Many-Core Acceleration of Common Graph Programming Frameworks

We propose creating a new company to deliver accelerated graph-based analytics for modern manycore systems. The work combines advanced multithreading and work-moving typical of HPC but targeted at cloud environments with insight into the structure of the graph to improve spatial locality and an attribute-first query system to dramatically reduce the work required to solve practical graph problems. We propose to focus on analyst and graph-framework programmer productivity to allow the construction of graphs in a form natural to the users understanding of the problem they are solving, while leveraging the construction and query information passed into the framework to dynamically build data structures more amenable to the block and cache oriented structure of modern manycore cache and memory hierarchies. We view graph processing in the context of real-world workflows that include not just the performance of graph queries but data ingest and additional post-processing. The work will be done in the context of social network, cybersecurity, and medical informatics graphs of relevance to national security and commercial enterprises. The results from Phase I will yield an approach that meet the 4-5 order of magnitude performance improvement objectives of the program and government and private sector commercial needs.