Massively Parallel Computing
For calculations made in the previous section, the input files for the
fragments were submitted to one specified queue, and each corresponded
to an independent job ID. Additionally, it is possible to submit MFCC
jobs by interfacing with the LSSCF module of
BDF42,44,86-88 and X-Pol.42 In the
present study, we used GridMol version 2.0 to supply the fragmentation
scheme (e.g., monomers and metal-ligand dimers) and then interfaced with
the X-Pol package for massively parallel calculations using the LSSCF
module in the BDF package. This allowed the hybrid MPI/OpenMP scheme to
be used for accelerating the calculations according to optimized
load-balance and task-tracking. Due to limited computing resources, we
selected only part of 3FMH system, including 1020 atoms, for testing
using PBE and STO-3G as the functional and basis set, respectively.
Table 2 shows the acceleration ratio of 49,152 cores relative to 12,288
cores at 53%, indicating the fragment-based method as a better choice
for large molecule systems according to accuracy and efficiency.