DeepTuner : a system for search technique recommendation in program autotuning

Abstract

OpenTuner can help users achieve better or more portable performance in their specific domain through program autotuning. A key challenge for users seeking good autotuning performance in OpenTuner is selecting a search approach appropriate for problem. However, not only are current in-situ learning search approaches not robust enough to handle all search spaces, but there are also too many possible search approaches for a user to examine manually after factoring in composable techniques. In this thesis, we introduce DeepTuner, a system for search approach recommendation operating across OpenTuner autotuning sessions to facilitate development of robust transfer learning search approaches. By utilizing historical autotuning data via DeepTuner's technique recommendation endpoints, the new search approaches can efficiently explore the space of possible search approaches and the autotuning space simultaneously, resulting in an adaptive, self-improving search approach. We demonstrate the robustness that recommendation brings on nine problems spread over three domains for a variety of initial technique sets. In particular, we show that the new Database Initialized Recommendation Bandit Meta-technique is highly robust, performing on par or significantly better than various old in-situ search approaches in OpenTuner. We achieve up to 3.7x performance improvement over the old default in-situ search approach for OpenTuner in the TSP domain.