Optimization of the Generalized Valence Bond Method for Vector Processors

Abstract

AbstractThe Generalized Valence Bond [GVB(pp)] multiconfiguration self‐consistent field wave function optimization program has undergone substantial revision in migration from serial scalar computers (like IBM and VAX) to the FPS‐164, a minisupercomputer. Many of the code optimization strategies applied to the GVB program are portable and applicable to a wide range of scientific application codes. In benchmark calculations, the optimized FPS‐164 program, GVB164, is 6.5 times faster than the initial FPS‐164 implementation, both programs storing the two‐electron integral list on disk. GVB164 runs 30 times faster in wall time (limited by the FPS‐164 disk access speed) than a reference version of GVB2P5 on the VAX 11/780. The performance of GVB164 nearly doubles again when all of the two‐electron integrals can be stored in main memory.