GPU Algorithm Of The LBM And Its Application In Hemodynamics Of Cerebral Aneurysm

Cerebral aneurysms are localized pathological dilatation of arterial walls, they are par-ticularly dangerous for the high morbidity and mortality when they rupture. A new therapy to treat the aneurysm is implanting a porous stent across the neck of the aneurysm, which is viewed as a promising and minimally invasive treatment modality. However, the effect of the stent is affected by many factors such as the shape of the aneurysm and the structure of the stent, thus need to be thoroughly investigated based on patient-specific aneurysm geome-tries. For this kind of problem, clinical observation, animal experiment and in vitro study are greatly restricted. On the other hand, hemodynamics is thought to play an important role in the pathogenesis and treatment of cerebral aneurysms. Thus numerical simulation is a powerful and efficient tool to study the effect of the stent by investigating the change in hemodynamic characteristics of blood flow in the aneurysm. However, the blood properties, the complex geometry of vessel, as well as large-scale computation limit traditional nu-merical methods in studying the hemodynamics of cerebral aneurysms. Therefore, efficient computational method and advanced parallel computing technology are needed.The lattice Boltzmann method (LBM) is a novel method for simulating fluid flow and modeling physics in fluids, and has been proved to be powerful in many areas. The LBM has many advantages, such as fully parallelism, easy implementation of boundary conditions, and simple coding. Based on these advantages, the LBM is quite suitable for modeling and simulation of blood flow in cerebral aneurysms. In addition, the LBM is quite suitable for large-scale parallel computing. At the same time, GPU (graphics processing unit) computing is becoming rapidly popular in recent years. This is mainly due to the fact that, for many applications, the GPU based algorithm could achieve a higher performance, usually at least one order of magnitude faster than that of the CPU based algorithm. The fully parallelism of the LBM makes it quite compatible with GPU for large-scale parallel computation, and GPU based LBM could achieve rather remarkable performance.However, the GPU based LBM has not been studied extensively and thoughtfully, es-pecially for simulations of fluid flow in complex geometries such as blood flow. In addition, for LBM based hemodynamic studies of cerebral aneurysms, only two-dimensional or ide-alized aneurysm model are considered in previous work. In view of this, this thesis will first further develop the GPU based LBM, and then employ it to study hemodynamic character-istics of blood flow in patient-specific cerebral aneurysms. This thesis is composed of two parts: (1) In term of GPU based LBM, the LBM is firstly implemented on GPU using the compute unified device architecture (CUDA), then the role of memory access optimization and other optimization technologies in programming and the performance of algorithm are analyzed in detail. The performance of the presented algorithm is higher than those proposed in previous work, and compared with CPU algorithm, GPU algorithm can achieve a speedup close to50using Tesla C1060. Secondly, based on the characteristic of complex geometries and the structure of the GPU, two algorithms utilizing the common full matrix mode and sparse matrix mode are presented respectively, and then the performance of the two algo-rithms are detail compared for fluid flow in various complex geometries. The sparse matrix algorithm could achieve a MLUPS above250, which is higher than the results reported in previous work. Finally, the algorithms are implemented on multi-GPU to gain a higher per-formance and to overcome the limitation of the GPU memory size. On a single-node cluster equipped with four Tesla C1060, the performance of full matrix algorithm shows a linear increase with the number of GPUs, and for the sparse algorithm, a parallel efficiency close to90%is achieved.(2) In term of hemodynamic investigation of cerebral aneurysm, the non-Newtonian effect on hemodynamic characteristics of blood flow is first studied. Previous researches usually regard the blood flow as Newtonian flow, but this assumption is questionable for blood flow in aneurysms, especially in stented aneurysms. In this thesis, hemodynamic characteristics derived with Newtonian and non-Newtonian models are studied, and com-pared in detail. The results show that the non-Newtonian effect gives a great influence on hemodynamic characteristics of blood flow in stented cerebral aneurysm, especially in small necked ones. Beside this, the thesis presents a pipeline of blood flow simulation in patient-specific cerebral aneurysm, including three-dimensional reconstruction of the aneurysm and virtual deployment of the stent. Then numerical studies based on patient-specific cerebral aneurysm were performed, and found that the small-necked aneurysm implanted with high porosity stent is sensitive to the position offset of the stent; the stent with a lower porosity in the proximal aneurysm neck could affect the aneurysmal inflow more significant than the stent with a lower porosity in the distal aneurysm neck.In conclusion, the GPU based LBM is systematically studied in this thesis, especially for fluid flow in complex geometries. Meanwhile, hemodynamic studies on blood flow in cerebral aneurysm are performed via the LBM, and extend to realistic patient-specific cerebral aneurysm model. The work can be viewed as a necessary basis for future studies.