Tumbling down the GraphBLAS rabbit hole with SHMEM

Abstract

In this paper we present shgraph, a SHMEM implementation of the GraphBLAS standard, which enables the user to redefine complex graph algorithms in terms of simple linear algebra primitives. GraphBLAS offers many nice features such as type abstractions, the ability to perform generalized matrix/vector operations over a semiring, and executing graph operations out-of-order (non-blocking mode). shgraph seeks to efficiently manage and process billion-edge or greater sparse graphs on an HPC system. We walk through sample GraphBLAS code and discuss the shgraph development process. In particular, we explain how SHMEM was used and where it was necessary to tweak the GraphBLAS specification to be compatible with a distributed system. Additionally, we analyze some preliminary performance results, map out next steps, and suggest potential applications.