GPU acceleration of hybrid FETI solver for problems of transient nonlinear dynamics

Abstract

FETI methods, which build on the Finite Element Method, are utilized for large-scale engineering simulations. They use domain decomposition techniques to divide a large domain into many smaller subdomains, which can be processed in parallel. Current trends in HPC focus on GPU-accelerated clusters. To utilize them efficiently, FETI solvers should be able to use these accelerators. Recent developments have demonstrated that the fundamental component of the FETI methods, the dual operator, can be successfully offloaded to the GPU.In this paper, we focus on GPU acceleration of the Hybrid FETI variant. It reduces the size of the projector by using a two-level decomposition, thus allowing for a significantly higher number of compute nodes to be efficiently utilized. In turn, it allows us to split the problem into a larger number of smaller subdomains, which improves single-process performance. We demonstrate the performance on a real-world problem of transient nonlinear dynamics that requires reassembling of the dual operator, preconditioner, and projector during each call of the solver. On the MareNostrum 5 supercomputer, using Nvidia H100 GPUs, we achieved a speedup of 2.9 for the whole Hybrid FETI solver compared to a CPU-only run.