Investigation of Different Parallel Models for DNS of Turbulent Channel Flow

Abstract

The chapter investigates the efficiency of different parallel implementations and targets at direct numerical simulation (DNS) of high Reynolds number channel flow. The chapter presents the investigation of different parallel implementations, including different parallel models and benchmark results, and reports DNS simulation results at high Reynolds number, Re*=1000. Detailed statistics and visualization are also provided in the chapter. Because DNS at high Reynolds number is time and memory consuming, there is a need to optimize the code to achieve the highest possible parallel efficiency to use supercomputers. To obtain maximum optimization, several different parallel models are implemented in the chapter. Their performance is benchmarked and the best model is chosen. The first model uses domain decomposition in the stream-wise direction, and it is the easiest way to implement. However, this model has potential limitation on the maximum number of processors that can be used, because at least two planes need to be allocated to each processor to do de-aliasing in the nonlinear step.