A distributed-memory parallel lattice Kinetic Monte Carlo algorithm for crystal growth applied to barite (001) face

• Published in: Crystal research and technology (1979) 2016-10, Vol.51 (10), p.575-585
• Main Authors: Abujas-Pereira, Jerónimo, Martin-Bragado, Ignacio, Pina, Carlos M., Pizarro, Joaquín, Galindo, Pedro L.
• Format: Article
• Language: English
• Subjects: Algorithms; Barite; Computer programs; Computer simulation; Crystal growth; Crystal growth simulation; Distributed computing; Distributed memory; Kinetic Monte Carlo; Monte Carlo methods; Software; Surface structure
• ISSN: 0232-1300; 1521-4079
• DOI: 10.1002/crat.201600141

This work presents a parallel approach of the Kinetic Monte Carlo (KMC) algorithm using a distributed memory architecture. The resulting computer software was tested by conducting crystal growth simulations on barite (001) face. Execution times, simulated times and crystallization velocities are compared with a shared memory parallel KMC software (MMonCa). Finally, a ≈ 1 μm2 crystal growth simulation is performed and compared with atomic force microscopy crystal growth experiments. The capability of this approach is demonstrated: a) a significant reduction of parallel overhead is achieved when comparing to the shared memory parallel version of the software, b) a distributed memory approach achieves an increase in memory resources enough to perform simulations with lattice sizes about 1 μm2, allowing the study of larger structures than those in shared memory or sequential implementations, c) this approach should be used only with large scale simulations to take advantage of the distributed memory architecture, d) further improvements are needed for parallel KMC to be faster than serial KMC in small scale simulations, e) the KMC algorithm used is able to adequately simulate two‐dimensional nucleation on large areas of barite (001) faces. A distributed memory approach of the Kinetic Monte Carlo algorithm is presented in this work and applied to crystal growth simulations on barite (001) face. This approach achieves simulation sizes of ≈ 1 μm2, allowing the study of larger structures than those in shared memory or sequential implementations, and also achieves a significant reduction of parallel overhead compared to the shared memory parallel version of the software.