(161c) Using Unified Memory and Nvlink for Node-Level Strong Scaling of Molecular Dynamics Simulations

In this presentation, I will outline recent additions to the HOOMD-blue simulation toolkit, which leverage the parallelism of dense multi-GPU nodes e.g. of the Summit supercomputer to speed up simulations of complex Molecular Dynamics simulations. HOOMD-blue has been extended to take advantage of CUDA unified memory, a virtual address space that allows coherent access between the CPU and several connected GPUs to it, for optimal performance using the recently introduced NVLINK interconnect. We demonstrate that our implementation can achieve an effective speedup of at least 2x over single GPU execution with up to three GPUs per compute node, for systems that include rigid constraints, pair potentials, and electrostatics. This the first demonstration of the benefit of CUDA unified memory and NVLINK for materials simulation, and we will provide science examples made possible through these capabilities.