GraphBLAST Targets GPU Graph Analytics Performance Issues
Since its introduction into the public domain five years ago, GraphBLAS has been widely adopted across commercial, scientific, and computational research communities. Heralded for its contributions to the long-standing Basic Linear Algebra Subprograms (BLAS) library, the open-source GraphBLAS framework applies linear algebra to high-performance graph analytics to improve scalability for a broad range of applications, from artificial intelligence and cybersecurity to computational science and parallel computing. It’s also been used to develop graph and combinatorial algorithms for exascale systems through the U.S. Department of Energy’s Exascale Computing Program.
Now a new iteration, “GraphBLAST” – developed by researchers from Lawrence Berkeley National Laboratory (Berkeley Lab) and the University of California, Davis (UC Davis) – further enhances the performance of this popular collection of graph algorithm building blocks by overcoming design and performance challenges specific to GPU processors.
“This is the first GPU-based graph framework that has the same performance as some of the traditional vertex-centric graph frameworks on the GPUs,” said Carl Yang, a software engineer at UC Davis and Berkeley Lab and lead GraphBLAST developer. He is also first author on a comprehensive paper describing GraphBLAST and its capabilities published in the journal ACM Transaction on Mathematical Software.
This article originally appeared on Computing Sciences Research at Lawrence Berkeley National Laboratory.
Kathy Kincade is a computing sciences editor/writer at the Lawrence Berkeley National Laboratory.