A study of the effects of machine geometry and mapping on distributed transpose performance.

This paper describes a parallel strategy to extend the scalability of a small 3D FFT on thousands of Blue Gene/L processors. The approach is to execute the intermediate phases of the 3D FFT on smaller processor subsets. Performance measurements of the standalone 3D FFT on two communication protocols, MPI and BG/L ADE are presented. While the performance of the 3D-FFT with MPI-based and BG/L ADE-based implementations exhibited qualitatively similar behavior, the BG/L ADE-based version has lower communication cost than the MPI based version for small message sizes. Measurements also show that the proposed approach is effective in improving Particle-Mesh-based N-body simulation performance significantly at the limits of scalability.