A Comparison of Distributed Memory Fast Fourier Transform (FFT) Library Packages
Authors: Dr. Samar Aseeri (King Abdullah University of Science and Technology)
Abstract: The FFT is used in many different applications but has poor scalability due to its high communication volume. It is used because of its low computational cost when accurate results are required. There are many methods to use for implementing Fourier Transforms, and it is not clear what will work best on a given distributed platform. To allow for comparisons, FFT developers and users will discuss design considerations and requirements of FFT libraries. It is an opportunity to learn about packages from developers, to be able to differentiate between the libraries, and to determine best FFT library design practices.
Long Description: The BOF will take the form of several short presentations, targeting the wider high performance computing community. Each participant in the BOF will address the following questions:
1) Why did you write your own FFT?
2) What considerations are important for you in an FFT implementation?
3) What might you look for if there were to be a unified FFT interface (similar to BLAS, LAPACK and SCALAPACK interfaces)?
4) How important are performance, portability, and scalability for you?
5) Will FFT be needed in exascale computing and if so how will it be achieved?
6) What would be a good FFT benchmark or a good way to include the FFT in a high-performance computer benchmark?
Following these questions, there will be a general discussion between the presenters along with audience interaction. Contributions will highlight existing or potential bottlenecks in the different FFT packages; possible methods to overcome them as well as FFT interface design.
There will be contributions from 9 FFT developers and users:
Daisuke Takahashi University of Tsukuba, Japan
Franz Franchetti Carnegie Mellon University, USA
Steve Plimpton Sandia National Laboratories, USA
Mikael Mortensen University of Oslo, Norway
Michael Pippig Intenta GmbH Chemnitz, Germany
Yu Feng Berkeley Center for Cosmological Physics, USA
Jens Henrik Goebbert Jülich Supercomputing Centre, Germany
Sunita Chandrasekaran University of Delaware, USA
Amir Gholami UC Berkeley, USA
The targeted audience are those who seek performance in a code that utilizes an FFT library, hardware designers and supercomputer architects who need to ensure their platforms support FFT calculations efficiently and software developers who would need to implement their own application specific or hardware specific FFT and need to obtain the highest possible performance.
Note: 7 of the participants will be present with one remote participation. The rest two participants will either be present or provide a summary slide for the session.
Conference Presentation: pdf
Birds of a Feather Index