A Parallel Preconditioned Conjugate Gradient Solver for the Poisson Problem on a Multi-GPU Platform

摘要

We present a parallel conjugate gradient solver for the Poisson problem optimized for multi-GPU platforms. Our approach includes a novel heuristic Poisson preconditioner well suited for massively-parallel SIMD processing. Furthermore, we address the problem of limited transfer rates over typical data channels such as the PCI-express bus relative to the bandwidth requirements of powerful GPUs. Specifically, naive communication schemes can severely reduce the achievable speedup in such communication-intense algorithms. For this reason, we employ overlapping memory transfers to establish a high level of concurrency and to improve scalability. We have implemented our model on a high-performance workstation with multiple hardware accelerators. We discuss the mathematical principles, give implementation details, and present the performance and the scalability of the system.