Performance of inverse atomistic scale fracture modeling on GPGPU architectures

The present work has been motivated by the continuous growth of General Purpose Graphic Processor Unit (GPGPU) technologies as well as the necessity of linking usability with multiscale materials processing and design. The inverse problem of determining the phenomenological interparticle Lenard-Jones potential governing the fracture dynamics of a two dimensional structure under tension, is used to examine the feasibility and efficiency of utilizing GPGPU architectures. The implementation of this inverse problem under a molecular dynamics framework provides verification of this methodology. The main contribution of this paper is a performance evaluation driven sensitivity analysis that is contacted on GPGPU-enabled hardware in order to examine efficiency relative to various combinations of GPGPU and Central Processing Unit (CPU) cores as a function of problem size. In particular, speedup factors are determined relative to various number of core combinations of a quad core processor.

Research highlights
► Demonstrated the implementation and feasibility of using a MD framework for solving an inverse problem on both conventional multi-core CPUs and GPGPU hardware for a 2D fracture problem on the atomic level.
► Two inverse problems were solved: One determining both L-J and lattice parameters, while the other two L-J parameters.
► Numerical experiments demonstrated a 8-fold higher performance of the one GPU over all other combinations for GPGPU and multi-core configuration as expected.
► Using more GPUs across host bus did not scale up as one would desire, while using multiple-cores of the Intel i7 processor scaled very well.