Code Synthesis for High-Performance Graph Analytics

Graph analytics platforms are widely deployed across diverse domains to analyze large, unstructured and semi-structured data-sets. While existing tools are clearly effective, there is a persistent need to perform deeper analyses on larger data-sets and retrieve information more quickly. Low data-locality and a high communication to compute ratio means that existing distributed frameworks are unlikely to be sufficient.Graphics processors (GPUs) have the potential to deliver dramatically better performance but realizing that potential necessitates algorithms and data-structures tailored the target device. Specializing algorithms by hand is impractical and prohibitively expensive.We propose to build a tool for generating specialized, high-performance graph algorithms for GPUs starting from high-level Python code.Code will be generated using program transformation and automated search to explore the design space of algorithms and representations. The tool will leverage an existing analytics framework, such as GraphLab, to yield an integrated, high-productivity, high-performance graph analytics capability.