Regensburg 2016 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

CPP: Fachverband Chemische Physik und Polymerphysik

CPP 28: Poster: Computational Physics of Soft Matter

CPP 28.1: Poster

Dienstag, 8. März 2016, 18:15–21:00, Poster B2

Dimensionality aware method for the domain decomposition of homogeneous and inhomogeneous molecular dynamics simulations — •Horacio V. Guzman, Torsten Stuehn, and Kurt Kremer — Ackermannweg 10, 55128 Mainz

Molecular simulations are a scientific tool that has enabled theoretical and computational methods to tackle soft matter research. This is reflected in the permanent development and enhancement of new methodologies within scientific codes[1,2,3]. One of the challenges in molecular dynamics simulations is to improve efficiency in the parallelization of domain decomposition schemes. Here, we present a new iterative method to optimize the decomposition of the simulation box which combines the dimensionality of the system and the level of parallelization to be used. In addition the proposed method also includes a shifting domain walls feature for an adpatative load balancing of inhomogeneous systems[4]. We have implemented the method in ESPResSo++[1] which allowed us to benchmark both homogeneous[5] and inhomogeneous systems. By comparing the existing domain decomposition scheme[1] with the new dimensionality aware method, we prove that the latter enables a computationally more efficient domain decomposition for both homogeneous and inhomogeneous systems, and hence a speedup increase, as well as, a higher upper bound scaling. [1] J. D. Halverson et al., Compt. Phys. Comm. 184 (2013) [2] W. M. Brown et al., Comput. Phys. Comm. 195 (2015) [3] B. Hess et al., J. Chem. Theory Comput. 4 (2008) [4] M. Praprotnik et al., Phys. Rev. E 73 (2006) [5] L. Moreira et al., Macromol. Theory Simul. 24(2015)