Book Description

The advancement of supercomputing technology enables scientific simulations at extreme scale and thus scientists' capability to study more complex problems at greater accuracy, leading to important scientific discoveries and breakthroughs. To validate and evaluate a simulation, the conventional approach is to output the simulation data to a massive storage device for post hoc analysis. One problem with this approach is that both the I/O capability and storage capacity are lagging behind a simulation's ability to generate data. As a result, scientists can only output a small fraction of the data, usually by sacrificing temporal resolution. Otherwise, a majority of the supercomputing time would be devoted to I/O. As supercomputing moves from the current petascale (1015) towards exascale (1018), the gap between the computing power and I/O capability is expected to become even wider, and therefore more data is expected to be discarded, which defeats the original purpose of performing the simulation at extreme scale. It has become clear that some simulated phenomena can only be captured at high precision and fine detail during the simulation before any data is discarded. This essentially suggests in situ processing. In situ processing is a promising concept, but there are still certain limitations to the current generation of in situ methods. Computational fluid dynamics (CFD) is an application domain that particularly benefits from in situ techniques. Common in situ tasks during CFD simulations include computing flow statistics, direct rendering of the flow field, and extracting flow features of interest. Performing these tasks in situ often requires the scientists to know a priori the specific aspects of the simulated phenomena to be captured or extracted. Often, such knowledge is not available before analyzing the data generated from simulations. Furthermore, many in situ methods pose a limitation; that is, the resulting geometric or imagery representations of the flow feature prohibit exploration. Such limitation greatly reduces the possibility of subsequent scientific analyses. This dissertation aims to enhance the explorability of in situ visualization in order to alleviate the requirement of a priori knowledge, and also evaluate the usability of in situ approach using two state-of-the-art combustion simulation codes. The basis of my approach is to allow in situ visualization parameters to be specified in a way to broaden the extent of post hoc exploration and analysis. Each of my techniques balances the explorability of the outputs against the costs to generate it. To support scalar field visualization, I derive a design embedding multiple layers of depth information into the in situ outputs, which not only enables post hoc isosurfaces visualization, but also supports 3D feature extraction and tracking without needing the original simulation data. To support vector field visualization, I derive a design based on an image-based, point-sprite cloud representation of flow movement to provide pathtubes exploration in multiple dimensions. To support post hoc statistical analysis, I generate regional probability distribution functions in situ, with which I conduct a long-term user-centered design study to iteratively enhance the usability of the solution. This dissertation introduces several new approaches to in situ visualization and analysis and successfully demonstrates their potential values to the simulation scientists.