Parallel performance analysis of coupled heat and fluid flow in parallel plate channel using CUDA

TLDR: In this article, technique applied for parallelization of in-house developed generic code using CUDA and OpenMP paradigm is discussed and it is found that the GPU provides immense speedup and outperforms OpenMP largely.

Abstract: The heat transfer analysis coupled with fluid flow is important in many real-world application areas varying from micro-channels to spacecraft’s. Numerical prediction of thermal and fluid flow situation has become very common method using any computational fluid dynamics software or by developing in-house codes. One of the major issues pertinent to numerical analysis lies with immense computational time required for repeated analysis. In this article, technique applied for parallelization of in-house developed generic code using CUDA and OpenMP paradigm is discussed. The parallelized finite-volume method (FVM)-based code for analysis of various problems is analyzed for different boundary conditions. Two GPUs (graphical processing units) are used for parallel execution. Out of four functions in the code (U, V, P, and T), only P function is parallelized using CUDA as it consumes 91% of computational time and the rest functions are parallelized using OpenMP. Parallel performance analysis is carried out for 400, 625, and 900 threads launched from host for parallel execution. Improvement in speedup using CUDA compared with speedup using complete OpenMP parallelization on different computing machines is also provided. Parallel efficiency of the FVM code for different grid size, Reynolds number, internal flow, and external flow is also carried out. It is found that the GPU provides immense speedup and outperforms OpenMP largely. Parallel execution on GPU gives results in a quite acceptable amount of time. The parallel efficiency is found to be close to 90% in internal flow and 10% for external flow.