CSSI Element: SI2-SSE: Gunrock: High-Performance GPU Graph Analytics

We present a plan to develop the “Gunrock” programmable, high-performance graph analytics library for programmable graphics processors (GPUs) from a working prototype to a robust, sustainable, open-source component of the GPU computing ecosystem. We believe Gunrock is the most fully-featured and highest-performing programmable graph library today for single-GPU and single-node, multiple-GPU graph analytics. However, we see numerous exciting opportunities for a coherent program of research and development to improve Gunrock and make it more valuable for the computing community, including supporting greater scalability, expanding the Gunrock stack with both higher-level APIs and additional Gunrock back ends, adding more core operations and primitives to Gunrock, and exploring graph data structure challenges. Gunrock has grown considerably since this grant's inception and we present new foundational work that moves graph analytics for irregular problems beyond bulk synchronous parallel models and on to asynchronous task-based models more amenable to multi-GPU and multi-node computing environments. In addition we present some lessons learned while maintaining an open source project targeting both industry and academia.