Seminars

back Back to all Seminars

CS Seminar: Development of an efficient FFT library and its scaling up to 65,536 cores of BlueGene/P and 196,608 cores of Cray XC40

Start Date: October 14, 2018
End Date: October 14, 2018

​By Dr. Anando G. Chatterjee (Indian Institute of Technology Kanpur, India)
 
The Fast Fourier transform (FFT) is a crucial kernel of many scientific simulation codes, including pseudo-spectral codes for partial differential equations. For a three-dimensional grid of size N^3, the FFT has a near optimal time complexity of O((N^3 ) log_2 (N^3 ))[1]. Hence, parallel algorithms have been devised to compute the FFT for large N, finely resolved grids. For performing high resolution simulations of various systems in periodic, free-slip, or mixed boundary conditions we have developed an FFT library named FFTK. In this talk the parallelization strategies used in FFTK and its comparison with popular FFT libraries will be presented. We will present strong and weak scaling of FFTK to 65,536 cores of BlueGene/P (Shaheen-I at KAUST) and 196,608 cores of Cray XC40 (Shaheen-II) and discuss how the all-to-all routines of MPI (Message Passing Interface) affect scaling[2]. On top of FFTK, we have built a pseudo-spectral open-source code, TARANG, to perform simulations of turbulent fluid flows, Rayleigh-Benard convection, magnetohydrodynamics, and magnetoconvection with periodic or free-slip boundary conditions[3]. Turbulence is a ubiquitous phenomenon in nature characterized by a wide range of spatial and temporal scales. Fundamental turbulence problems are often studied for incompressible and homogeneous systems. To solve multi-scale phenomena in turbulent flows, pseudo-spectral and spectral simulations are found to be efficient and accurate and focus attention on the FFT for scalable performance. [1] James W. Cooley, John W. Tukey. An algorithm for the machine calculation of complex Fourier series. Math. Comput. 19: 297–301 (1965). [2] A. G. Chatterjee, M. K. Verma, A. Kumar, R. Samtaney, B. Hadri, and R. Khurram, Scaling of a Fast Fourier Transform and a pseudo-spectral fluid solver up to 196608 cores, J. Parallel Distrib. Comput. 113, 77 (2018). [3] M. K. Verma, A. Chatterjee, K. S. Reddy, R. K. Yadav, S. Paul, M. Chandra, and R. Samtaney, Benchmarking and scaling studies of pseudospectral code Tarang for turbulence simulations, Pramana, 81, 617 (2013).
 
Biography: Anando G. Chatterjee is pursuing Ph.D. from Department of Physics, Indian Institute of Technology Kanpur, India. His main work was to enhance a pseudo-spectral code Tarang, develop a Fast Fourier Library FFTK and showing similarities between fluid turbulence and RBC so that fluid turbulence models, such as LES, could be applied to RBC as well. His main scientific interests are: high-performance computing, distributed and parallel programming, and GPGPU computing. He is a lead developer of the library FFTK that has been scaled up to 1,96,608 cores of Cray XC40 and supports Sin/Cos/Fourier mixed transforms.
 

More Information:

For more info contact: Prof. David Keyes; email david.keyes@kaust.edu.sa
 
Date: Sunday 14th Oct 2018
Time:12:00 AM - 01:00 PM
Location: Building 4, Level 5, Rm 5220
Light refreshments will be served