high performance computing on graphics processing units: hgpu.org

This thesis deals with a particular problem out of the research field of computational fluid dynamics, the numerical simulation of fluids containing soluted rigid particles. Such problems arise within a variety of applied sciences, such as medicine, ecology and engineering and need to be studied in detail in three-dimensions. So far most scientific publications on this topic either dealt with 2D phenomena only or the behavior of a very few particles in 3D. The reason for this being most of all performance constraints of the used simulation techniques and hardware. This thesis aims to accomplish a high-performance simulation of a large number of particles in 3D by using co-processor technology. The landscape of such accelerators is reviewed thoroughly and assessed according to their versatility in scientific computing in general and the above problem in particular. Graphics processing units (GPU) are identified as technology of high potential, promising powerful computing abilities while being increasingly easier to program. In order to fully benefit from the advantages of this new kind of hardware, the complete simulation development process is reviewed. A model for particulate flows is derived from the very basis of continuum mechanics leading to a constraint Navier-Stokes problem. The solving methodology is chosen to fit the GPU’s architecture and achieve best possible performance results. The implementation of the numerical method is done on the basis of intensive evaluation of the GPU used, by means of key scientific computation kernels. Final benchmarks of components of the particulate flow simulation demonstrate the successful use of GPUs to accelerate the computation by at least a factors of 3 and up to 20 in selected sub-routines.