Towards Interactive Pattern Search in Massive Graphs

Tahsin Reza, Matei Ripeanu, Geoffrey Sanders, Roger Pearce

Abstract: We present the design overview of a pattern matching engine for labeled graphs that supports interactive search: the user, based on feedback received from the search system, repeatedly revises her search template until s/he is satisfied with the results. To this end, we have developed a distributed memory solution that supports human-in-the-loop processing. Our solution embraces a number of design principles to offer high-performance, scalability and efficiency: (i) fast parallel processing - we adopt a vertex parallel computation model; (ii) aggressive search space reduction - using lightweight routines, we identify and prune away the non-matching part of the graph early; (iii) redundant work elimination - a revised query is likely to share label(s) and/or substructure(s) with its predecessor(s); therefore, whenever possible, we avoid redundant computation by reusing (partial) match information from earlier searches. Our preliminary evaluation highlights the effectiveness of the proposed approach: using a 257 billion edge real-world webgraph, on a 128 node (4,608 cores) deployment, we demonstrate the advantage of our technique over a naive approach (that uses an exact matching solution to independently search the original query and each of its revisions).