Large-Scale Algorithm Design for Parallel FFT-based Simulations on GPUs

We describe and analyze a co-design of algorithm and software for high-performance simulation of a partial differential equation (PDE) numerical solver for large-scale datasets. Large-scale scientific simulations involving parallel Fast Fourier Transforms (FFTs) have extreme memory requirements and high communication cost. This hampers high resolution analysis with fine grids. Moreover, it is difficult to accelerate legacy Fortran scientific codes with modern hardware such as GPUs because of memory constraints of GPUs. Our proposed solution uses signal processing techniques such as lossy compression and domain-local FFTs to lower iteration cost without adversely impacting accuracy of the result. In this work, we discuss proof-of-concept results for various aspects of algorithm development.