Acceleration of a finite-difference method with general purpose GPUs - Lesson learned

Abstract

Modern massively parallel graphics cards (GPGPUs) offer a promise of dramatically reducing computation times of numerically-intensive dataparallel algorithms. As cards that are easily integrated into desktop PCs, they can bring computational power previously reserved for computer clusters to the office space. High performance rates make GPGPUs a very attractive target platform for scientific simulations. In this paper we present the lessons learned during the parallelization of a finite-difference time-domain method, an inherently data-parallel algorithm frequently used for numerical computations, on the state of the art graphics hardware.

Accelerating a three-dimensional finite-difference wave propagation code using GPU graphics cards

“…molecular dynamics simulations (Yang et al 2007; Anderson et al 2008), fluid dynamics simulations (Brandvik & Pullan 2007; Elsen et al 2008), or astrophysical calculations (Nyland et al 2007). Regarding FD, several applications have been ported to GPUs as early as 2004 (Krakiwsky et al 2004a,b; Baron et al 2005; Humphrey et al 2006; Adams et al 2007; Abdelkhalek 2007; Inman et al 2007; Price et al 2007; Balevic et al 2008a,b; Valcarce et al 2008; Inman & Elsherbeni 2008; Micikevicius 2009; Abdelkhalek et al 2009). Some of the other numerical techniques mentioned above for seismic wave propagation have recently been successfully ported to GPUs, for instance the spectral‐element method by Komatitsch et al (2009) in the case of one GPU and by Komatitsch et al (2010a,b) in the case of a cluster of many GPUs used in parallel, and the discontinuous Galerkin method by Klöckner et al (2009).…”

The Case for Higher Computational Density in the Memory-Bound FDTD Method within Multicore Environments

“…In GPU computing, this can cause a serious performance penalty in the form of long idle times for the computing cores as they wait for data to process. Most of the efforts reported in the GPU/FDTD literature [3][4][5][6][7][8][9][10][11][12][13] were focused on how to best use the available GPU memory bandwidth through exploiting its on-chip shared and texture memory banks, as opposed to the much slower off-chip global memory. To a slightly lesser degree, this memory-bandwidth limitation issue applies to today's high-end CPUs as well.…”