Algebraic multigrid employing mixed structured–unstructured data on manycore hardware

This paper outlines investigations in computing performance when deploying a Computational Fluid Dynamics solution on CPU and GPU resources. Critical to the performance is an algebraic multigrid solver that concurrently utilizes mixed structured and unstructured data for solution. Software organization for manycore computing and data storage patterns for efficient memory access is described in detail along with performance testing on practical flow cases. It is shown that structured data blocks of sizes greater than 1 million, are solved more than 25× faster on a GPU compared to when solved using a single CPU thread. On the other hand unstructured data blocks does not reach more than 10× speedup for the same comparison. Consequently maximizing use of structured data blocks in a mixed data configuration allows a more efficient utilization of GPU acceleration while still benefiting from flexibility of unstructured blocks for grid generation purposes. Speedup obtained for mixed data block problems varies depending on the block configurations where an average 3× speedup is reported for a (76% structured, 24% unstructured) submarine incident flow problem in comparison with a same fully unstructured problem.