For many scientific applications, the fast Fourier transformation (FFT) of multi-dimensional data is the kernel that limits scalability on a large number of processors. This paper investigates the extent of performance improvements for a parallel three-dimensional FFT (3D-FFT) implementation when using customized MPI task mappings. The MPI tasks are mapped in a customized fashion from the two-dimensional virtual processor grid of the algorithm to the physical hardware of a system with a mesh interconnect. We compare and analyze the outcomes on Blue Gene/P with those from previous investigations on Blue Gene/L. The performance analysis is based on bandwidth considerations. The results demonstrate that on Blue Gene/P, a carefully chosen MPI task mapping with regards to the network characteristics is more important compared to Blue Gene/L and yields significant improvement.