠′𝑛 is the size of the task after processing, 𝑡𝑛 indicates the task generation time, and 𝑇𝑛𝑚𝑎𝑥 defines the task deadline for processing. The processing of tasks is restricted to 𝑧𝐸𝐶𝑈𝑚 and 𝐻. We consider that task scheduling schemes (described in Section IV) dynamically assign tasks to computing units within the IVN. When a task 𝑓𝑛 is executed on a processing unit different from where it was generated, the processed result with size 𝑠′𝑛 must be transmitted back to its source unit. The 𝑇𝑛𝑚𝑎𝑥 of task 𝑓𝑛 accounts then for the transmission time to move the task to the assigned processing unit, the processing duration, and the time required to transmit the processed result back to its source unit. The in-vehicle computing units have different processing speed, denoted as 𝑃𝑥 measured in GHz, and a maximum processing capacity 𝐶𝑥𝑚𝑎𝑥 over a time period T with 𝐶𝑥𝑚𝑎𝑥 = 𝑃𝑥 ∙ 𝑇, where 𝑥 𝜖{𝑧, ℎ} refers to the type of processing unit, i.e., {𝑧𝐸𝐶𝑈𝑚 , 𝐻}, respectively. The time required to process a task 𝑓𝑛 on a computing unit 𝑥𝜖{𝑧, ℎ} is given by: 𝑐𝑛 𝑡𝑝𝑛 = . (1) 𝑃𝑥
2025 IEEE/CIC International Conference on Communications in China (ICCC), Shanghai, China, 2025, DOI: 10.1109/ICCC65529.2025.11149335. We consider that the IVN topologies depicted in Fig. 1 can IV. DETERMINISTIC TASK SCHEDULING be fully wired or hybrid wireless-wired. In both cases, we This study proposes a deterministic task scheduling represent with 𝐸 the set of links between the set of the IVN scheme for IVNs. The Deterministic scheme prioritizes elements (𝑆𝑁𝐴𝑠𝑚 , 𝑧𝐸𝐶𝑈𝑚 , 𝐻 ). 𝑑𝑖𝑗 represents the distance maximizing the number of tasks completed within their between nodes 𝑖 and 𝑗 in the IVN. We consider that nodes that deadlines (i.e., 𝑇𝑛 ≤ 𝑇𝑛𝑚𝑎𝑥 ), making it particularly suitable can be reached directly are closer than those that require for guaranteeing the timely execution of critical vehicle passing through other nodes. functions within strict bounded time constraints. The scheme For wired links, we consider Ethernet-like connections dynamically adjusts task completion times based on varying with a data rate 𝑅𝑤 and no transmissions errors (i.e., the deadlines, enabling flexible management and balanced reliability 𝜌 is equal to 1), ensuring reliable and consistent workload distribution across the IVN. The objective function data transmission. The transmission time over a wired link is formulated as: 𝑤 𝜖 𝐸 can then be computed as: 𝑇𝑛 (7) 𝑠𝑛 𝑚𝑖𝑛 ∑ 𝛽 ( 𝑚𝑎𝑥 ) , 𝑡𝑐𝑛 = . (2) 𝑇𝑛 𝑛 𝑅𝑊 where 𝛽(𝜉) is a penalty function defined as: In the hybrid wireless-wired IVN scenarios, we restrict (8) 1 − ∏ 𝑥𝑖𝑗 . 𝜌𝑖𝑗 , 0 ≤ 𝜉 ≤ 1, the use of wireless connectivity to links between sensors and 𝛽(𝜉) = { (𝑖,𝑗)∈𝐸𝑛 actuators and the IVN units to which they can connect. The 1, 𝜉 > 1, wireless connections therefore depend on the topology (see where 𝑥𝑖𝑗 is a binary decision variable which is equal to 1 if Fig. 1). Wireless links are prone to transmission errors, and the task is scheduled over the link between node 𝑖 and node are characterized by a reliability 𝜌 < 1. We consider that the wireless links utilize an Orthogonal Frequency Division 𝑗, and 𝜌𝑖𝑗 is the reliability of the link between node 𝑖 and Multiple Access (OFDMA)-based radio access interface. A node 𝑗. By introducing the reliability of the IVN links in (8), dedicated band with bandwidth 𝐵𝑊𝑚 is assigned for each of this scheme seeks selecting the most reliable path possible the 4 in-vehicle zones. Additionally, communications from the multiple available paths when allocating the task between sensors/actuators of all zones and the HPCU in the from the source to the computing unit. If the selected set of centralized mesh IVN topology (Fig. 1.d) use a dedicated links used to schedule the task has a reliability of 1, no band of bandwidth 𝐵𝑊ℎ . Each 𝐵𝑊𝑥 (𝑥 𝜖 {𝑧, ℎ}) is divided penalty is applied. Conversely, if the combined reliability is into 𝐾𝑥 orthogonal resources. Then, the data rate available at less than 1, a penalty (a value between 0 and 1) is introduced, any given time for communication resource 𝑘𝜖{𝐾𝑥 } in the inversely proportional to the overall reliability of the selected (𝑘) links. Eq. (8) also ensures that if a task exceeds its deadline, wireless link 𝑙 𝜖 𝐸 is denoted as 𝑟𝑙 (𝑡): (𝑘) a penalty of 1 is imposed, discouraging deadline violations. (3) 𝑟 (𝑡) = 𝐵𝑊 ∙ 𝑙𝑜𝑔 (1 + 𝛾 (𝑡))(1 − 𝐵𝐸𝑅). 𝑘
𝑙
2
𝑙
In (3), 𝐵𝑊𝐾 represents the bandwidth of the communication resource 𝑘 , 𝛾𝑙 (𝑡) denotes the Signal-toInterference plus Noise Ratio (SINR) at time 𝑡 of the wireless link 𝑙, and BER is the bit error rate, which depends on the modulation and coding scheme employed in the communication resource 𝑘. To model channel fading effects, we assume a Rayleigh distribution. The total data rate of the link 𝑙 is calculated as the sum of the data rates for all communication resources 𝑘 utilized in the link: (𝑘) (4) 𝑅 (𝑡) = ∑ 𝑟 (𝑡). 𝑙
𝑘
𝑙
The transmission time over the wireless link 𝑙 is then: 𝑠𝑛 (5) 𝑡𝑐𝑛 = , 𝑅̅𝑙 1 where 𝑅̅𝑙 = ∫Δ𝑡 𝑅𝑙 (𝑡)dt represents the average of 𝑅𝑙 (𝑡) Δ𝑡 over the time until the end of time slot that Δ𝑡. 𝑅̅𝑙 ≥ 𝑠𝑛 . Similarly, the transmission time over wireless communication links for the processed result of a task 𝑓𝑛 with size of 𝑠′𝑛 can be expressed as 𝑡′𝑛𝑐 and is computed following (5) using 𝑠′𝑛 instead of 𝑠𝑛 . The total execution time 𝑇𝑛 required to complete a task 𝑓𝑛 includes the communication time to transmit the task to the processing unit (𝑡𝑐𝑛 ), the processing time at the computing unit (𝑡𝑝𝑛 ), and the communication time to return the processed result (𝑡′𝑛𝑐 ). The total execution time is given by: (6) 𝑇𝑛 = 𝑡𝑐𝑛 + 𝑡𝑝𝑛 + 𝑡′𝑛𝑐 .
A. Constraints A first binary task scheduling constraint is defined as follows to ensure that task 𝑓𝑛 is assigned to a single computing unit and cannot be split among multiple units: 𝑀+1 (9) (𝑖) ∑ 𝑎𝑛 = 1, ∀𝑛 , 𝑖=1
where 𝑀+1 represent the total number of computing units (𝑖) (i.e., 𝑀 ECUs and 1 HPCU), and 𝑎𝑛 is a binary variable equal to 1 if task 𝑓𝑛 is allocated to the computing unit 𝑖. The second constraint, formulated in eq. (10), is only applicable to wireless links of the hybrid wireless-wired IVN topologies. Following OFDMA principles, this constraint ensures that communication resources from each band are allocated to only one communication link at a time. This prevents transmission collisions and enables interferencefree communication between different zones and the HPCU. 𝑁 (10) (𝑘) ∑ 𝑏𝑙,𝑛,𝑧 = 1 , ∀𝑘, 𝑙, 𝑧, (𝑘)
𝑛=1
where 𝑏𝑙,𝑛,𝑚 is a binary variable equal to 1 when communication resource 𝑘 is allocated to transmit task 𝑓𝑛 in link 𝑙 of band 𝐵𝑊𝑧 (𝑧 𝜖 {𝑚, ℎ}). The third constraint in eq. (11) ensures that the transmission rate for all tasks sharing a link does not exceed the link’s maximum achievable data rate. 𝑁 (11) ∑ 𝑅𝐸,𝑛 (𝑡) ≤ 𝑅𝐸 (𝑡), ∀𝑙, 𝑛=1
THIS IS AN AUTHOR-CREATED POSTPRINT VERSION
2025 IEEE/CIC International Conference on Communications in China (ICCC), Shanghai, China, 2025, DOI: 10.1109/ICCC65529.2025.11149335. between 5 and 15 Mcycles with an average of 10 Mcycles, and where 𝑅𝐸,𝑛 (𝑡) is the data rate of link 𝐸 = 𝑙 ∪ 𝑤 for task 𝑓𝑛 , their size ranges between 0.5 and 1.5 Mbits, with an average and 𝑅𝐸 (𝑡) is the maximum possible data rate of link 𝐸. size of 1 Mbits. The size of the processed result for each task Finally, the fourth constraint in eq. (12) ensures that the is set to 15% of its original size. According to the total processing workload of different tasks allocated to a requirements for in-vehicle functions identified in [7], task computing unit within a specific time interval does not deadlines (𝑇𝑛𝑚𝑎𝑥 ) are randomly assigned within the 40 to 100 exceed the unit’s maximum processing capacity. ms range. 𝑁 (12) (𝑥) When hybrid wireless-wired IVN topologies are ∑ 𝑐𝑛 𝑎𝑛 ≤ 𝐶𝑥𝑚𝑎𝑥 , ∀𝑥, 𝑛=1 considered, the dedicated total bandwidth 𝐵𝑊 for wireless inwhere 𝑐𝑛 denotes the computing demand of the task 𝑓𝑛 , and vehicle communication is 100 MHz, divided into five 𝐶𝑥𝑚𝑎𝑥 is the maximum processing capacity of unit 𝑥 𝜖{𝑧, ℎ}. segments: each zone is assigned a bandwidth 𝐵𝑊𝑚 of 20 MHz, and the wireless connection to the HPCU has a B. State-of-The-Art Schemes dedicated bandwidth 𝐵𝑊ℎ of 20 MHz. OFDMA The Deterministic proposal is compared against three communications are configured with a subcarrier spacing benchmark schemes. The Baseline task scheduling scheme (SCS) of 30 kHz and a time slot duration of 0.5 ms following follows a traditional static approach, where tasks are allocated 3GPP TS 38.211. Based on empirical in-vehicle wireless to predefined computing units [1]. For the IVN topologies measurements in [10], we assume that wireless links within defined in Section II, this means that tasks generated by the vehicle maintain an average SINR of 30 dB, with channel sensors/actuators are allocated to the ECUs within the same fading modeled using a Rayleigh distribution. The reliability zone, while the ECUs and HPCU process their own tasks. 𝜌 is considered to randomly vary in the range (0.95–1) for the The Shortest task scheduling scheme follows the classic wireless links between the sensors/actuators and their zonal shortest-path approach [4][5] and focuses on minimizing the ECU, and in the range (0.90–1) for the connections to crossphysical distance between the task source unit and the zone ECUs and the HPCU due to the largest distances and computing unit. Its objective function is formulated as: presence of blocking elements [10]. The wired links are (13) modeled with an Ethernet-based data rate of 1 Gbps and 𝜌=1. min ∑ 𝑑𝑖𝑗 ∙ 𝑥𝑖𝑗 , We implement a genetic algorithm to solve the NP-hard (𝑖,𝑗)∈𝐸 optimization problems of task scheduling as in [7]. The algorithm starts with 1,000 candidate solutions, retaining the where 𝑑𝑖𝑗 represents the distance between node 𝑖 and node 𝑗. top 20% for the next generation while generating the 𝑥𝑖𝑗 is a binary decision variable equal to 1 if the task is remaining 80% through crossover. Over ten generations, a scheduled over the link between node 𝑖 and node 𝑗, and equal 20% mutation rate introduces random variations to enhance to 0 otherwise. 𝐸 = 𝑙 ∪ 𝑤 is the set of IVN links. diversity and prevent premature convergence. This configuration balances performance and computational The Minimum task scheduling scheme follows a common complexity, achieving near-optimal solutions. The proposed strategy used in task offloading processes [6]. Its objective is and benchmark schemes are compared for the same number to allocate communication resources and computing units to of generations, ensuring a fair comparison by maintaining minimize task execution time. This scheme transmits tasks by identical run times. selecting jointly the fastest available path based on network topology and communication resources and fastest computing VI. RESULTS unit according to available processing capacity. The optimization function for this strategy is: We first evaluate the ability of the task scheduling (14) schemes under evaluation to successfully support in-vehicle 𝑚𝑖𝑛 ∑ 𝑇𝑛 , tasks across different IVN topologies. A task is considered 𝑛 successfully supported if it is executed before its deadline. where 𝑇𝑛 is defined in (6). Fig. 2 depicts the average satisfaction ratio as a function of the For fairness, the Shortest and Minimum schemes are number of generated tasks for the fully wired implementation defined with the same four constraints as Deterministic. Only of the four IVN topologies. The satisfaction ratio represents the first three constrains apply for the Baseline scheme since the proportion of tasks completed before their deadlines it follows a predefined assignment of the computing units. relative to the total number of tasks. Note that non-satisfied tasks are also completed, but after their deadlines have passed. V. EVALUATION SCENARIO The results show that the Baseline scheme, which relies on We analyze the impact of task scheduling on the pre-assignment of tasks to computing units, can support all performance of the IVNs following the topologies described generated tasks in scenarios with up to 25 tasks independently in Section II. The IVN consists of 36 sensors/actuators equally of the IVN topology. Its performance significantly degrades distributed in the 4 areas of the vehicle. Each area is controlled with higher workloads. A similar trend is observed for the by a zone ECU, and central computing is performed in the Shortest task scheduling scheme even if it can dynamically HPCU. While tasks can be generated by any element of the schedule tasks across the IVN. This is the case because it does IVN, only ECUs and the HPCU can handle processing. The so considering only the physical topology of the IVN to find processing power of the ECUs and the HPCU is set to 1 GHz the shortest path, and does not account for the computing and 4 GHz, respectively, based on the existing capabilities of capabilities and workloads of the units or the status of the links off-the-shelf IVN processing units [8]. Within the vehicle, in the IVN. The Minimum task scheduling scheme does take 70% of tasks are generated by sensors, 15% by ECUs, and into account this information to schedule tasks across the IVN 15% by the HPCU. We consider task processing workloads to minimize task execution time. This approach outperforms and sizes following the characterization of in-vehicle the Baseline and Shortest schemes across all IVN topologies, functions in [9]. In particular, we consider that tasks require
THIS IS AN AUTHOR-CREATED POSTPRINT VERSION
2025 IEEE/CIC International Conference on Communications in China (ICCC), Shanghai, China, 2025, DOI: 10.1109/ICCC65529.2025.11149335.
Fig. 2 Task satisfaction ratio for different schemes (wired topologies).
Fig. 4 Usage ratio of computing units (ECUs – left, HPCU – right) in the hybrid wireless-wired centralized mesh IVN topology.
Fig. 5 Latency in different topologies for the Deterministic (left) and Minimum (right) schemes. Fig. 3 Task satisfaction ratio for different schemes in wired & hybrid cross-zone (left), and wired cross-zone & hybrid centralized (right).
and achieves a satisfaction ratio above 95% in scenarios with up to 35 tasks. However, like the Baseline scheme, it can only fully execute all tasks within their deadlines in scenarios with up to 25 tasks. On the other hand, Fig. 2 shows that the Deterministic scheme can fully satisfy a higher workload and achieves a satisfaction ratio above 95% in scenarios with up to 45 tasks (50 tasks in the centralized mesh). Under this load, the Deterministic scheme increases the ratio of satisfied tasks by {26.8%, 27.1%, 8.8%}, {27%, 27.4%, 9%}, {29.7%, 30%, 6.5%} and {30.8%, 30.9%, 6.3%} compared to the {Baseline, Shortest, Minimum} schemes for the tree-based, basic mesh, cross-zone mesh and centralized mesh topologies, respectively. These results clearly demonstrate that a deterministic task scheduling approach can better guarantee deterministic service levels and can support increasing invehicle computational workloads. In addition, deterministic task scheduling can better leverage advancements in the IVN –such as cross-zonal connections in the cross-zone mesh topology (see Fig. 1.c)– by flexibly managing task completion deadlines to efficiently schedule tasks across the IVN. We also analyze the impact of introducing wireless links in the cross-zone mesh IVN topology, specifically in the connections between sensors/actuators and the ECUs. Fig. 3left compares the satisfaction ratio achieved with the fully wired and hybrid wired-wireless implementations of the topology. The figure shows that all task scheduling schemes experience a reduction in the ratio of satisfied tasks with the introduction of wireless links. However, the reduction is the smallest with the Deterministic scheme. For instance, under all considered task loads the reduction it experiences remains below 1.5%, while it increases to 4.3%, 8.6% and 3% for the Minimum, Shortest and Baseline schemes, respectively. This is because Deterministic takes the reliability of the links into account when scheduling tasks to computing units across the IVN. The introduction of wireless links improves the capacity to establish new links within the IVN, and facilitates the
THIS IS AN AUTHOR-CREATED POSTPRINT VERSION
flexibility and reconfigurability sought with SDVs. For example, it would be possible to evolve a fully wired crosszone mesh IVN topology to a hybrid wired-wireless implementation of the centralized mesh topology by adding a wireless link between sensors/actuators and the HPCU (Fig. 1). In this case, Fig. 3-right demonstrates that, with the hybrid centralized mesh IVN topology, the Deterministic scheme compensates the performance degradation resulting from the introduction of wireless connections in the hybrid cross-zone mesh IVN topology, and even achieves higher satisfaction ratios compared to the wired cross-zone mesh IVN topology. This is not actually the case for all the other schemes that fail to mitigate the impact of wireless connections, resulting in lower satisfaction ratios with the hybrid centralized mesh IVN topology than with the wired cross-zone mesh IVN topology. The results also show that the Deterministic scheme is the only scheme that achieves a satisfaction ratio above 95% in the scenario with 50 tasks in the hybrid centralized mesh IVN topology, outperforming alternative task scheduling schemes by 12.9% to 49.9%. The higher satisfaction ratios achieved with the Deterministic scheme stem from its better scheduling and more balanced workload and resource utilization across the IVN. This is illustrated in Fig. 4, which depicts the average ratio of utilized computing resources of computing units in the hybrid wireless-wired implementation of the centralized mesh topology; similar trends are observed in the other topologies. The left figure shows the average usage ratio of the zone ECUs, while the right figure depicts the usage ratio of the HPCU. Fig. 4 shows that the Baseline and Shortest schemes saturate the ECUs in the scenarios with 40 tasks or more, while the HPCU experiences a low usage ratio (below 25%) even when the ECUs are saturated. This saturation of the ECUs leads to the drop in the satisfaction ratio shown in Fig. 3 for the Baseline and Shortest schemes. The Deterministic and Minimum schemes distribute tasks across different computing units, making better use of the HPCU’s high processing power compared to the Baseline and Shortest
2025 IEEE/CIC International Conference on Communications in China (ICCC), Shanghai, China, 2025, DOI: 10.1109/ICCC65529.2025.11149335. schemes. Comparing the Deterministic and Minimum vehicle zonal E/E architectures with centralized computing. schemes, Fig. 4 shows that the Minimum scheme tends to Our analysis has demonstrated that a deterministic approach utilize more the HPCU to minimize the task execution times to task scheduling can better guarantee deterministic service in scenarios with low to medium task loads. However, under levels than alternative approaches, and can satisfactorily higher task loads, it utilizes the ECUs more than the support increasing in-vehicle computational workloads and Deterministic scheme. In contrast, the Deterministic scheme tasks. This is achieved thanks to a more balanced workload follows the opposite trend, relying more on the HPCU at distribution and resource utilization across the IVN. These higher task loads, which helps avoid bottlenecks in the ECUs trends have been validated across a variety of IVN topologies, by distributing the load more efficiently. This more balanced and also considering the introduction of wireless connectivity distribution results in the higher satisfaction ratios shown in in hybrid IVN topologies. Fig. 3 for the Deterministic scheme. REFERENCES Finally, Fig. 5 compares the latency or total execution time 𝑇𝑛 experienced by the Deterministic and Minimum schemes under different IVN topologies. Results are reported for the wired implementation of the tree-based, basic mesh, and cross-zone mesh IVN topologies, and the hybrid wirelesswired implementation of the centralized mesh topology. Fig. 5 shows that Deterministic experiences higher latency than Minimum in scenarios with low to medium task loads because it prioritizes maximizing the number of tasks completed before their deadlines (Fig. 2-Fig. 3) over minimizing latency. On the other hand, Deterministic reduces the latency under higher-load scenarios thanks to its capacity to efficiently adapt the tasks' scheduling to the computing workload, as shown in Fig. 4. Results in Fig. 3 showed that Deterministic was the task scheduling approach that could better handle the introduction of wireless links. This is also visible in Fig. 5 that shows that Deterministic reduces the latency in the centralized mesh IVN topology by up to 16.3% and 15.6% compared to the tree/basic mesh and cross-zone mesh IVN topologies, while Minimum only reduces it by 8.3% and 5.5%, respectively. VII. CONCLUSION This study has introduced a novel deterministic task scheduling scheme for in-vehicle networks, and has demonstrated its potential to exploit the capabilities of in-
THIS IS AN AUTHOR-CREATED POSTPRINT VERSION
[1]
V. Bandur, et al., "Making the Case for Centralized Automotive E/E Architectures", IEEE Trans. Veh. Technol., vol. 70, no. 2, pp. 12301245, Feb. 2021. [2] P. Laclau, et al., "Enhancing Automotive User Experience with Dynamic Service Orchestration for Software Defined Vehicles", IEEE Trans. on Intell. Transp. Syst., vol. 26, no. 1, pp. 824-834, Jan. 2025. [3] Robert BoschGmbH, “The next step in E/E architectures”, Aug. 2023. Accessed: Mar. 2025. [Online]. Available: https://www.boschmobility.com/en/mobility-topics/ee-architecture/ [4] A. Berisa, et al., “AVB-aware Routing and Scheduling for Critical Traffic in Time-sensitive Networks with Preemption”, Proc. ACM 30th RTNS, pp. 207-2018, Paris (France), 7-8 June 2022. [5] B. Xu, et al., “A Joint Routing and Time-Slot Scheduling Load Balancing Algorithm for In-Vehicle TSN”, IEEE Trans. Consum. Electron. (early access on IEEE Xplore since Feb. 2025). [6] J. Cai et al., “Multitask multi objective deep reinforcement learningbased task offloading method for industrial Internet of Things”, IEEE Internet Things J., vol. 10, no. 2, pp. 1848–1859, Sep. 2023. [7] S.D. McLean, et al., “Configuring ADAS platforms for automotive applications using metaheuristics”, Front. Robot. AI, vol. 8, Jan. 2022. [8] Advantech, "ECU-4784," Accessed: Mar. 2025. [Online]. Available: https://www.advantech.com/en-us/products/1-369nwl/ecu4784/mod_18553282-e8f5-4b32-a64b-1083f7182d36. [9] Y. Deng, et al., “Multi-hop task routing in vehicle-assisted collaborative edge computing”, IEEE Trans. Veh. Technol., vol. 73, no. 2, pp. 2444–2455, 2023. [10] E.A. Vitucci, et al. “6G-SHINE D2.3: Radio propagation characteristics for in-X subnetworks”, Dec. 2024.