Modern algorithms in numerical simulation need to combine efficient mathematical methods with concepts from computer science for nowadays computer architectures. In high-performance computing, memory access is a crucial factor and more important than pure cpu power. This memory boundedness can be reduced by utilizing modern hierarchical memory structures. However, many classical multigrid methods lack the necessary data locality. Thus, we use a space-filling curve, the Peano curve, for the discretization and linearization of the domain. Then, we employ a system of stacks for processing the grid points linearly in a cache efficient way. Based on an additive multigrid method, an adaptive full multigrid cycle using a posteriori error estimates is developed. Additionally, higher order methods can be integrated by using τ-extrapolation. Thus, while keeping the cache-efficiency on a very high level, one is able to apply advanced mathematical tools and ideas.
«
Modern algorithms in numerical simulation need to combine efficient mathematical methods with concepts from computer science for nowadays computer architectures. In high-performance computing, memory access is a crucial factor and more important than pure cpu power. This memory boundedness can be reduced by utilizing modern hierarchical memory structures. However, many classical multigrid methods lack the necessary data locality. Thus, we use a space-filling curve, the Peano curve, for the discr...
»