high performance computing on graphics processing units: hgpu.org

Over the last few decades, computer modelling and computer simulation have become an invaluable tool for computational chemists interested in advancing their research and experiment in a more efficient, cost effective way with new molecules. As computer capabilities increase the demand for more accurate models and faster simulations has also grown. Some of these models have proved more successful than others with regards to their predictive power, and therefore experienced widespread adoption and support. One of these models in particular, the Merck Molecular Forcefield 94 (MMFF94), has been chosen as a study research subject for this work. The MMFF94 model and its parallelization using multicore and GPU technologies is presented in this work, using as a study frame, the implementation provided in OpenBabel, an open source cheminformatics software, that uses MMFF94 internally to compute the energy of a molecule, among other applications. The work dissects OpenBabel MMFF94 implementation with respect to its parallelization, proposes a software architecture to test and compare between single-threaded, multicore and GPUparallelized versions of MMFF94. Implementation and benchmarking were carried out for Eigen, OpenMP and OpenCL. Results of the benchmarking are discussed in the context of three different applications within OpenBabel: obenergy, obminimize and obconformer. Each of these applications scaling properties are presented together with a discussion on bottlenecks and implementation drawbacks with regard to their parallelization. In the only case where an application performance gain has been achieved (obconformer), the enabling code has been contributed back to the OpenBabel project.