LBM的GPU算法及其在颅内动脉瘤血流动力学中的应用

发布时间：2018-08-27 10:17

【摘要】：颅内动脉瘤是一种常见的颅内凶险疾病,其破裂后产生的蛛网膜下腔出血具有极高的致残率和致死率。支架置入术作为一种治疗颅内动脉瘤的最新方法受到广泛关注,该方法具有微创性、恢复快、疗效好等优点。但是,支架的治疗效果受到动脉瘤形状、支架结构等多种因素的影响,需要针对个体真实动脉瘤模型进行深入研究。对于此类问题,现有的临床观测、动物实验和体外实验等研究手段受到了很大的限制。另一方面,血流动力学被普遍认为是影响动脉瘤发生、生长、破裂及治疗的主要因素之一,因而利用数值模拟研究动脉瘤内血流动力学参数的变化是探讨支架作用的极佳选择。传统数值方法在研究颅内动脉瘤血流动力学方面会受到血液流体特性、血管复杂几何结构以及大规模计算量等诸多瓶颈的限制,因此需要发展高效的计算方法和先进的并行计算技术。 20世纪80年代后期发展起来的介观格子Boltzmann方法(LBM)兼有微观与宏观模型的优点,在模拟流体的相互作用和处理复杂边界等方面有着传统方法难以比拟的优势,特别适合用于研究颅内动脉瘤内的血流流动。此外,LBM具有计算效率高且天然并行等优点,特别适合利用先进的并行计算技术进行加速。与此同时,基于图形处理器(GPU)的并行技术近几年来获得迅猛的发展。相比基于传统CPU的应用,基于GPU的应用可获得一到两个数量级的加速效果。LBM的天然并行性使得其与GPU具有良好的匹配性,基于GPU的LBM可获得较理想的加速效果。但是,目前基于GPU的LBM研究还未够深入,特别是对于血液流这种复杂流场内流动问题的LBM的实现与优化缺乏系统深入的讨论。此外,基于LBM的颅内动脉瘤血流动力学相关研究中,未实现对置入支架的个体化真实动脉瘤的数值模拟。鉴于此,本文将首先进一步发展基于GPU的LBM,在此基础上,展开基于医学图像的个体化真实颅内动脉瘤血流动力学研究。论文主要工作包括以下两个部分： (1)在基于GPU的LBM方面,本文首先以方腔流等简单流场为例,着重探讨存储器访问优化等优化技术的作用,并对程序的性能进行了详细分析,探讨了在GPU处理器上影响LBM程序性能的因素。所设计的算法性能优于国际上相关算法,与经过充分优化的CPU程序相比,在Tesla C1060上最高可达50倍加速比。在此基础上,针对不同复杂流场的特点,分别基于完全矩阵方法和稀疏矩阵方法提出了两种算法,并通过大量的数值模拟分析各自的适用工况。其中稀疏矩阵算法在Tesla C1060上取得250以上MLUPS(每秒百万网格更新),优于国际上相关结果(170MLUPS)。最后,将设计的算法进一步推进至多GPU平台,对数据访存和通信过程等方面进行优化。在配备4块Tesla C1060的平台上,完全矩阵算法可取得接近100%的并行效率,稀疏矩阵算法可取得接近90%并行效率。 (2)在颅内动脉瘤血流动力学方面,针对已往大部分研究将血液简化假设为牛顿流体的现状,本文利用不同的动脉瘤模型、不同孔隙率的支架等工况来对比牛顿模型与非牛顿模型得出的结果,以此探讨这种简化假设的合理性,分析非牛顿效应对研究结果的影响。研究结果表明,窄颈动脉瘤和置入低孔隙率支架的动脉瘤更易受到非牛顿效应的影响。在此之后,本文从医学图像出发,通过动脉瘤模型的三维重建、支架的虚拟配置等处理,实现个体化真实动脉瘤的数值模拟,并对相关血流动力学问题展开了研究,发现支架置入位置对窄颈动脉瘤和置入高孔隙率支架的动脉瘤有较大影响,血流进入瘤内方向发生明显改变；发现在动脉瘤瘤颈近端对支架加密可达到较好效果,与整体加密的支架功效相近。总之,本文推进了基于GPU的LBM的相关研究,对复杂流场内流动问题的模拟进行了细致而深入的优化。同时,本文利用LBM对颅内动脉瘤血流动力学相关问题展开了研究,实现了基于个体化真实动脉瘤模型的数值研究。
[Abstract]:Intracranial aneurysms are a common and dangerous intracranial disease. Subarachnoid hemorrhage after rupture of intracranial aneurysms has a high morbidity and mortality rate. Stent implantation as a new treatment of intracranial aneurysms has attracted wide attention. This method has the advantages of minimally invasive, rapid recovery and good curative effect. To study the effect of aneurysm shape, stent structure and other factors, it is necessary to study the real aneurysm model in detail. One of the main factors in rupture and treatment is the use of numerical simulation to study the hemodynamic parameters in aneurysms, which is the best choice to explore the role of stents. Therefore, it is necessary to develop efficient computing methods and advanced parallel computing technology.
The mesoscopic lattice Boltzmann method (LBM), developed in the late 1980s, has the advantages of both microscopic and macroscopic models. It has many advantages over traditional methods in simulating fluid interactions and dealing with complex boundaries. It is especially suitable for studying blood flow in intracranial aneurysms. At the same time, parallel technology based on graphics processor (GPU) has developed rapidly in recent years. Compared with traditional CPU-based applications, GPU-based applications can achieve speedup of one to two orders of magnitude. LBM's natural parallelism makes it possible to speed up with GPU. With good matching, LBM based on GPU can get an ideal acceleration effect.
However, the research of LBM based on GPU has not been thorough enough, especially the realization and optimization of LBM in complex flow field. In addition, in the related research of intracranial aneurysms hemodynamics based on LBM, the numerical simulation of individual real aneurysms implanted with stents has not been realized. In this paper, firstly, we will develop a GPU-based LMB, and on this basis, we will launch a personalized real intracranial aneurysm hemodynamics research based on medical images.
(1) In the aspect of GPU-based LBM, firstly, taking the simple flow field such as square cavity flow as an example, this paper focuses on the role of memory access optimization and other optimization techniques, analyzes the performance of the program in detail, and discusses the factors that affect the performance of the LBM program on the GPU processor. Compared with the optimized CPU program, the maximum acceleration ratio can reach 50 times on Tesla C1060. On this basis, two algorithms based on complete matrix method and sparse matrix method are proposed for different characteristics of complex flow field, and a large number of numerical simulations are carried out to analyze their applicable conditions. The sparse matrix algorithm is selected on Tesla C1060. More than 250 MLUPS (millions of grid updates per second) are obtained, which is superior to the relevant international results (170MLUPS). Finally, the algorithm is further promoted to the multi-GPU platform to optimize the data access and communication process. On the platform equipped with four Tesla C1060 blocks, the complete matrix algorithm can achieve nearly 100% parallel efficiency and the sparse matrix algorithm. You can get close to 90% parallel efficiency.
(2) In terms of hemodynamics of intracranial aneurysms, in view of the fact that most previous studies have simplified blood as Newtonian fluid, this paper compares the results of Newtonian model and non-Newtonian model by using different aneurysm models and stents with different porosity, so as to explore the rationality of this simplified hypothesis and analyze non-Newtonian model. The results show that aneurysms with narrow carotid artery and those with low porosity stents are more susceptible to non-Newtonian effects. From this point of view, we simulate the real aneurysms individually by three-dimensional reconstruction of aneurysm model and virtual placement of stents. Relevant hemodynamic problems were studied. It was found that stent placement had a great effect on narrow carotid aneurysms and high porosity stent placement aneurysms, and the direction of blood flow into aneurysms changed significantly.
In a word, this paper advances the research of LBM based on GPU, and optimizes the simulation of complex flow field in detail. At the same time, this paper uses LBM to study the related problems of intracranial aneurysm hemodynamics, and achieves the numerical research based on the individualized real aneurysm model.
【学位授予单位】：华中科技大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：R739.41

【参考文献】