Efficient Search-Space Encoding for System-Level Design Space Exploration of Embedded Systems

For Design Space Exploration (DSE) of embedded systems as a combinatorial Multi-Objective Optimization Problem (MOP), metaheuristic optimization approaches are typically employed to determine high-quality solutions within limited optimization time. This requires the encoding of implementations from the design space in a search space which represents the available degrees of freedom for the optimization approach. Determining an encoding that ensures all design constraints are met by construction is, however, impossible for multi-/many-core DSE problems, so that the search space contains infeasible solutions. While state-of-the-art DSE techniques repair infeasible solutions, little to no attention has been paid to the efficiency of the resulting encoding w.r.t. its suitability for the employed optimization approach. Therefore, we formally define requirements for an efficient search space and analyze the drawbacks of automatically generated inefficient encodings. We furthermore present efficient search-space encodings for a state-of-the-art hybrid optimization approach suitable for a wide range of MOPs. The proposed encodings significantly reduce the required degree of repair, allowing us to introduce a feedback loop from repaired solutions in the design space to the respective encoded solutions in the efficient search space to further improve the optimization. The positive effects of the proposed efficient encoding and design-space feedback are demonstrated for system-level DSE using benchmarks from the domains of embedded many-core as well as networked automotive systems. Compared to inefficient search spaces from literature, significant enhancements in both optimization quality and time are observed. Furthermore, we propose metrics to quantify search-space efficiency which provide novel insights into the interdependence of search space and design space for multi-/many-core DSE.