Many-Core Acceleration of Common Graph Programming Frameworks

This proposal describes a program to substantially accelerate graph analytic algorithms by leveraging many-core computation, and to allow seamless use of these accelerated implementations within popular graph ecosystems. We will (1) improve, adapt, and extend our Gunrock multi-GPU graph analytic engine with additional algorithms -- including tailored randomized algorithms for community detection and pattern matching -- and optimizations, including optimization for new hardware. We will also (2) implement translation layers allowing seamless use of this world-class engine with the APIs of popular frameworks GraphX (Spark) and GraphLab; code written for those frameworks will be sped up without modification. We will accomplish a 10,000 times speedup in algorithms such as Page Rank over existing solutions like Apache Giraph. Additionally, we will (3) make the improved high-performance graph analytics available through our algorithms as a service platform. This platform will make the capabilities accessible not only for users of graph frameworks, but also for those who have never used any graph framework, via a simple API. Users may leverage our cloud, negating the need to purchase hardware, or use their own cluster. We are seeing commercial demand for such an offering, across a range of industries.