Efficient parallelization of analytic bond-order potentials for large-scale atomistic simulations

Optimized parallel simulations of analytic bond-order potentials on hybrid shared/distributed memory with MPI and OpenMP

“…As a result, it is necessary to perform significant changes in the main initialization routines to provide the spatial subdivision of the array of atoms for each domain. Considering the use of equal orthogonal domains and the use of Cartesian processor topology given by the MPI implementation (Teijeiro et al, 2016a), the creation of subdomains follows the same approach: every MPI process defines the same internal three-dimensional Cartesian topology, so that every domain is subdivided into equal orthogonal subdomains. In order to provide an efficient processing of path computations, the main atom information associated with each subdomain is stored consecutively and is marked with a start and a stop index.…”

“…Additionally, every domain defines an overlap zone with its neighbour domains, so that every process stores this redundant information in order to perform a local computation of paths. The analysis of different algorithms has proven the possibility of avoiding redundant path computations and minimizing inter-process communications at the same time (Teijeiro et al, 2016a). However, this pure MPI approach exhibits some limitations in flexibility for an efficient exploitation of many-core systems.…”

“…The main idea behind this algorithm is to combine the work distribution between threads with the possibility of overlapping computations and communications specifically for the case of hybrid MPI/OpenMP simulations. Therefore, this implementation is explicitly based on the BONDCOMMS algorithm (Teijeiro et al, 2016a) for MPI parallelization, which defines the outer part of a domain as subject to communications of force contributions with the corresponding neighbour domains. In order to provide the desired functionality, there is an explicit classification of the atoms in every domain in two groups: (1) bulk-domain atoms, which are located in the core of the domain and whose force contributions are fully computed by the domain without any communication, and (2) near-overlap atoms, which are located next to the overlap zone between domains, and whose force contributions should be considered for MPI communications.…”

“…The simulated systems have been similar to previous works (Teijeiro et al, 2016a, 2016b): 3D cubic systems with regular bcc-W crystals (i.e. 5 valence orbitals per atom, which defines 5 × 5 bond matrices for path computations) and periodic boundary conditions on all dimensions.…”

“…Therefore, a possible parallelization can be conveniently based on domain decomposition. According to this, previous works (Teijeiro et al, 2015(Teijeiro et al, , 2016a have presented and analysed several parallel implementations for BOPs with Message Passing Interface (MPI) Forum (2017). However, a pure MPI approach can cause a non-optimal use of computational resources for the case of highly parallel simulations, in which the full memory of a computing node is exhausted.…”