Scientific Visualization System And HFFT Over Non-Tensor Product Domains

Scientists nowadays have unprecedented computing and instrumental capability for studying natural phenomena at greater accuracy, resulting in an explosive growth of data. However, current data handling and visualization capacities still seem orders of magnitude too small for scientists to interpret the voluminous and complex data they are capable of producing routinely. This dissertation presents our efforts on this topic.First, we propose a visualization system model based on large-format display technology and so called Client-Commware-Server structure. The large-format display system based on clusters and projectors , which is now an area of active research, can not only provides high-resolution displays but also is suitable for parallel processing. Compared with the traditional Client-Server structure which is widely used in scientific visualization systems, our Client-Commware-Server structure decreases the connection between the server for computing and the client for visualization and makes the whole system more independent and flexible, which brings great convenience for system maintenance, expansion and reuse.We have accomplished two visualization systems: PRIS-Explorer (a postprocessing system for reservoir simulation) and SMSP(Scalable Materials Simulation Package). They are not only the implementations of the model we give above, but also meaningful for developing the practical software with our own copyrights. The key technologies for developing PRIS-Explorer are how to handle the large-scale data set produced by parallel programs and the implementation of the Object-Oriented technology. For SMSP, the main difficulty is how to build a Molecule-Building System based on OpenGL.Data storage,compression,transmission and processing are the key elements in large-scale visualization and FFT is one of the most important algorithms to handle these.In fact.FFT is one of the top 10 algorithms in the last century, and it plays an important role in lots of areas. But it can only be used in tensor-product domains, and how to generalize the approach into high dimensions, beyond box domain, is always an open problem.In this dissertation we propose a set of fast algorithms for computing the DGFT (Discrete Generalized Fourier Transforms) over the hexagon domains and the parallel dodecahedron domains and give the parallel implementation on Linux clusters.We call this set of fast algorithms over 2-dimension with 3-directions, 3-dimension with 4-directions, and generally m-dimension with m+1-directions domains HFFT(High-Dimension Fast Fourier Transformations).HFFT is a typical data-dense computation, and the communicaticfas are dominant during the parallel processing, which makes the efforts on how to reduce the data communication the key element to obtain high performance. By giving a specific recursive data structure we given and making full use of the parallelity of HFFT, we propose a parallel HFFT algorithm on distributed memory^ computers. The experiments on a Linux cluster up to 128 CPUS show that our algorithm has good parallel performance.The main contributions of this dissertation are In chapter 2, we propose a scientific visualization system modal which is based on large-format display technology and Client-Commware-Server structure. In chapter 3 and chapter 4, we accomplish two large-scale scientific visualization systems. : In chapter 5 and chapter 6, we propose HFFT on hexagon and parallel dodecahedron domains, and reduce the computational complexity fromO(N4) and O(N6) to O(N2logN) and O(N3logN). Lastly, we propose a parallel HFFT on distributed memory computers.