A novel strategy for compact finite difference evaluation on gpu-accelerated clusters

Abstract eng:
In this work, we present a novel strategy to solve the tridiagonal systems arising in such numerical schemes as compact finite differences and alternating direction implicit methods on graphics processing units (GPUs). We demonstrate the impact of the simple matrix structure on the cyclic reduction algorithm, and show that precomputation of coefficients appearing in the algorithm becomes feasible and efficient. We demonstrate that an implementation using our approach is able to outperform the NVIDIA CUSPARSE and multithreaded Intel MKL solvers on GPU and CPU, respectively. We apply this tridiagonal solver to the solution of compact finite differences on multiple GPUs distributed in a cluster, and show scaling for up to 64 GPUs.

• Export as: BibTeX | MARC | MARCXML | DC | EndNote | NLM | RefWorks
• Add to your basket