ASCI Tera-scale Scientific Data Analysis and Visualization at LLNL

The Accelerated Strategic Computing Initiative (ASCI) vision is to rapidly shift from nuclear test-based Stockpile Stewardship to a high-performance computing and data analysis environment with performance, reliability and confidence which is unprecedented in numerical simulation. ASCI applications will use extremely high fidelity computer models (on the order of 1 billion cells) to generate terabytes of raw data, which will be analyzed by physicists who rely on the visualization expertise of the three defense programs laboratories, Los Alamos National Laboratory, Lawrence Livermore National Laboratory, and Sandia National Laboratories. The challenge of analyzing and visualizing the tera-scale datasets is being addressed with a combination of high-performance storage and networking with a scaleable visualization architecture which permits interactive exploration of large volumes of data.

Visualization Architecture
Tera-scale visualization architecture is driven by the need to visualize the results of simulations which may have been computed on the desktop, down the hall, or thousands of miles away. Flexibility and adaptability of data access mechanisms are required in order to accommodate orders-of-magnitude variations in processing capability, memory, and storage. This flexibility is realized through a segmented architecture which separates data services from the interactive front-end. The data server and the front-end are designed such that they can inter-operate in a variety of configurations.

Data Services
Intelligent data storage and caching are key to the ability to interact with tera-scale data. The data server lies conceptually between the simulation code and the visualization application, providing access to the portions of the data required for processing. The principal responsibilities of the data server are to satisfy client queries for sub-meshes and slices, both at full resolution and at adaptive resolutions. Hierarchical data representations permit spatial queries and selective refinement based on preprocessed error metrics which are stored in the hierarchy. The data server may be located on the MPP running a simulation, or on an auxiliary server or cluster of workstations with large aggregate memory and access to high-performance storage.

Visual Browser
The data browser allows interactive navigation of tera-scale scientific data by acting as a client of the data server. The user interactively manipulates slices and isocontours of the data while gaining a global view of the time dynamics of the simulation through an intuitive interface for navigating the time dimension. The browser may be integrated into existing visualization systems in order to leverage the comprehensive visualization tools available.

Compression
In order to make efficient use of available memory and storage, we are evaluating approaches to caching of compressed data in primary memory, in addition to paging and high-speed transmission of compressed representations for distance visualization. Research is currently being conducted in 3D and 4D wavelet compression with adaptive sparse decompression and paging capabilities.

Adaptive Isocontouring
The data server is currently being used to extract adaptive meshes which conform to a user-selected set of isovalues. The resulting adaptive mesh is transparently downloaded from the data server to a desktop running MeshTV, a visualization package developed at Lawrence Livermore National Laboratory. The adaptive data selection is performed through octree approximation of a regular grid of volume data. The resulting adaptive mesh is composed of mixed cell-types and enforces continuity of the approximate scalar field. Figure 1 illustrates the resulting adaptive mesh and a contour extracted from the mesh. This example illustrates a common situation, in that the desktop visualization tool frequently cannot deal with the original tera-scale data, however adaptive approximations and subsets bring the data to a level which is manageable for existing tools.

Fig 1: Visualization of a Rayleigh-Taylor Instability simulation in MeshTV. A continuous, adaptive mesh extracted in preprocessing for isocontouring (above) and the resulting adaptive isocontour (below)

Fig 2: Detail of the adaptive mesh in Figure 1

Interactive Techniques
Scaleable visualization requires efficient algorithmic techniques for accessing only the data which is required for a particular visualization.

Preprocessing of hierarchical data representations allows the computation of view-dependent error estimates from the hierarchical error bounds on the data. We are using these view-dependent error measures to adapt the rendering of meshes, contours and slices to the current user view during interactive exploration of the data. The eventual goal is to develop rendering approaches which are bound by the complexity of the output images, rather than the input data, while maintaining a guaranteed level of accuracy at the pixel level. The image in Figure 3 demonstrates adaptive triangulation of a large terrain mesh based on viewer perspective, developed collaboratively with researchers at Los Alamos National Laboratory. As the viewer location or orientation are changed, the triangulation is updated in time proportional to the number of changes rather than the size of the data, while error bounds in the geometry distortion in screen space are minimized.

Fig 3: Adaptive view-dependent meshing of a terrain (top) and the viewer perspective (bottom)

We are also addressing the need for interactive computation as well as interactive display. Current isocontour extraction approaches are output sensitive in computational complexity, however the resulting surface lacks the hierarchical properties that are required for view-dependent rendering. Figure 4 illustrates a typical surface (composed of several million triangles) extracted from an ASCI turbulence simulation. We are currently developing an integrated approach which performs both computation and rendering in a view-dependent, output-sensitive manner. For increased interactivity during modification of the isovalue, computation may be short-circuited by increasing a variable error tolerance in the output image.

Fig 4: Full resolution contour surface extracted using an accelerated value-space cell search technique. The surface is colored by the gradient magnitude of the scalar field at each point.

Fig 5: Close-up of the isocontour in Fig 4