Accelerating Algorithms on GPUs in SCIRun: the Conjugate Gradient Case Study
Northeastern University, Boston, Massachusetts
Symposium on Application Accelerators in High Performance Computing, 2010
@article{yablonskiaccelerating,
title={Accelerating Algorithms on GPUs in SCIRun},
author={Yablonski, D. and Leeser, M. and Brooks, D.},
booktitle={Application Accelerators in High Performance Computing, 2010 Symposium, Papers},
year={2010}
}
The goal of this research is to integrate graphics processing units (GPUs) into SCIRun, a biomedical problem solving environment, in a way that is transparent to the scientist. We have developed a portable mechanism that allows seamless coexistence of CPU and accelerated GPU computations to provide the best performance while also providing ease of use. Features include integration into the existing graphical user interface (GUI) as well as easy extensibility of GPU processing to future algorithm development. As a case study, we focus on the linear solving algorithms of SCIRun for sparse data, including the conjugate gradient (CG) with Jacobi preconditioner. Acceleration of nearly 6x was achieved using NVIDIA’s CUDA with sparse matrices, demonstrating the performance of our approach. The linear solving algorithms were chosen for their suitability for acceleration on the parallel processing architecture that NVIDIA’s GPUs exhibit and illustrate a mechanism that can be extended to other existing and future algorithms in SCIRun.
February 17, 2011 by hgpu
