high performance computing on graphics processing units: hgpu.org

The goal of this work is to explore novel approaches to CUDA accelerated breadth-first search (BFS) algorithm and analyze their application in a state-of-the-art algorithm for graph decomposition into strongly connected components via CUDA capable devices, i.e. GPUs. A previous method [7], as will be shown, does not reasonably work on real-world graphs. Therefore, we extend a data set for an experimental evaluation of the algorithm and provide reasoning behind an applicability of the original algorithm and its improvements to the individual types of graphs. Furthermore, we attempt to identify circumstances which the most significantly affect a performance of the algorithm. Finally, we show that the nature of the chosen algorithm is an ultimate barrier to its effective usage on all types of graphs what opens the space for further research.