Generating Large-Scale Heterogeneous Graphs for Benchmarking

Graphs have emerged as an important genre of data that are found in a wide class of applications. The most dominant benchmark for graph data today is Graph 500 that generates a Stochastic Kronecker graph of various sizes, and reports the time to perform a breadth-first search. Apache Giraph uses Pagerank computation as an algorithmic benchmark for large graphs, but does not provide the mechanism to generate graph data. Other forms of graph benchmarks have been developed by smaller communities and are not known widely. However, most benchmarking data for graphs are derived from a single structure generation model, and therefore does not capture the variability of structure and content. To this end, we propose heterogeneous graphs, a mixed model graph structure that combines several existing generation techniques into a single benchmark. It is a hybrid that constructs edge-labeled multigraphs with multiple components, which can be hierarchical, power-law graphs, community-forming graphs, and a new class of graphs formed by motif composition. The user can use a simple set of 4 parameters to specify the graph, but has the option to use several more parameters to have a finer control of the hybrid structure. We define the generation process for heterogeneous graphs and propose an initial set of query operations against the generated data.