Performance study of mapping irregular computations on GPUs

Abstract

Recently, Graphical Processing Units (GPUs) have become increasingly more capable and well-suited to general purpose applications. As a result of the GPUs high degree of parallelism and computational power, there has been a great deal of interest directed toward the platform for parallel application development. Much of the focus, however, has been on very regular applications that exhibit a high degree of data parallelism, as these applications map well to the GPU. Irregular applications, such as the Breadth First Search discussed in this paper, have not been as extensively studied and are more difficult to implement in an efficient fashion on the GPU. We will present both an implementation of the Breadth First Search algorithm as well as that of a Matrix Parenthesization algorithm. These pair of algorithms showcase similar synchronization behavior when implemented on a GPU using CUDA, enabling a more direct comparison between them. The results obtained can be used to showcase some of the synchronization issues present with irregular algorithms on the GPU.