A visualization and analysis platform for performance tuning

Abstract

With a framework like OpenTuner, one could build domain-specific multi-objective program auto-tuners and gain significant performance improvements. But explaining why and interpreting the results are often hard, mainly due to the large number of parameters and the inability to figure out how each parameter affects the performance improvement. We have a solution that can explain the performance improvements by identifying key parameters while providing better insights on the tuning process. Our tool uses machine learning techniques to identify parameters which account for a significant performance improvement. A user could utilize different methods provided in the tool to further experiment and verify the accuracy of such findings. Further, our tool uses multidimensional scaling to display all the configurations in a two dimensional graph. This interface allows users to analyze the search space closely and identify clusters of configurations with good or bad performance. It also provides real-time information of tuning process which would help users to optimize the tuning process.