Multi-Level Parallelization of the Fragment Molecular Orbital Method in GAMESS

The parallelization of the fragment molecular orbital (FMO) method implemented in GAMESS is reviewed for Cray/Intel, IBM Blue Gene, and Fujitsu supercomputers. Various strategies for load balancing are compared. A basic OpenMP implementation of FMO is described and its parallel efficiency is reported. The two-level generalized distributed data interface is extended to an arbitrary number of levels and applied using three levels to an FMO Hessian calculation at the level of second-order Moller–Plesset (MP2) theory.