Conference Presentation (After Call)

FZJ-2024-04850

http://join2-wiki.gsi.de/foswiki/pub/Main/Artwork/join2_logo100x88.png Porting mpi4py-fft to GPU

Baumann, T. (Corresponding author)FZJ* ; Speck, R.FZJ*

2024

This record in other databases:

Please use a persistent id in citations: doi:10.34734/FZJ-2024-04850

Abstract: The mpi4py-fft library enables distributed fast Fourier transforms on CPUs with an easy to use interface and scales very well. We attempt to port this to GPUs, which significantly outperform the CPU counterpart at a given node count. While the porting is straightforward for the most part, the best communication strategy is still an open question for us.The algorithm relies on MPI alltoallw. Even with CUDA-aware MPI, this exhibits very poor performance on the Juelich computers. By replacing it with a custom communication strategy, throughput can be increased at a slight loss of generality. We would like to discuss optimising the strategy, or even if the performance of alltoallw can be increased by some measure.

---

Contributing Institute(s):

1. Jülich Supercomputing Center (JSC)

Research Program(s):

1. 5112 - Cross-Domain Algorithms, Tools, Methods Labs (ATMLs) and Research Groups (POF4-511) (POF4-511)
2. JLESC - Joint Laboratory for Extreme Scale Computing (JLESC-20150708) (JLESC-20150708)
3. RGRSE - RG Research Software Engineering for HPC (RG RSE) (RG-RSE) (RG-RSE)

Appears in the scientific report 2024

---

Database coverage:

---