A Systematic Approach for Mapping ForSyDe Applications to NVIDIA Holoscan

Sammanfattning

Model-based design provides a systematic approach for developing complex embedded systems by enabling early abstraction and formal analysis. Formal System Design (ForSyDe) is a well-established framework in this domain and is developed as a system design methodology for heterogeneous embedded systems. In parallel, Holoscan serves as an execution framework originally developed by NVIDIA to enable low-latency, real-time streaming applications on embedded and edge computing systems. At the moment, there is no systematic mapping methodology that connects formal high-level modeling frameworks like ForSyDe with the practical execution framework like Holoscan. This thesis addresses the problem of how applications modeled in ForSyDe, specifically using the Synchronous Dataflow (SDF) model of computation, can be reliably mapped to the NVIDIA Holoscan framework. The challenge lies in bridging the static, token-based execution semantics of SDF with the dynamic, message-driven execution and scheduling mechanisms of Holoscan, while preserving functional correctness and predictable behavior. In this thesis, a set of explicit mapping rules are proposed, which clarify how ForSyDe components can be represented within Holoscan execution model. Common SDF communication patterns are summarized as point-topoint, one-to-many and many-to-one connections and corresponding rules are proposed for each scenario. The proposed mapping is validated through a series of controlled experiments that examine execution order, token handling and runtime behavior under different Holoscan scheduling strategies. A performance analysis further illustrates how different schedulers influence the runtime behavior and semantic preservation of the mapped applications. Finally, a complete image processing case study demonstrates the applicability of the proposed approach in a realistic scenario.