Towards design space exploration of industrial distributed cyber-physical systems

Abstract

The increasing complexity and performance requirements of industrial distributed cyber-physical systems (dCPS) create a growing need for efficient and accurate early-stage design evaluation. Design Space Exploration (DSE) supports this process by enabling system designers to evaluate alternative hardware and software configurations before implementation. However, existing DSE approaches are generally not suitable for large-scale industrial dCPS because they lack sufficiently abstract system-level models, depend heavily on manual model construction, and provide limited support for scalable performance evaluation.

This thesis introduces CompDSE, a structured methodology for system-level DSE of the computing components of industrial dCPS. CompDSE combines abstract system modeling, automated model derivation, model transformation, code generation, and simulation-based performance analysis within a coherent workflow. The methodology is based on separate application workload, hardware platform, and workload-to-platform mapping models. An additional environmental influence model represents timing effects caused by untraced subsystems interacting with the physical environment.

The models are automatically derived from application-level and system-level execution traces collected from an existing industrial system. The derived models are transformed into executable simulation entities and simulation-ready code, enabling efficient evaluation of alternative hardware architectures and software-to-hardware mappings. CompDSE is evaluated using the ASML TWINSCAN lithography machine as a representative industrial dCPS. The evaluation shows that the automatically derived models reproduce the measured system behavior with a difference of 2.7% in average wafer processing time, while a single design configuration can be evaluated in under one minute. These results demonstrate that the proposed methodology provides an effective balance between accuracy, and simulation efficiency. The thesis further introduces a refined application workload model that captures both inter-process and intra-process dependencies. This model enables the identification of software tasks that can be safely reordered or executed in parallel without requiring significant modifications to the existing software infrastructure. The resulting approach supports the analysis and exploitation of parallelism in complex industrial software workloads.