Publications
Algorithms and architectures for high performance analysis of semantic graphs
Semantic graphs offer one promising avenue for intelligence analysis in homeland security. They provide a mechanism for describing a wide variety of relationships between entities of potential interest. The vertices are nouns of various types, e.g. people, organizations, events, etc. Edges in the graph represent different types of relationships between entities, e.g. 'is friends with', 'belongs-to', etc. Semantic graphs offer a number of potential advantages as a knowledge representation system. They allow information of different kinds, and collected in differing ways, to be combined in a seamless manner. A semantic graph is a very compressed representation of some of relationship information. It has been reported that the semantic graph can be two orders of magnitude smaller than the processed intelligence data. This allows for much larger portions of the data universe to be resident in computer memory. Many intelligence queries that are relevant to the terrorist threat are naturally expressed in the language of semantic graphs. One example is the search for 'interesting' relationships between two individuals or between an individual and an event, which can be phrased as a search for short paths in the graph. Another example is the search for an analyst-specified threat pattern, which can be cast as an instance of subgraph isomorphism. It is important to note than many kinds of analysis are not relationship based, so these are not good candidates for semantic graphs. Thus, a semantic graph should always be used in conjunction with traditional knowledge representation and interface methods. Operations that involve looking for chains of relationships (e.g. friend of a friend) are not efficiently executable in a traditional relational database. However, the semantic graph can be thought of as a pre-join of the database, and it is ideally suited for these kinds of operations. Researchers at Sandia National Laboratories are working to facilitate semantic graph analysis. Since intelligence datasets can be extremely large, the focus of this work is on the use of parallel computers. We have been working to develop scalable parallel algorithms that will be at the core of a semantic graph analysis infrastructure. Our work has involved two different thrusts, corresponding to two different computer architectures. The first architecture of interest is distributed memory, message passing computers. These machines are ubiquitous and affordable, but they are challenging targets for graph algorithms. Much of our distributed-memory work to date has been collaborative with researchers at Lawrence Livermore National Laboratory and has focused on finding short paths on distributed memory parallel machines. Our implementation on 32K processors of BlueGene/Light finds shortest paths between two specified vertices in just over a second for random graphs with 4 billion vertices.