Scalable Evaluation of Polarization Energy and Associated Forces in Polarizable Molecular Dynamics: I. Toward Massively Parallel Direct Space Computations
Scalable Evaluation of Polarization Energy and Associated Forces in Polarizable Molecular Dynamics: I. Toward Massively Parallel Direct Space Computations
复制标题
DOI：
10.1021/ct401096t
复制
发表时间：
2014-04-01
影响因子：
5.5
通讯作者：
Piquemal, Jean-Philip
中科院分区：
文献类型：
--
作者：
Lipparini, Filippo;Lagardere, Louis;Piquemal, Jean-Philip
在本文中，我们研究了各种数值策略来计算的直接空间极化能量和相关的力量的上下文中的点偶极子近似（包括阻尼）用于极化分子动力学。我们提出了一个仔细的数学分析的算法，已实现在流行的生产包，并适用于大型测试系统。我们表明，不应使用经典的Jacobi过度松弛方法（JOR），因为它的收敛需要适当的松弛参数值，而应首选其他策略。在一个单一的。节点，预条件共轭梯度法（PCG）和雅可比算法加上迭代子空间直接反演（JI/DIIS）提供可靠的稳定性/收敛性，大约是JOR的两倍。此外，这两种算法都适合大规模并行实现。在进程通信方面的较低要求使JI/DIIS成为MPI和混合型OpenMP/MPI范例用于真实的寿命测试的首选方法。此外，使用预测步骤作为猜测沿着分子动力学模拟提供了另一种廉价但非常有效的收敛加速形式。总体而言，与初始JOR单节点方法到最终PGC或JI/DIIS并行方法结合相比，可以获得两到三个数量级的时间。预测器MD细化。这样的加速跟踪极化分子动力学的高性能实现的新路线，因此扩展了该技术的适用性，因为它将有利于未来的多尺度QM/MM/连续计算。
In this paper, we investigate various numerical strategies to compute the direct space polarization energy and associated forces in the context of the point dipole approximation (including damping) used in polarizable molecular dynamics. We present a careful mathematical analysis of the algorithms that have been implemented in popular production packages and applied to large test systems. We show that the classical Jacobi Over-Relaxation method (JOR) should not be used as its convergence requires a proper value of the relaxation parameter, whereas other strategies should be preferred. On a single. node, Preconditioned Conjugate Gradient methods (PCG) and Jacobi algorithm coupled with the Direct Inversion in the Iterative Subspace (JI/DIIS) provide reliable stability/convergence and are roughly twice as fast as JOR. Moreover, both algorithms are suitable for massively parallel implementations. The lower requirements in terms of processes communications make JI/DIIS the method of choice for MPI and hybrid OpenMP/MPI paradigms for real life tests. Furthermore, using a predictor step as a guess along a molecular dynamics simulation provides another inexpensive, yet very effective, form of convergence acceleration. Overall, two to three orders of magnitude in time can be gained compared to the initial JOR single node approach to the final PGC or JI/DIIS parallel one combined with the. predictors MD refinements. Such a speedup traces a new route for the high performance implementation of polarizable molecular dynamics and therefore extends the applicability of the technique as it will facilitate future multiscale QM/MM/continuum computations.