SCALABLE ALGORITHMS FOR COMPUTATIONAL CHEMISTRY

Ian T. Foster, Jeffrey L. Tilson, Albert F. Wagner, Ron Shepard, Argonne National Laboratory; Robert J. Harrison, Rick A. Kendall, Rik J. Littlefield, Pacific Northwest Laboratory

One of the most critical problems facing the world community today is cleaning up hazardous wastes. Computer scientists and chemists at Argonne National Laboratory and Pacific Northwest Laboratory have teamed up to develop new algorithms, software, and diagnostics for solving this problem.

Their efforts have focused on electronic structure codes that help predict the structure, spectroscopy, and reactivity of species involved in environmental remediation. Currently, they are developing a scalable Hartree-Fock code, a foundation method in electronic structure theory upon which many other methods are built. Their goal is to develop methods that are capable of solving problems orders of magnitude larger than currently possible.

To achieve this goal, the group is exploring parallel decomposition algorithms that can run effectively on a wide range of parallel computer systems, from networks of multiprocessor workstations to the massively parallel teraflops machines of the future. These algorithms will be implemented and evaluated on a variety of machines, including the Intel Delta, the IBM SP, and other MIMD machines as they become available. Prototype codes have already been written and tested on several homologous series of molecules (e.g., linear alkanes and silicon oxide clusters) using primarily the Delta, with testing commencing on the SP.

A variety of software tools and techniques are being used in the prototyping and evaluation efforts, including the TCGMSG and p4 communication libraries and the Upshot and ParaGraph performance visualization displays. In addition, new tools and techniques are being developed that provide scalable linear algebra operations, portable support for globally addressable, distributed data structures, and built-in performance modeling and analysis.

The prototype Hartree-Fock code achieved speedups of 80% to 90% of the ideal speedup if the number of processors were selected to be less than the "natural" number for the problem (related to a basis set expansion size). The latest algorithm demonstrates much-improved scaling, with an approximately constant efficiency of about 90%.

The code is currently restricted to determining molecular energy as a function of molecular geometry. Further efforts are being directed at obtaining derivatives of the energy with respect to molecular geometry and incorporating fast summation of long-range interactions. The group is also testing the code on "proof of principle" problems developed by industrial partners in the project, which include researchers at Allied Signal, Inc., Amoco Chemical Co., E. I. du Pont de Nemours Co., Exxon Research and Engineering Co., and Phillips Research Center.

Table of Contents