A novel approach to evaluating compact finite differences and similar tridiagonal schemes on GPU-accelerated clusters

Compact finite difference schemes are widely used in the direct numerical simulation of fluid flows for their ability to better resolve the small scales of turbulence. However, they can be expensive to evaluate and difficult to parallelize. In this work, we present an approach for the computation of compact finite differences and similar tridiagonal schemes on graphics processing units (GPUs). We present a variant of the cyclic reduction algorithm for solving the tridiagonal linear systems that arise in such numerical schemes. We study the impact of the matrix structure on the cyclic reduction algorithm and show that precomputing forward reduction coefficients can be especially effective for obtaining good performance. Our tridiagonal solver is able to outperform the NVIDIA CUSPARSE and the multithreaded Intel MKL tridiagonal solvers on GPU and CPU respectively. In addition, we present a parallelization strategy for GPU-accelerated clusters, and show scalabality of a 3-D compact finite difference application for up to 64 GPUs on Clemson’s Palmetto cluster