Content area
Numerous modern engineering applications rely heavily on finite element analysis to perform informative solid mechanics simulations. The majority of these simulations rely on low-order methods and assembled sparse matrices, which offer poor computational efficiency, especially on GPUs, compared to high-order matrix-free methods.
We perform efficiency analyses to show the reliability, efficiency, and scalability of matrix-free p-multigrid methods with algebraic multigrid coarse solvers for large deformation hyperelastic simulations of multiscale structures. We investigate accuracy, cost, and execution time for multi-node jobs on an variety of GPU architectures for simulations with millions to billions of degrees of freedom, resulting in order of magnitude efficiency improvements over a broad range of scales.
Additionally, we extend these methods to finite-strain poroelasticity, showing that the efficiency benefits of high-order matrix-free methods carry over to advanced materials. We include the mathematical formulation and matrix-free implementation for both small-strain (linear) and finite-strain (nonlinear) poroelastic models, based on Biot's and mixture theory. We use the finite-strain model to perform efficiency studies to show that the efficiency gains of matrix-free formulations and p-MG extend to complex and multiphysics materials and can outperform a hyperelastic model with sparse assembled matrix implementation.
These methods are all available as open source libraries that handle all architecture-specific implementations for the user, allowing for more rapid adoption of the tools. The demonstrated solver setup and data structures allow for both newer and legacy code bases to benefit from higher GPU performance with minimal GPU-specific code needed from the user.
