Finite difference iterative solvers for electroencephalography: serial and parallel performance analysis
Finite difference iterative solvers for electroencephalography: serial and parallel performance analysis
复制标题
DOI：
10.1007/s11517-008-0344-9
复制
发表时间：
2008-09-01
影响因子：
3.2
通讯作者：
Ng, Kwong T.
中科院分区：
文献类型：
--
作者：
Barnes, Derek N.;George, John S.;Ng, Kwong T.
目前脑电（EEG）研究中使用的头部模型的分辨率受到正向求解器速度的限制。在这里，我们提出了一个并行有限差分技术，可以减少解决时间的泊松方程的头部模型。多个处理器用于同时处理问题，以加快解决方案并为解决大型问题提供内存。原始计算域被划分为多个矩形分区。然后，每个分区被分配给一个处理器，该处理器负责与该特定分区中的节点相关联的所有计算和处理器间通信。由于正向求解时间主要花在求解相关的矩阵方程上，因此需要找到最佳矩阵求解器。对从MRI图像构建的各向同性和各向异性真实头部模型进行了各种迭代求解器的详细比较。共轭梯度法（CG）预处理与先进的几何多重网格技术被发现提供最佳的整体性能。对于具有256 x 128 x 256个单元的各向异性模型，该技术在32个处理器上提供了508的串行CG解决方案的加速比，通过多重网格预处理和并行化分别提供了20.1和25.3的加速比。
Currently the resolution of the head models used in electroencephalography (EEG) studies is limited by the speed of the forward solver. Here, we present a parallel finite difference technique that can reduce the solution time of the governing Poisson equation for a head model. Multiple processors are used to work on the problem simultaneously in order to speed up the solution and provide the memory for solving large problems. The original computational domain is divided into multiple rectangular partitions. Each partition is then assigned to a processor, which is responsible for all the computations and inter-processor communication associated with the nodes in that particular partition. Since the forward solution time is mainly spent on solving the associated matrix equation, it is desirable to find the optimum matrix solver. A detailed comparison of various iterative solvers was performed for both isotropic and anisotropic realistic head models constructed from MRI images. The conjugate gradient (CG) method preconditioned with an advanced geometric multigrid technique was found to provide the best overall performance. For an anisotropic model with 256 x 128 x 256 cells, this technique provides a speedup of 508 on 32 processors over the serial CG solution, with a speedup of 20.1 and 25.3 through multigrid preconditioning and parallelization, respectively.