Performance of unstructured Navier Stokes solvers on multi-core cluster architectures

By its nature, the cache optimization of a computational fluid dynamics (CFD) code is most readily assessed on relatively simple geometries and on problems that only require core routines. Having previously conducted cache optimization on the unstructured two-/three-dimensional code UNCLE, we are now examining in detail what effect these improvements have on the performance of this code. L1, L2 and TLB cache performance is compared between different versions of the Uncle code. Then, the performance characteristics of this code are compared to another unstructured code, icoFoam, which is part of the OpenFOAM (Open Field Operation and Manipulation) CFD toolbox. The analysis considers the effects of running these codes on single core processors and multi-core commodity processors in parallel. Code modifications to improve code performance in these flows through cache-usage optimization will be presented, and the overall ability to translate code performance from core routines and simple geometries to widespread practical applications are discussed.