The following HTML text is provided to enhance online
readability. Many aspects of typography translate only awkwardly to HTML.
Please use the page image
as the authoritative form to ensure accuracy.
Computational and Theoretical Techniques for Materials Science
such libraries, each optimized for that vendor's MPP, thus providing optimal performance. Such calls to platform-specific library routines and functions will be easy to change (if needed) to run on different platforms. These libraries include such widely used routines as fast Fourier transforms, linear algebra, various solvers, random-number generators, and communication primitives.
Visualization. Visualization is an important part of code development, debugging, and data analysis and is, therefore, crucial for migration of code to MPPs. One reason for going to MPPs is to treat much bigger problems much faster, and by the very nature of this goal, immense data sets will be generated. Visualization of these data sets will not only be desirable but in fact vital to understanding the phenomena being studied. This presents some large problems that vendors are addressing. One such problem is the communication rate. Shipping huge data sets to graphics workstations is already becoming impractical because the transfer rates tend to be low, the available memory insufficient, and the processing speed of the workstations too slow to process such huge data sets in a timely way. Therefore, the tendency will be to do the visualization on the MPPs themselves, which will require parallel graphics packages. These packages will of course need high-bandwidth parallel I/O, which is already available from some vendors and will be available from others. Parallel I/O allows the user to access the data rapidly and the nodes of the MPP to start the visualization. For some remote users who may have difficulty transferring the full three-dimensional rotatable visualization to their screens (because of the network connection to the MPP), data reduction will be the way to solve this problem. They can, for example, transfer one isosurface at a time, or even (if the data set is exceedingly large) a single two-dimensional slice at a time. Such reduced data sets will allow remote users to benefit from the power and versatility of MPPs and parallel visualization over the network. The arrival of very fast network connections will alleviate many of the connection problems faced by remote users and the data transfer problems among computers.