Data parallel optimizations on GPU architectures for molecular dynamic simulations

Molecular dynamic (MD) simulations are algorithmic frameworks that are created to examine and investigate the physical movements of different type of molecules in coherence with the calculation of the forces that are effective over them. Computational power that is available to the simulation was always a barrier for the molecular dynamic simulations. Recently, Graphic Processing Units, originally developed for rendering real-time effects in computer games, started to provide considerable amount of computational power for many applications. Unfortunately, there are many steps that should be investigated in order to adopt molecular dynamic simulations to the GPU architecture. In addition to that, since dividing the molecular data among different computational units to enable concurrent execution is a quite difficult task in the related domain, the execution can only be carried out in single computational unit. In this work, we developed a molecular dynamic simulation that is entirely executed on GPU as well as a Planar Division method which can be used to increase the data parallelism of the simulation. In our benchmarks we observed that GPU implementation is dominating the CPU execution especially on higher workloads. Additionally, the Planar Division Algorithm that we have proposed is quite useful to overcome the algorithmic complexity that might be difficult to manage on huge data sets by dividing the data to different computational units.