Cyclically parallelized treecode for fast computations of electrostatic interactions on molecular surfaces
Cyclically parallelized treecode for fast computations of electrostatic interactions on molecular surfaces
复制标题
DOI：
10.1016/j.cpc.2020.107742
复制
发表时间：
2021-03
期刊：
影响因子：
--
通讯作者：
Jiahui Chen;Weihua Geng;D. Reynolds
中科院分区：
文献类型：
--
作者：
Jiahui Chen;Weihua Geng;D. Reynolds
我们研究了一种柔性阶笛卡尔树码算法的并行化，用于评估带电粒子系统的静电势，其中N粒子位于生物分子（如蛋白质）的分子表面。当满足良好分离条件时，treecode算法使用远场Taylor展开计算O （N log N）个粒子-簇相互作用来代替O （n2）个粒子-粒子相互作用。该算法采用消息传递接口（Message Passing Interface， MPI）标准，通过在每个任务的内存中创建相同的树结构来实现并行计算。我们设计了一种循环顺序方案，将空间封闭的目标粒子均匀分布到所有可用的任务中，显著提高了并行负载平衡。我们还研究了并行效率受三码参数的影响，如泰勒展开阶p、每叶最大粒子数N 0和最大可接受准则θ。这种循环并行的树码可以解决多达数千万个粒子之间的相互作用。然而，如果问题大小超过每个任务的内存限制，则可以使用使用正交递归对分（ORB）树的可扩展域分解（DD）并行树码来代替，除了有效地计算带电粒子的n体问题外，我们的方法还可以加速GMRES迭代求解边界积分泊松-玻尔兹曼方程。
We study the parallelization of a flexible order Cartesian treecode algorithm for evaluating electrostatic potentials of charged particle systems in which N particles are located on the molecular surfaces of biomolecules such as proteins. When the well-separated condition is satisfied, the treecode algorithm uses a far-field Taylor expansion to compute O (N log N) particle–cluster interactions to replace the O (N 2) particle–particle interactions. The algorithm is implemented using the Message Passing Interface (MPI) standard by creating identical tree structures in the memory of each task for concurrent computing. We design a cyclic order scheme to uniformly distribute spatially-closed target particles to all available tasks, which significantly improves parallel load balancing. We also investigate the parallel efficiency subject to treecode parameters such as Taylor expansion order p, maximum particles per leaf N 0, and maximum acceptance criterion θ. This cyclically parallelized treecode can solve interactions among up to tens of millions of particles. However, if the problem size exceeds the memory limit of each task, a scalable domain decomposition (DD) parallelized treecode using an orthogonal recursive bisection (ORB) tree can be used instead In addition to efficiently computing the N-body problem of charged particles, our approach can potentially accelerate GMRES iterations for solving the boundary integral Poisson–Boltzmann equation.