Parallel Computing in Quantum Chemistry

Parallel computing is one of the greatest challenges in computational Quantum Chemistry and also one of the greatest promises. The possibilities for increasingly realistic simulations of chemical systems have expanded dramatically over the last years. The computer systems which can be used for quantum chemical calculations span a very wide range nowadays starting from high-end PCs over workstations to supercomputers. The enormous computational power of supercomputers is not so much based on faster single-processor units but on the combination of dozens to several hundreds of processors and on more powerful interprocessor communication. Multi-reference configuration interaction (MRCI) (and related procedures like MR-ACPF and MR-AQCC) are very useful and flexible quantum chemical methods. In many cases they are the only possibilities to treat complicated chemical problems like bond dissociation and electronic excitations satisfactorily. These methods are, however, computationally very demanding and one can easily come into the situation that one wants to solve a very interesting problem, but that it is simply not manageable with standard computational facilities like workstations or similar equipment. It is the purpose of this project to extend the range of applications for CI methods drastically by using parallel computing techniques. We are using the COLUMBUS program which has been developed by us in cooperation with I. Shavitt from the Ohio State University and R. Shepard from the Argonne National Laboratory. In the previous project P10681-CHE we had developed a very powerful, massively parallel program for the CI section of COLUMBUS. The program runs very efficiently (99% parallel efficiency and more) on 128 to 256 nodes on major parallel computers like the IBM SP2 and the Cray T3E and CI calculations with up to 300 million configurations could be performed and those with one billion configurations are in progress. A major success in the parallelization of COLUMBUS has been achieved. However, in order to complete the whole parallelization project successfully a number of additional steps have to be set. We want to complete the high- quality calculations on the Chromium dimer and on ozone in order to demonstrate the capabilities of the program. The new COLUMBUS program package shall be made available as public domain software. Moreover, there are two major program development steps necessary (doubly-direct CI and internally contracted CI) to make new classes of applications available. The first feature will make calculations on much bigger molecules possible and the second feature will make the representation of the wave function much more effective.

The parallelization of the quantum chemical program package COLUMBUS, which has been developed in collaboration with Ron Shepard from the Argonne National Laboratory and Isaiah Shavitt from the Ohio State University, created new possibilities for precise predictions of molecular energies and other molecular properties. Even nowadays, the efficient use of parallel computer systems is a big challenge since the interplay between individual components such as computer hardware, operating system and application software has to be tuned much better in this case than for single-processor applications. The parallelization was carried out on the basis of the program package "global arrays" developed at the Pacific Northwest Laboratory. This program system was chosen since it allowed flexible access to joint data by individual processes beyond standard synchronous message- passing. Our parallelization work on COLUMBUS aimed at efficient implementations on various parallel computer platforms. Special emphasis was laid on Linux PC clusters. As one of the application, the valence and Rydberg excitations in formaldehyde were investigated. COLUMBUS is available as public domain program package (http://www.itc.univie.ac.at/~hans/Columbus/columbus.html) and currently has more than 180 registered users.