Parallel Computing of KIVA-4 Using Adaptive Mesh Refinement

Parallel computing schemes were developed to enhance the computational efficiency of engine spray simulations with adaptive mesh refinement (AMR). Spray simulations have been shown to be grid dependent and thus fine mesh is often used to improve solution accuracy. In this study, dynamic mesh refinement adaptive to spray region was developed and parallelized in KIVA-4. The change of cell and node numbers and the local characteristics in the dynamic mesh refinement posed difficulties in developing efficient parallel computing schemes to achieve low communication overhead and good load balance. The present strategy executed AMR on one processor with data scattering among processors following the adaptation, and performed AMR every ten computational timesteps for enhanced parallel performance. The re-initialization was required and performed at the minimized cost. The present computational schemes were used to simulate transient sprays in three different engine geometries including a cylindrical chamber, a 2-valve engine, and a direct-injection gasoline engine. The implementation was validated by comparing the predicted spray penetrations and structures under different simulation conditions. Various spray conditions and grid resolutions were tested in the three geometries to assess the computational efficiency. Reasonable speed-ups were obtained for the test cases.