Spring 2009

Dr. Manuel Tiglio, a physicist at the University of Maryland interested in applying high performance computing techniques to gravity-based problems, and Dr. Ramani Duraiswami of the Computer Science department recommended I work on this project.

The research is based on Dr. Tiglio's recent work on the propagation of gravitational waves in the field of a Kerr black hole [1]. It makes heavy use of certain mathematical methods that map well to manycore, parallel architectures, like Chebyshev/Legendre-based differentiation and fourth-order Runge-Kutte integration. Further information can be found in the literature review [2].

In short, my project uses Nvidia's CUDA to accelerate these parallelizable functions using CUDA-capable graphics cards. Furthermore, it leverages FLAGON to interface CUDA and legacy Fortran code used for Dr. Tiglio's initial paper. A report on the project gives further details [3].

We saw an overall increase in speed, due to the manycore architecture of GPUs, but a decrease in accuracy, due to GPUs' speedy single-precision floating point arithmetic operations. These results are graphed below.

Local Speedup

Error Propagation

Timings Screenshot

[1] Tiglio, M. Kidder, L. Teukolsky, S. 2007. High accuracy simulations of kerr tails: coordinate dependence and higher multipoles. (Download .pdf)
[2] Literature Review (Download .pdf)
[3] Dickerson, J. 2009. Using GPUs to Accelerate Analysis of Kerr Tails. (Download .pdf)