Elements: Software: Multidimensional Fast Fourier Transforms on the Path to Exascale

点击翻译按钮获取中文摘要

This project will contribute to software available freely to the public, providing solution of an important class of problems in computational science, namely the multidimensional Fast Fourier transforms (FFTs). This software will enable scientists to run certain classes of computer experiments on powerful state-of-the-art supercomputers, known as Exascale machines. A wide range of science fields will benefit from such public software; thus, it will contribute to innumerable discoveries related to fundamental understanding of nature, protecting the environment, efficient energy and transportation, designing new materials and improving medicine. It will also contribute to the computational ecosystem by virtue of being a building block that can be used and reused by other programs. This project also includes an educational effort in training a wide range of supercomputer users, from undergraduate interns to graduate/postdoctoral researchers and faculty. Effort will be made to reach out to underrepresented groups when recruiting the interns.The National Strategic Computing Initiative (NSCI), in particular Strategic Objective 4 "An Enduring National HPC Ecosystem", outlines the crucial need for development of foundational algorithms and software for next generation supercomputers, and for easing access to the next-generation compute resources to wide classes of users. Fast Fourier Transforms (FFT) is a ubiquitous tool in scientific simulations, from Computational Fluid Dynamics to plasma physics, astrophysics, ocean modeling, materials research, medical imaging, molecular dynamics and many others. This project will fill the gap in highly efficient software for multidimensional FFTs for use on Exascale platforms. While running full FFT at exa-scale appears prohibitive due to strong interconnect bandwidth dependence, steps must be taken towards this goal due to the importance of the algorithm. Building on the previous work with P3DFFT, an open-source numerical library used by many applications in diverse science fields, this proposal aims to push the envelope in terms of adapting multidimensional FFT to new architectures and aggressively scale its performance, without losing the portability and the practical ease of use. The project will employ both novel and proven state-of-the-art tools, such as auto-tuning, overlap of communication with computation and GPU implementation, which will help reduce the bandwidth bottleneck. In addition, features will be added that will make the software appealing to a wider user base. The software will become an integral part of the National Cyberinfrastructure and will aid in numerous scientific and engineering advances, including in areas such as environmental research, energy, efficient transportation, new drugs and new materials. This project includes training of a new generation of scientific software developers through XSEDE training webinars, conference training workshops and supervision of undergraduate interns. The project plan includes reaching out to underrepresented groups when recruiting undergraduate interns.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.