Introducing the Graph 500

Abstract

In the words of Lord Kelvin, “if you cannot measure it, you cannot improve it”. One of the longlasting successes of the Top 500 list is sustained, community-wide floating point performance improvement. Emerging large-data problems, either resulting from measured real-world phenomena or as further processing of data generated by simulation, have radically different performance characteristics and architectural requirements. As the community contemplates scaling to large-scale HPC resources to solve these problems, we are challenged by the reality that supercomputers are typically optimized for the 3D simulation of physics, not large-scale, data-driven analysis. Consequently, the community contemplating this kind of analysis requires a new yard stick for evaluating future platforms. Since the invention of the von Neumann architecture, the physics simulation has largely driven the development and evolution of High Performance Computing. This allows scientists and engineers to test hypotheses, designs, and ask “what if” questions. Emerging informatics and analytics applications are different both in purpose and structure. While physics simulations typically are core-memory sized, floating point intensive, and well-structured, informatics applications tend to be out of core, integer oriented, and unstructured. (It could be argued that physics simulations are moving in this direction.) The graph abstraction is a powerful model in com-