Files

No attached file found for this publication.

Details

Authors
Abstract
Massively parallel Fourier transforms are widely used in computational sciences, and specifically in computational fluid dynamics simulations of incompressible external flows which involves unbounded Poisson problems. In this talk, we present a library of unbounded Poisson solvers (FLUPS) for 2D and 3D distributed uniform grids, tailored for massively parallel architectures. The multidimensional Fourier transform is performed as a succession of 1D FFTs using FFTW, while we propose three different communication strategies for the reordering and parallel redistribution of the data between each transform: one all-to-all implementation and two non-blocking implementations using manual packing or MPI Datatype. The proposed software is validated against analytical solutions for unbounded, semi-unbounded, and periodic domains. The performance of the approaches is then compared against another distributed FFT implementation, using a periodic problem. Finally, the performance metrics of each implementation are analyzed and detailed on various top-tier European systems up to 49,152 cores. This work brings FLUPS to a fully production-ready and performant distributed FFT library, featuring many possible types of FFTs and with flexibility in the data-layout. The code is available under a BSD-3 license at github.com/vortexlab-uclouvain/flups.
Affiliations

Citations

  • APA
  • Chicago
  • FWB

Balty, P., Caprace, D.-G., Chatelain, P., & Gillis, T. (2023). FLUPS: a versatile and performant FFT-based library of unbounded Poisson solvers. SIAM Conference on Computational Science and Engineering (CSE23), Amsterdam. https://hdl.handle.net/2078.5/103591