high performance computing on graphics processing units: hgpu.org

Nowadays, scientists are often disappointed by the outcome when parallelizing their simulations, in spite of all the tools at their disposal. They often invest much time and money, and do not obtain the expected speed-up. This can come from many factors going from a wrong parallel architecture choice to a model that simply does not present the criteria to be a good candidate for parallelization. However, when parallelization is successful, the reduced execution time can open new research perspectives, and allow to explore larger sets of parameters of a given simulation model. Thus, it is worth investing some time and workforce to figure out whether an algorithm is a good candidate to parallelization. Automatic parallelization tools can be of great help when trying to identify these properties. In this paper, we apply an automatic parallelization approach combining Java and OpenCL on an existing Gap Model. The two technologies are linked with a library from AMD called Aparapi. The latter allowed us to study the behavior of our automatically parallelized model on 10 different platforms, without modifying the source code.