Efficiency of classical molecular dynamics algorithms on supercomputers

Abstract

High performance computing hardware is developed faster than the algorithms for fundamental mathematical models such as classical molecular dynamics are adapted. A wide variety of choice makes it necessary to determine clear criteria based on the computational efficiency of a specific algorithm on a particular hardware. The LINPACK benchmark can no longer serve this purpose. In this paper, we analyze the solution time–peak performance metric based on practical considerations. In this metric, we compare different hardware (both current and obsolete) based on the example of the LAMMPS benchmark, which is widely used for atomistic simulations. It is shown that the considered metric can be used for unambiguous comparison of different combinations of CPUs, accelerators, and interconnection.