Accelerating the molecular dynamics program LAMMPS using graphics cards' processors and the Nvidia Cuda technology

摘要

This bachelor thesis discusses the possibilities and limits of extending the widely-used Molecular Dynamics program LAMMPS (the Large Scale Atomic/Molecular Massive Parallel Simulator) in order to accelerate simulations by making use of modern graphic processing units in the framework of NVIDIA's Cuda technology. An own implementation serves as basis for the discussion. The first chapters summerizes selected aspects of Molecular Dynamics, LAMMPS' design and the Cuda technology for graphics card programming. On that basis, details of the above mentioned implementation are discussed, comparing different force calculation strategies and showing methods to avoid frequent data transfers from and to the graphics card. The following part of the work deals with benchmarks and the influence of factors like numerical precision and system size in typical Molecular Dynamics simulations, always comparing the results to the original version of LAMMPS with no graphics card support. Finally, conclusions of this work are presented and the implementation is compared with other similar projects.