基于GPU的分子动力学模拟并行化及实现

Parallel Algorithm and Implementation for Molecular Dynamics Simulation Based on GPU

分子动力学模拟作为获得液体、固体性质的重要计算手段,广泛应用于化学、物理、生物、医药、材料等众多领域。模拟体系的复杂性和精确性的需求,使得计算量巨大,耗费时间长。并行计算是加速大规模分子动力学模拟的重要途径。GPU以几百GFlops甚至上TFlops的运算能力,为分子动力学模拟等的计算密集型应用提供了新的加速方案。提出了一种基于GPU的分子动力学模拟并行算法——oApT-AD,并在OpenCL和CUDA框架下加以实现。性能测试显示,在Tesla C1060显卡上,该算法在OpenCL框架下的实现相对于CPU的串行实现,最高达到120倍加速比。通过对比发现,该算法在CUDA上的性能与OpenCL基本相当。同时,该算法还可以扩展到两块及以上的GPU上,具有良好的可扩展性。

Molecular Dynamics Simulation is an important method for acquiring liquid and solid atoms＇ properties.This method has been widely used in the fields of chemistry,physics,biology,medicine and materials.The complexity and accuracy demand causes enormous workloads.Parallel computing is a feasible way to speedup large-scale molecular dynamics simulation.With hundreds of GFlops or even TFlops performance,GPU can speed up computing-intensive applications.This paper presented a parallel algorithm named oApT-AD,and we implemented it on GPU under OpenCL and CUDA Framework.The experiment results show that the oApT-AD algorithm can achieve 120 speedup on GPU Tesla C1060 under OpenCL Framework,compared to that on CPU.And we also implemented the oApT-AD algorithm on GPU under CUDA Framework.The implement under OpenCL Framework provides almost the same performance as the implement under CUDA Framework.Moreover,our algorithm can be extended to two or more GPUs,with good scala-bility.