DualSPHysics_v2.0 用户手册：模拟与流体动力学

需积分: 50 65 浏览量更新于2024-07-23 1 收藏 4.11MB PDF 举报

"DualSPHysics_v2.0_GUIDE 是 DualSPHysics 代码的用户手册，由 University of Vigo、The University of Manchester 和 Johns Hopkins University 的研究人员合作开发。这个项目是 SPHysics 项目的一部分，专注于 Smooth Particle Hydrodynamics (SPH) 方法的应用。该手册发布于2012年，得到了多个机构的支持，包括 Xunta de Galicia、European Regional Development Fund (ERDF)、ESPHI 项目、Research Councils UK (RCUK) 研究奖学金和 EPSRC Grant EP/H003045/1。此外，Orlando Garcia Feal 提供了技术帮助。" 详细说明: DualSPHysics_v2.0 是一个基于 Smooth Particle Hydrodynamics (SPH) 方法的开源软件，用于模拟流体动力学问题。SPH 是一种数值计算方法，用于解决连续介质的力学问题，例如流体流动、结构动态和多体相互作用等。该方法通过将流体或固体离散化为一组移动的质点来近似连续体，从而避免了网格生成的复杂性。在 DualSPHysics_v2.0 用户指南中，用户可以找到关于如何使用该软件进行各种模拟的详细步骤。这份文档可能涵盖了软件安装、基本概念、参数设置、边界条件、求解器的使用以及结果后处理等方面的内容。它还可能包含实例研究，展示如何应用 DualSPHysics 解决实际问题，比如水坝破裂、海洋波浪、火山爆发等。这份手册特别强调了 DualSPHysics 的开发得到了多方面的支持，其中包括 Xunta de Galicia 在不同项目下的资助，以及欧洲社区委员会的 ESPHI 项目和 Marie Curie Actions 的资金支持。此外，Research Councils UK (RCUK) 的研究奖学金和英国工程与物理科学研究理事会 (EPSRC) 的赠款也对项目发展起到了关键作用。在技术层面，Orlando Garcia Feal 对项目的成功贡献了技术支持，这可能涉及到软件开发、优化或问题解决。对于用户来说，这样的支持意味着他们可以在遇到困难时寻求帮助，确保软件的有效使用。 DualSPHysics_v2.0_GUIDE 是一份全面的文档，旨在帮助用户理解和利用 DualSPHysics 进行复杂的流体动力学模拟。它不仅包含了软件的技术细节，还记录了项目背后的研究背景和资助来源，是 SPH 研究者和工程师的重要参考资料。

Figure 3-1. The volume searched using cells of side 2h (left panels) is bigger than using cells of

side h (right panels). Note that symmetry is applied in the interaction.

Multi-core programming with OpenMP

Current CPUs have several cores or processing units, so it is essential to distribute the

computation load among them to maximize the CPU performance and to accelerate the

SPH code. OpenMP is a portable and flexible programming interface whose

implementation does not involve major changes in the code. Using OpenMP, multiple

threads for a process can be easily created. These threads are distributed among all the

cores of the CPU sharing the memory. So, there is no need to duplicate data or to

transfer information between threads.

3.2 GPU Implementation

A GPU implementation initially focused on the force computation since following

Dominguez et al. [2010] this is the most consuming part in terms of runtime. However

the most efficient technique consists of minimising the communications between the

CPU and GPU for the data transfers. If neighbour list and system update are also

implemented on the GPU only one CPU-GPU transfer is needed at the beginning of the

simulation while relevant data will be transferred to the CPU when saving output data is

required (usually infrequently). Crespo et al. [2011] used an execution of DualSPHysics

performed entirely on the GPU to run a numerical experiment where the results are in

close agreement with the experimental results.

The GPU implementation presents some key differences in comparison to the CPU

version. The main difference is the parallel execution of all tasks that can be parallelised

such as all loops regarding particles. One GPU execution thread computes the resulting

force of one particle performing all the interactions with its neighbours. The symmetry

of the particle interaction is employed on the CPU reducing the runtime, but is not

applied in the GPU implementation since it is not efficient due to memory coalescence

issues.

DualSPHysics is unique where the same application can be run using either the CPU or

GPU implementation; this facilitates the use of the code not only on workstations with a

Nvidia GPU but also on machines without a CUDA-enabled GPU. The main code has a

memory access and keep the hardware as busy as possible. The number of active warps

depends on the registers required for the CUDA kernel (defined in the file

DualSPHysics_ptxasinfo), the GPU specifications and the number of threads per block.

Using this optimization, the size of the block is adjusted according to the registers of the

kernel and the hardware specifications to maximise the occupancy.

Simplifying the neighbour search

During the GPU execution of the interaction kernel, each thread has to look for the

neighbours of its particle sweeping through the particles that belong to its own cell and

to the surrounding cells, a total of 27 cells in 3-D since symmetry cannot be applied.

However, this procedure can be optimized simplifying the neighbour search. This

process of neighbour searching can be removed from the interaction kernel if the range

of particles that could interact with the target particle is previously known. Since

particles are reordered according to the cells and cells follow the order of X, Y and Z

axis, the range of particles of three consecutive cells in the X-axis (cell

x,y,z

, cell

x+1,y,z

cell

x+2,y,z

) is equal to the range from the first particle of cell

x,y,z

to the last of cell

x+2,y,z

Thus, the 27 cells can be defined as 9 ranges of particles. The interaction kernel is

significantly simplified, when these ranges are known in advance. Thus, the memory

accesses decrease and the number of divergent warps is reduced.

Figure 3-3. Interaction cells using 9 ranges of three consecutive cells (right) instead of 27 cells

(left). Note that symmetry is not applied in the GPU interaction.

Division of the domain into smaller cells

As described in the optimization applied in the CPU implementation, the procedure

consists in dividing the domain into cells of size h instead of size 2h in order to increase

the percentage of real neighbours. Using cells of size h on the GPU implementation, the

number of pair-wise interactions decreases.

剩余82页未读，继续阅读

枫剑逐雪

粉丝: 0
资源: 1

DualSPHysics_v2.0 用户手册：模拟与流体动力学

DualSPHysics

SPHYSICS流体力学仿真模拟程序的动态链接库编译及C#混合编程方法（源码）

一个公开程序里面的粒子类方法

dualsphysics4.4windows安装

DualSPHysics固体代码是什么

怎么看dualsphysics是否再GPU上运算

dualsphysics怎么提取波高

dualsphysics怎么输出规则波

dualsphysics教程

dualsphysics怎么模拟波浪

最新资源