Quantum Molecular Dynamics in the Post-Petaflops Era

Nichols A. Romero, Aiichiro Nakano, Katherine M. Riley, Fuyuki Shimojo, Rajiv K. Kalia, Priya Vashishta, Paul C. Messina
2015 Computer  
As the scale of quantum molecular dynamics simulations has grown in time and system size, QMD codes must increase intranode and instruction-level parallelism to take advantage of emerging supercomputer architectures. The authors present one promising parallelization approach and illustrate its success on one of the world's most powerful systems. Q uantum molecular dynamics (QMD) simulations are a critical tool in various scienti c domains including condensed-matter physics, materials science,
more » ... emistry, and biology. As the scale of QMD simulations has grown beyond the canonical metrics of time and system size, there has been a paradigm shift to more "naturally" parallel methods such as divide-and-conquer, equilibrium and nonequilibrium statistical sampling, and high-throughput screening. The scienti c community has also needed to adapt to the short ve-year life cycle of new computer architectures. We explore how QMD simulations have evolved over the years and then focus on a novel approach that extends divide-and-conquer. We also present two examples of successes achieved with this method on Mira (www.alcf.anl.gov/mira), the -petaflops IBM Blue Gene/Q system at Argonne National Laboratory (ANL) currently ranked fifth on the TOP supercomputer list.
doi:10.1109/mc.2015.337 fatcat:rxm2uiybgjdhpfppnkvdwk5fbu