CS224W: Methods of Parallelized Kronecker Graph Generation

The question of generating realistic graphs has always been a topic of huge interests. This topic has gained huge attention over the past few years with the advent of massive real-world network data that re generated by large software companies like Facebook and Google, along with the increase in the computation power that makes anyone capable of processing them. With real graphs at massive scale and parallelized frameworks to analyze them, network analysis became a major topic of scientific research. As the need to analyze these networks grew, the question of modeling and generating a real-world network graph at the same scale also became a topic of interest. Out of many approaches to model real-world networks, Stochastic Kronecker Graph (SKG) generation and its predecessor R-MAT generation have attracted interest in the network analysis community, due to their simplicity and their abilities to capture the properties of real-world networks. Along with such algorithms, a new programming methods to process large graphs called vertex-centric BSP with the implementations such as Pregel [3], Apache Giraph, GPS, and Apache Hama have become increasingly popular as an alternative to MapReduce and Hadoop, which are ill-suited to run massive scale graph algorithms [3]. The SKG, R-MAT and vertex-centric BSP, however, are not well-suited for each other. The obvious approach of parallelizing SKG, which is to generate edges in parallel, is not “vertexcentric” in nature and therefore is unnatural to program and runs inefficiently in vertex-centric