利用GPU实现大规模复杂体系反应分子动力学模拟的方法

Quantum chemistry (mechanics) is currently the basic method for understanding reactive events in molecular simulation, but it is still extremely computationally expensive that the modeling problem size is usually limited within hundreds of atoms on a single computation node at acceptable computational costs. ReaxFF is a reactive force field with a novel bond order potential by explicitly modeling bond activity and charge equilibration that makes it potentially a powerful tool for modeling complex molecular systems with chemical reaction. ReaxFF based molecular dynamics (MD) simulation increases the ability to modeling reactive systems containing over 1000 atoms with faster computing speed compared with quantum chemistry program. However, ReaxFF MD computation is tens times slower when compared with conventional MD simulation. Computational implementation on a single computational node for better performance is playing a crucial role for ReaxFF MD applications in meeting the challenge of handling complex reactive atomistic system that is more realistic for practical applications. ReaxFF is much more demanding for computing performance and memory of computer hardware than conventional molecular force fields. The outstanding theoretical Floating-Point Operations per Second and memory band width of GPU (graphics processing unit) over CPU provides new possibilities for improving computation implementation of ReaxFF MD. This proposal aims at accelerating ReaxFF MD computation by taking advantage of GPU on single computation node. To achieve high computing efficiency, the computational tasks of ReaxFF will be decomposed and mapped to GPU threads. The computational implementation will be optimized by careful hierarchical organization of threads and memory to balance the instruction throughput with memory throughput, and by proper decomposition of computing tasks between CPU and GPU on a single node for maximized parallel execution. The proposed approach to improve computational implementation of ReaxFF MD on single computing node with GPU is a novel one, which should of help in applying ReaxFF MD to more complex reactive atomistic systems such as coal pyrolysis that can hardly handled by quantum chemistry computation.

目前量子化学计算是分子模拟中可探索化学反应机理的主要方法，但计算极其耗时，单计算节点适宜模拟的体系规模约~100个原子，无法应用于复杂体系。基于ReaxFF的反应分子动力学方法(ReaxFF MD)将单节点适宜模拟的规模提高到>1000个原子、计算速度也显著高于最快的量子化学方法，但应用于高度复杂的体系如煤、高分子、生物质时，其计算规模和效率仍然是最大的瓶颈。ReaxFF算法对计算性能和存储器要求远高于其他常用分子力场，本项目拟利用GPU突出的理论计算峰值和存储器带宽解决这一棘手问题。通过将计算任务映射为大规模GPU线程、精细控制计算在CPU和GPU之间的划分与协调、细致平衡计算吞吐与数据吞吐，将单节点ReaxFF MD模拟的适宜规模提升为>10000原子，并应用于煤热解复杂反应机理的探索。利用GPU加速ReaxFF MD模拟的思路创新，可望使分子层次、高度复杂体系的化学反应模拟成为可能

反应分子动力学(ReaxFF MD)可模拟分子模型中的化学反应，计算速度和规模显著高于应用广泛的DFT方法，且计算准确度接近于DFT。特别是无需预先定义反应路径，使其有潜力研究复杂反应过程如煤热解。但其计算性能成为应用的瓶颈。由于反应MD模拟的时间步长比经典MD的步长小一个量级(0.1 fs)，且每个时间步都要对原子电荷进行动态优化，加上为平滑描述原子成断键引入的势能校正，ReaxFF MD的计算速度比经典MD慢数十倍。本项目的研究内容是探索基于GPU提高单一计算节点上ReaxFF MD计算性能的方法，目标是实现~10000原子规模煤热解的ReaxFF MD高效模拟。.针对ReaxFF MD对计算能力和数据存储能力均有高要求的特性，为了充分利用GPU大规模线程并行的优势，本项目将ReaxFF MD的主要计算置于GPU端，CPU端主要完成多GPU并行时计算任务的划分。对于共轭梯度法求解动态电荷的瓶颈步骤，采取了T-thread优化线程调度；对双精度的浮点原子加和瓶颈则采用了规约与原子加和计算量的权衡优化；对基于MPI多GPU并行的通信瓶颈则优化初始区域分解的均衡性，并尽量减少需要通信的原子“边界宽度”。本项目对单卡所有计算步骤的精细优化相当充分。所建立的GMD-Reax程序，采用国际上著名的LAMMPS上的CPU代码8核运行性能为基准进行了计算性能测试，对于1378 - 27283原子的煤模型而言，单GPU C2050卡的双精度版本计算速度是优化充分的C程序的1.5–4倍，是FORTRAN程序的3 - 7.8倍；与国际上仅有的另一个GPU程序(PuReMD-GPU)发表的计算性能相当。在K40的双精度计算性能比C2050可提升1.7-2.2倍。GMD-Reax可实现大规模煤分子模型(~10,000原子)热解模拟的高效计算，单GPU卡运行是性价比最好的模拟运行模式。对万原子规模煤模型的ReaxFF MD模拟应用表明，大规模模拟可获得微观尺度煤热解过程本征反应随温度和时间演化的全景式图景，是一种研究煤热解复杂过程的有前景的新方法。.特别是GMD-Reax的模拟应用已超出项目计划，所建立的GMD-Reax是一个通用的程序，模拟应用已扩展至生物质热解、高分子热解、生物油氧化、燃料的热解与燃烧、工业废水中有机物的深度脱除机理等复杂体系的反应机理研究。