Multi-GPU Electrostatics and Rigid Bodies in HOOMD-Blue 2.0

We report on the addition of a scalable particle-particle particle-mesh (PPPM) implementation to HOOMD-blue 2.0, which is capable of calculating the long-ranged Coulomb interaction efficiently on several GPUs using spatial domain decomposition and the Ewald summation. The method is based upon a scalable implementation of the Fast Fourier Transform (FFT) in MPI and on the CUDA FFT library (CUFFT) and uses internal block-cyclic data representations and vector radix transforms to speed up the calculation. We also present on a redesigned capability of HOOMD-blue to integrate the equations of motion for rigid bodies composed of smaller particles, which readily scales to many GPUs and allows for easy creation of copies of identical rigid bodies.