Sparse Measurement Algorithm Execution Time Prediction on Heterogeneous Edge Devices for Early Stage Software-Hardware Matching

The design and implementation of edge computing solutions needs elaborate knowledge about the inter-dependencies of software and hardware, which are also often developed parallel to meet performance requirements like time constraints. However, the suitability of hardware to execute a given algorithm and vice versa is often assessed by time-consuming trial-and-error approaches. For algorithm assessment, the execution time is crucial. However, existing execution time estimation approaches usually rely on either thorough timing models for the underlying hardware or vast amounts of measurements. Consequently, these approaches are not feasible for an early design stage, where an assessment with limited effort is crucial. Thus, this paper tries to overcome those limitations by comparing a parametric and a non-parametric execution time prediction approach suitable for an early design stage. The evaluation with four edge devices out of heterogeneous categories using measured data is a first attempt at generalizability. A selection of four different algorithms applicable in smart manufacturing and benchmarking ensures the approaches' broad applicability. The parametric and non-parametric model comparison shows the trade-off between source code analysis and performing measurements.