Due to the demand for both computing power and environmental performance, the use of GPUs in scientific computing is advancing. CUDA libraries are tools that make it easy to convert existing calculations to GPUs, and they also handle sparse matrix operations, but their support for parallel computing is still lagging. These libraries achieve distributed parallel execution by internally reordering matrices to construct independent patterns. In structural analysis, linear equations are constructed in each domain based on mesh decomposition. Since there is overlap between the regions, the solution cannot be obtained when using the direct method without not only solving each equation but also handling the dependencies. We examined the implementation of parallel computing of direct methods on GPUs, which can be applied to problems that use distributed meshes based on node-based partitioning as input and can handle a wide range of analyses.
Kono et al. (Wed,) studied this question.