Abstract We present a distributed memory many-GPU implementation of the Physalis method for resolving spherical particles in disperse multiphase flow simulations. The current work extends a previous single-GPU computational procedure by implementing a distributed memory Poisson solver and distributed finite-size particle methods using MPI. We document the changes required to move to a distributed memory model for both the fluid and solid phases. We benchmark the code with up to one million resolved particles in a domain size of 192 0 3 on 216 GPUs at the Maryland Advanced Research Computing Center and present strong and weak scaling results. Finally, by taking advantage of the realization that the solution procedure for the pressure Poisson equation can be implemented using a symmetric matrix, we are able to replace the biconjugate gradient stabilized algorithm used in previous work with the conjugate gradient algorithm.

中文翻译：

---

在许多 GPU 上使用 Physalis 方法解析粒子模拟

摘要 我们提出了 Physalis 方法的分布式内存多 GPU 实现，用于解析分散多相流模拟中的球形粒子。当前的工作通过使用 MPI 实现分布式内存泊松求解器和分布式有限尺寸粒子方法，扩展了以前的单 GPU 计算过程。我们记录了迁移到流体和固相的分布式内存模型所需的更改。我们在马里兰州高级研究计算中心的 216 个 GPU 上使用域大小为 192 0 3 的多达 100 万个解析粒子对代码进行了基准测试，并展示了强和弱缩放结果。最后，利用压力泊松方程的求解过程可以使用对称矩阵实现这一认识，