On Search-Space Restriction for Design Space Exploration of Multi-/Many-Core Systems

Design Space Exploration (DSE) for embedded system design with its multi-objective nature and large search spaces typically prohibits exhaustive search and popularized the use of metaheuristic optimization techniques. Recent large-scale multi- and especially many-core architectures offering a multitude of application mapping possibilities create tremendously large search spaces which give rise to the question whether established metaheuristics are still efficient. In this work, we propose to employ a heuristic search-space restriction (SSR) approach based on the exploration of subsystems, which significantly reduces individual search-space size and, thus, exploration time. Knowing that this kind of restriction may miss global optima, we also investigate the use of high-quality solutions derived from subsystems as an initial population for the optimization of the complete system. Experimental results for tiled 8x8 to 24x24 many-core architectures and several benchmark applications show that the proposed SSR enables the metaheuristic to derive implementations of higher quality in a significantly reduced exploration time. Although not all global optima are exposed to the restricted problem, this work gives evidence that too complex search spaces may sacrifice the efficiency of metaheuristics drastically and, thus, serves as a motivation for future research into SSR or splitting techniques for DSE