Edge-Cloud-Assisted Real-Time Cyber-Physical Systems

Engineering notes

Key topics: autonomous driving, deployment, control. See the paper for implementation details and experimental results.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Cyber-Physical Systems (CPS) represent a type of interconnected systems where computational and physical processes are deeply integrated. These systems are increasingly prevalent across domains like automotive, aerospace, healthcare, and industrial automation. With increasing autonomy and complexity, CPS must not only be functionally correct but also meet stringent dependability requirements, especially when deployed in dynamic environments. As the demands for responsiveness and autonomy grow, CPS must meet strict *real-time constraints*, where not only functional correctness but also the timing of operations is critical. Real-Time CPS (RT-CPS) embody this by having time constraints, making them sensitive to delays and variations. In safety-critical domains, such as autonomous driving, healthcare, or industrial automation, violating these timing constraints can lead to degraded performance or system failure, underscoring the need for precise and predictable execution. To meet growing computational demands, CPS are increasingly being deployed on *edge-cloud platforms*. Cloud computing provides virtually unlimited computing capacity and scalability, while edge computing offers low-latency responsiveness by processing data near the source. The combination, known as the edge-cloud continuum, enables CPS to balance performance, scalability, and responsiveness. However, this architectural shift introduces new complexities. While the cloud provides centralised power, it suffers from variable latency; edge nodes, though responsive, have limited capacity. Integrating real-time guarantees across this continuum raises key issues in scheduling, orchestration, and predictability under uncertainty. This thesis addresses the convergence of *real-time* and *edge-cloud computing* in CPS. The integration of real-time guarantees in the edge-cloud continuum is particularly challenging as applications and workloads become more dynamic. The main challenges include ensuring temporal correctness under bursty workloads, preserving service continuity and fairness across applications, adapting to mobile environments, mitigating variable network conditions, and migrating legacy systems into modern container-based orchestration platforms like Kubernetes. The thesis contributes a suite of novel techniques and frameworks to address these challenges: - **Online job-level admission control for real-time applications with deadline guarantees:** We address the problem of guaranteeing timing correctness for jobs in dynamic and shared multicore server-based environments, where workloads are unpredictable and applications compete for processing capacity. We propose a runtime job-level admission policy that uses schedule-abstraction graphs to determine whether incoming jobs can meet their deadlines under non-preemptive fixed-priority scheduling. The policy provides lightweight runtime admission control that ensures per-job deadlines. - **Online admission control with deadline and admission ratio guarantees:** We address the problem of ensuring fairness and service continuity in multi-tenant environments, where bursty or overloaded conditions can compromise real-time constraints. We propose a job-level admission policy based on the m/k-firmness model, which requires that *m* out of *k* subsequent jobs of a task meet their processing deadlines. This policy ensures consistent service levels during bursty workloads by combining static m/k guarantees with dynamic, opportunistic admission of additional jobs, providing both deadline guarantees and a minimum level of service under varying load conditions. In addition, the contribution considers allocating optimal resources to tasks that cannot be scheduled on a single core, and distributing the runtime workload efficiently among the allocated resources. - **Managing dynamic resources and service migration:** We address the challenge of adaptive resource management in dynamic edge-cloud systems, where workloads and resource availability fluctuate due to mobility or workload variation. We propose algorithms for real-time service allocation and dynamic scaling across edge-cloud platforms, using the Linux `SCHED_DEADLINE` scheduler to provide temporal isolation. These algorithms adapt Quality of Service modes based on workload changes, cost, and resource availability, supporting dynamic scaling and service migration while preserving real-time guarantees. - **Real-time scheduling of Kubernetes containers:** We address the problem of scheduling containers with the `SCHED_DEADLINE` scheduler on a real-time Linux kernel. We propose a Kubernetes-based resource driver that enables the deployment of containers with real-time `SCHED_DEADLINE` reservations through native Kubernetes interfaces. - **Kubernetes-based real-time admission control and load distribution framework:** We show the feasibility of ensuring per-job deadlines for containerised applications in dynamic, multi-tenant environments in a realistic Kubernetes deployment. The deployment provides a scalable framework that integrates admission control based on the m/k-firmness model, load distribution, and `SCHED_DEADLINE` scheduling, ensuring robust real-time guarantees in modern cloud-native systems. These contributions address core challenges in enabling predictable, real-time performance for RT-CPS deployed over edge-cloud infrastructures, providing robust, fair, and predictable execution of CPS workloads under dynamic conditions. They bridge the gap between theory and practice, bringing formal real-time guarantees into containerised, cloud-native systems without sacrificing modularity, portability, or scalability.

Links and sources

• PDF from original source

Comments