Parallel computing in quantum chemistry -- Message passing and beyond for a general ab initio program system

One of the most prominent aims in Computational Chemistry is the modeling of chemical reactions and the prediction of molecular properties. Quantum chemical methods are used for the calculation of molecular structures, spectra, reaction energy profiles and many other interesting quantities. Nowadays, the accuracy of the theoretical calculations can compete to an increasing extent with the experimental one. A great variety of quantum chemical methods exist ranging from the standard Hartree-Fock theory to sophisticated electron correlation approaches. From a computational point of view all these methods require rather lengthy and complicated program codes and have to handle a large amount of data to be stored on external devices. In the simplest case, the Hartree-Fock (SCF) method, ``direct`` algorithms have eliminated the I/O and storage bottleneck and have opened the way to parallel implementations. For post-Hartree-Fock methods the situation is much more complicated as will be demonstrated below. Therefore, most of the previous attempts in parallelizing quantum chemical ab initio programs concentrated on SCF methods. The authors investigations presented here are a continuation of their previous work on the parallelization of the COLUMBUS program system. The COLUMBUS program is based on the multireference single- and double-excitation configuration interaction (MRSDCI) approach. It is very portable and runs on a large variety of computers including numerous Unix-based workstations, VAX/VMS minicomputers, IBM mainframes and Cray supercomputers