Hybrid OpenMP/MPI implementation of the EFMO method in GAMESS
Anastasia O. Gunina, Mark S. Gordon Ames
Laboratory of US DOE/Iowa State University, Ames, Iowa, USA
Steep computational scaling and significant memory requirements of conventional electronic structure methods both limit the size of systems that can be described and hinder transition towards exascale computing. One of the ways to overcome these problems is to use fragmentation approaches. Our approach of choice is the Effective Fragment Molecular Orbital (EFMO) method, which is essentially a combination of the effective fragment potential (EFP) and fragment molecular orbital (FMO) methods: ab initio description is utilized for short-range interfragment interaction, below a selected Rcut, while longrange interactions as well as many-body polarization are described by EFP. For each fragment in its specific geometry, an EFMO calculation requires obtaining EFP parameters, which are generated by a MAKEFP run.
The current EFMO implementation in the GAMESS program package is, however, not optimized nor fully parallelized. As a fragmentation method, EFMO naturally allows the user to put different fragments onto different nodes, creating a need to parallelize within a node for an individual fragment. In this study, a major effort is related to OpenMP parallelization and optimizing the bottleneck MAKEFP step. The resulting performance is demonstrated using a ~1700-atoms piece of a mesoporous silica nanoparticle (MSN).