Task Placement of Parallel FFTs on a Mesh Communication Network
Heike Jagode
TU Dresden Center for Information Services and High Performance Computing (ZIH) Dresden, Germany
October 11, 2007
10:00 AM
ORNL, 5600-J304
Host: Jeff Vetter
(vetter@ornl.gov
)
ABSTRACT:
For
many scientific applications, the fast Fourier transformation (FFT) of
multi-dimensional data is the kernel that limits scalability to large
numbers of processors. This talk summarizes the investigation of the
potential scope for performance improvements of a parallel FFT by
choosing different mappings between the virtual processor grid of the
algorithm and the physical hardware of a system with a mesh
interconnect. Based on bandwidth considerations, a simple model for the
performance of a large class of mappings is derived. The predictions of
the model are tested on an IBM eServer Blue Gene/L system. The results
demonstrate that a carefully chosen mapping pattern with regards to the
network characteristics yields significant improvement.
# # #