For several months, NCSA has been conducting scalability studies and cross-architecture comparisons of a number of academic and third-party applications codes in its multivendor scientific computing environment. The purposes of this effort are (1) to evaluate the effectiveness of various parallel architectures for large-scale scientific computing, (2) to determine the level of effort required to port realistic-sized codes to parallel environments, and (3) to determine the extent to which portability is achievable across diverse architectures (and at what price to performance ratio).
This project has focused on three major parallel platforms already in place at NCSA: the Thinking Machines CM-5, the CONVEX Exemplar, and the SGI Power Challenge. These three machines represent three major architectural paradigms: pure distributed memory, distributed memory with a global address space, and pure shared memory, respectively. Although NCSA's efforts will naturally continue to focus primarily on machines in-house, the center would expect to expand the list of architectures used through collaboration with other institutions.
Since much of the production work of NCSA's industrial users relies on third party application codes, we have initiated a new program to evaluate the most heavily used third party codes on NCSA's various architectures. Nine application teams have been created, one for each of our partners' most widely used application codes (ABAQUS, CHARMm, FIDAP, Flow3D, FLUENT, Gaussian, KIVA, NASTRAN, and Spectrum). Each team is anchored by an appropriate NCSA applications scientist and contains volunteers from among the largest partner users. So far, there are 16 strategic user volunteers signed up from six of NCSA's industrial partners. These teams will create nonproprietary, but relevant, application datasets of varying sizes, so that the scalability of each application code can be assessed. The results of these runs, made during the next six months, will be made available to NCSA's partners. For details, see the NCSA Application Performance Home Page on the World Wide Web via NCSA Mosaic.
In a similar manner, NCSA has called upon its academic research user community to collaborate in exploring and understanding the scalability of these systems as applications move from a single to hundreds of processors. Currently, there are efforts in the areas of astrophysics, biology, chemistry, earth sciences, gravitation, fluids, and structures. The goal is to optimize the working environment for users and to determine the most appropriate upgrade strategy for the future.
NCSA has also invited various computer vendors to collaborate in this effort. The initial focus has been on the SGI Challenge, as it offers the most mature development environment at present. With the arrival of the CONVEX Exemplar and upgrades to the operating system and processors of the SGI Challenge, our efforts will be expanding to these new systems.