Optimizing for Visual Cognition in High Performance Scientific Computing
|Title||Optimizing for Visual Cognition in High Performance Scientific Computing|
|Publication Type||Journal Article|
|Authors||Ware, C, Rogers, D, Petersen, M, Ahrens, J, Aygar, E|
|Publisher||Society for Imaging Science and Technology|
High performance scientific computing is undergoing radical changes as we move to Exascale (1018 FLOPS) and as a consequence products for visualization must increasingly be generated in-situ as opposed to after a model run. This changes both the nature of the data products and the overall cognitive work flow. Currently, data is saved in the form of model dumps, but these are both extremely large and not ideal for visualization. Instead, we need methods for saving model data in ways that are both compact and optimized for visualization. For example, our results show that animated representations are more perceptually efficient than static views even for steady flows, so we need ways of compressing vector field data for animated visualization. Another example, motion parallax is essential to perceive structures in dark matter simulations, so we need ways of saving large particle systems optimized for perception. Turning to the cognitive work flow, when scientists and engineers allocate their time to high performance computer simulations their effort is distributed between pre and post run work. To better understand the tradeoffs we created an analytics game to model the optimization of high performance computer codes simulating ocean dynamics. Visualization is a key part of this process. The results from two analytics game experiments suggest that simple changes can have a large impact on overall cognitive efficiency. Our first experiment showed that study participants continued to look at images for much longer than optimal. A second experiment revealed a large reduction in cognitive efficiency as working memory demands increased. We conclude with recommendations for systems design.