Berlin 2015 – wissenschaftliches Programm

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

MM: Fachverband Metall- und Materialphysik

MM 48: Methods in Computational Materials Modelling VI: Algorithms

MM 48.1: Vortrag

Donnerstag, 19. März 2015, 10:15–10:30, H 0106

Molecular-Dynamics Simulations on Many-Core Processors — •Ralf Meyer and Chris Mangiardi — Laurentian University, Sudbury, Canada

Molecular-dynamics (MD) simulation is one of the most important methods for the numerical studying of materials. This contribution discusses new algorithms for large-scale MD simulations on modern CPUs. Forthcoming many-core processors will soon integrate hundreds of compute cores in a single processor with SIMD units that operate on vectors of 8 or more double-precision numbers simultaneously. However, the full power of these devices will only be accessible with the help of novel algorithms.

The cell task method [1,2] uses a task-based programming approach for the parallelization of MD simulations on multi- and many-core architectures. The method avoids load balancing problems by using a large number of dynamically scheduled small tasks to distribute the workload among the processing cores. Furthermore, a tiling algorithm is used that increases the efficiency of wide SIMD vector units in the simulations.

Results from benchmark simulations on Xeon Phi co-processors are presented. The results show that the cell task algorithm scales well for large numbers of threads. In addition to this, the method outperforms the spatial decomposition approach for simulations of inhomogeneous (e.g. porous) systems.

[1] R. Meyer, Phys. Rev. E 88, 053309 (2013).

[2] R. Meyer, J. Phys.: Conf. Ser. 540, 012006 (2014).