Project reference : ANR-21-CE46-0005
Context
Visualization is a tool that is getting more and more widespread, and has become essential today in almost every scientific fields. It is embedded in most experimentation workflows, at different levels. Visualization tools are used to analyze datasets or extract information from them, to guide phenomenon modeling, to validate or invalidate models or as a tool for evaluating experimental results.
On the other hand, numerical simulations help to validate and optimise more and more complex systems and technologies, thus considerably decreasing risks and costs. Generally, their main objective is to predict a system's behaviour before it is effectively manufactured in practice. This process inevitably requires the use of modern visualization methods at different stages to validate the simulation codes and detecting conceptual or programming errors. However, coupling a visualization system with a parallel computation is not a trivial task: data structures must be adapted, large data amounts must be managed and the usually complex nature of the data makes that visualization is far from being a simple task. Furthermore, the access to increasingly powerful computing machines enables scientists to simulate ever larger and more complex phenomena. Large-scale simulations generally output time-varying multivariate volumetric data, modeled by volume meshes of increasingly complex nature on several aspects:
1. Their size: large-scale simulations can induce raw data that far exceeds the amount of available memory on standard PCs or medium size supercomputers;
2. Their composition: unstructured, hybrid (containing cells of different geometries), of high polynomial order, with non-convex cells and non-planar faces…
Direct volume rendering (DVR) is a well known method for visualizing volume data and its implementation on graphics processors (GPU), based on volume ray-casting algorithm [69], offers good rendering quality combined to good performance. However, such an implementation on simulated data presenting above-mentioned characteristics is a difficult problem that remains open. This can be explained by the fact that GPU memory is still much more limited than CPU memory and their SIMD parallel architecture works well for simple and regular data that some former simulations used to produce, but much less for unstructured and complex data that more advanced simulations now massively produce. A key challenge of research is to make visualization techniques follow up with this drastically increasing complexity.
On the other hand, the interest for scientists to use in-situ visualization approaches [43] for large-scale simulations is justified by their need to obtain high-performance visualizations to guide the simulation as it progresses. Nevertheless it also requires considering all the specificities of the complex nature of the data to be visualized, while leveraging a high- performance computing environment that can also be challenging to use and implement.
Objectives
We intend to overcome the above-mentioned issues to establish GPU- based modern volume ray-casting as a reliable tool for in-situ visualization of large-scale simulation that produces large and complex volume data. To this end, we propose to focus on two main aspects.
First, the development of a GPU-based visualization method capable of handling geometrically and topologically complex volume data and whose memory representation exceeds the physical capabilities of the machine used.
For this, we want to rely on an out-of-core approach. It must be able to efficiently access from the GPU, to the whole topology of an unstructured mesh i) of very large dimension (several hundreds of millions of cells or more) and ii) composed of any kind of polyhedron. We propose to integrate this system to the implementation of a volume ray tracer that takes up the key concepts established today to offer good performances with a good rendering quality (empty space skipping, adaptive sampling ...). To these aspects will be added the management of the visualization of physical quantities coming from multi-variate, nonlinear, nonconvex meshes, with non-simplex cells and non-planar faces, while working interactively with the previously mentioned out-of-core system.
