Parallel Graph Mining on Shared Memory Architectures

Mining graph based data sets has practical applications in many areas including molecular substructure discovery, web link analysis, fraud detection, and social network analysis. The discovery challenge in graph mining is to find all subgraphs which occur in at least graphs of a graph database, where is user specified parameter. Subgraph isomorphism, the enormous search space of candidate graph patterns, and the importance of interactive response times make addressing this challenge particularly daunting. In this work, building on the existing state-of-the-art, we propose a novel approach for parallelizing such algorithms on shared memory multiprocessor systems. We present several novel task partitioning and load balancing schemes, and evaluate their efficacy. We also propose several queuing models which afford dynamic task sharing. We show that dynamic partitioning and dynamic task allocation provide a synergy which greatly improves scalability. Our parallelization algorithm achieves excellent speedup of 27 times on 32 nodes for several real world datasets, as compared to a naive approach which affords only 5-fold speedup. We also discuss implications of this work in light of recent trends in micro-architecture design, particularly multiple core and multithreaded systems.