Triangulating molecular surfaces over a LAN of GPU-enabled computers

• The first multi-GPU marching cubes algorithm using a distributed OpenMPI–OpenMP–CUDA architecture.
• Triangulating and rendering of molecules with a large number of atoms (millions of atoms).
• A scalable solution to triangulate and render very-large datasets of molecules over a LAN populated with GPUs.

Standalone GPU-enabled computers are adequate to triangulate and rendering molecular datasets with some tens of thousands of atoms at most. But, a standalone GPU-enabled computer has a limited capacity to host programable graphics cards, which in turn have also their constraints in terms of memory space. Thus, in spite of the huge memory space made available and the tremendous processing power of the current GPU-based graphics cards, there remains a scalability problem when it is necessary to triangulate and render big molecules with hundreds of thousands to millions of atoms. In order to overcome this scalability problem we use an OpenMPI–OpenMP–CUDA solution that runs on a loosely-coupled GPU cluster over a LAN (Local Area Network). More specifically, we propose a fast, scalable, parallel triangulation algorithm for molecular surfaces that takes advantage of multicore processors of CPUs and GPUs available over a local network, with each CPU core working as the master of a single GPU. The main contribution of this paper is that likely introduces the first marching cubes algorithm that triangulates molecular surfaces on CUDA devices over a network of GPU-enabled computers.