Data Mining in System-Level Design Space Exploration of Embedded Systems

With increasingly complex applications and architectures, the task of determining Pareto-optimal implementations at the system level becomes a challenge even for state-of-the-art Design Space Exploration (DSE) methodologies. In this field, nature-inspired techniques such as Evolutionary Algorithms (EAs) are frequently employed, since they are well-suited to the multi-objective and hard-constrained nature of the DSE optimization problem. On the other hand, meta-heuristic approaches are problem-agnostic and are often observed to converge relatively quickly. Furthermore, this type of optimization lacks explainability, i.e. the way in which the optimization algorithm arrives at improved solutions as well as the individual contributions of design decisions to the resulting quality of a solution are not at all clear - and are consequently not utilized during DSE as of yet. To remedy this, we propose the integration of automated data-mining techniques into state-of-the-art DSE flows. Data mining is, thereby, used for (a) the automatic extraction and generation of previously untapped information from the optimization process to be (b) incorporated into the DSE to enhance optimization quality. We present a variety of ways to extract and include relevant knowledge during DSE, as well as (c) several possibilities to gain insight into the interdependence between decision variables and optimization objectives. Experimental results for benchmark systems for large-scale many-cores to networked embedded systems demonstrate the potential of the proposed techniques to improve the quality of the optimized implementations at no DSE-time overhead.