Real-Time Virtualization of Mixed-Criticality Heterogeneous Embedded Systems for Fusion Diagnostics and Control

The rise of complex, safety-critical systems, such as nuclear fusion reactors, has highlighted significant challenges in optimizing resource utilization on modern Multi-Processor Systems on chips. These platforms offer substantial computational power, but strict real-time guarantees often lead to overprovisioning. As a result, only a small fraction of available resources are effectively utilized, which can hinder the system's overall efficiency and longevity, especially when the hardware is expected to operate for decades in reactors. This dissertation addresses these challenges by exploring real-time virtualization as a key solution for improving hardware utilization in mixed-criticality systems. It begins with an in-depth analysis and testing of virtualization technologies, focusing on enhancing temporal isolation between applications that share the same hardware. A systematic Design-of-Experiments approach is used to assess and optimize the isolation provided by different virtualization configurations. Building on this foundation, the core contribution of this dissertation is the development of the Omnivisor, an innovative virtualization model that extends existing partitioning hypervisor architectures to manage asymmetric cores (such as ARM64, ARM32, and RISC-V) and control memory bandwidth allocation for accelerators, such as FPGAs and GPUs. The Omnivisor ensures isolation between virtual machines and predictable behavior across mixed-criticality applications, enhancing system reliability and resource efficiency. Additionally, the RPUGuard communication framework, integrated into the Omnivisor, provides fine-grained control over the communication between virtualized asymmetric processors, reducing interference and ensuring real-time performance. To ease the utilization of Omnivisor in distributed systems the RunPHI framework is introduced. RunPHI integrates the Omnivisor into a cloud-to-edge orchestration system, simplifying the deployment of distributed mixed-criticality systems across heterogeneous environments. The contributions of this work are demonstrated through real-world use cases related to nuclear fusion, where the proposed models and frameworks facilitate the reliable and efficient operation of safety-critical control applications. The evaluation phase shows how innovative virtualization techniques can allow critical controllers to co-exist with utility applications on the same hardware, ensuring they remain isolated while sharing resources.