About Us

Background

Research Focus

People

Repository

Workshops

Projects

Links

Contact MRI

CCS: Institutes: MRI: Background: HPC Rationale

• Introduction
• Scientific Rationale
• HPC Rationale
• Thrust of CCS-MRI
• Resources of CCS
• Next Steps
• Enabling National Science Goals

HPC Rationale

Measuring relative performance improvements is not as straightforward as measuring the increase in peak performance of the fastest super- computers because the scientific focus changes over the years. click for details

Materials scientists are already some of the largest consumers of HPC resources. Although many significant challenges remain, the award of the 1998 Nobel Prize to Professor Walter Kohn for his development of Density Functional Theory (DFT) - which provides the theoretical foundation for modern first principles quantum calculations of materials - is just one indication of the success and promise of this rapidly advancing field. Over the last three decades materials physicists/scientists have been in the vanguard of exploiting successive generations of computer hardware - scalar, vector, parallel - as these machines have become available to basic science. For example, the first principles LSMS code, a multiple scattering code being developed to study non-equilibrium and finite temperature properties of magnetic materials, was awarded the 1998 Gordon Bell Prize for "Best Performance on a Supercomputer Application" and continues to be the best performing code on the most recent IBM SP supercomputers.

The performance increase of computers over the last three decades has been staggering and the trend is not likely to end any time soon. However even more impressive is what computational materials science has achieved in relative performance increases through improved models and algorithms, and the creativity of scientists. This synergy between the increased capability and performance of computer hardware on the one hand and concomitant advances in theory and algorithms is illustrated in the sidebar. The tenant underlying CCS-MRI is that, in the future, this synergy can be best developed within an focused center in which theoretical and algorithmic advances are implemented in a software environment that is integrated with and optimized for state of the art high performance computing resources and that the whole is dedicated to the solution of the CMS community's most challenging problems.

As argued previously, nanoscience presents major theoretical and computational challenges that can be roughly categorized as (1) Integrated modeling of multiple length and time scales; (2) First principles based methods for correlations and excited states; (3) Far from equilibrium systems and non-linear response.

A common feature of these major themes is that making progress will require expertise that resides in multiple places and multiple disciplines. Consequently, addressing these problems also demands a new approach to software design. Ultimately, the current approach of developing monolithic codes to solve specific problems does not possess the flexibility, ease of use, or reusability of code required to solve this new class of problems. Thus, the CCS-MRI is not envisioned as simply a repository for the collection of existing monolithic codes running on high performance machines. Rather, it will provide a properly software engineered collection of elements that comprise a tool set whose components can be rapidly assembled into codes to solve complex problems.

Proper implementation of such a tool-set will rely on a suite of well-designed algorithms that emphasize efficiency, high performance on modern HPC platforms, and interoperability. These algorithms will necessarily cover a large range of topics in computer science and applied mathematics. Examples are parallel data structures, random numbers generation, graph theory, linear algebra, handling of large data structures and adaptive mesh generation. While in itself each of these topics may be well established and, to a certain extent, can be considered as "solved", it will now be necessary to apply these algorithms in a way that they function in the context of the CMS problems outlined above, in particular that they scale favorably when models are connected and systems sizes are increased. It is quite clear that such a project can only be realized in an environment that facilitates close collaborations between computational materials scientists, applied mathematicians, and computer scientists.