A04: Optimization of the AIREBO Many-Body Potential for KNL
Supervisor: Paolo Bientinesi (RWTH Aachen University)
Abstract: Molecular dynamics simulations are an indispensable research tool for computational chemistry and material science. Empirical many-body potentials promise high-fidelity simulations that capture bonding and reaction behavior accurately, providing a level of detail in between more classical molecular dynamics and quantum methods.
The AIREBO potential is one such example that provides forces and energies for molecular dynamics (MD) simulations of carbon and carbohydrate structures. Allowing many-body potentials to profit from the recent architectural advances still poses a challenge due to deeply nested, short loops. We develop an optimized, vectorized AIREBO implementation for Intel's Xeon and Xeon Phi (co)processors and integrate it into the open-source LAMMPS molecular dynamics code. By both introducing improvements to the code and vectorization, we achieve a sustained real-word speedup of two on Broadwell, and a speedup of four on KNL. The optimized code will be distributed with each LAMMPS download as part of the USER-INTEL package.
ACM-SRC Semi-Finalist: no
Two-page extended abstract: pdf