DynaGraph

Abstract

In this paper, we present DynaGraph, a system that supports dynamic Graph Neural Networks (GNNs) efficiently. Based on the observation that existing proposals for dynamic GNN architectures combine techniques for structural and temporal information encoding independently, DynaGraph proposes novel techniques that enable cross optimizations across these tasks. It uses cached message passing and timestep fusion to significantly reduce the overhead associated with dynamic GNN processing. It further proposes a simple distributed data-parallel dynamic graph processing strategy that enables scalable dynamic GNN computation. Our evaluation of DynaGraph on a variety of dynamic GNN architectures and use cases shows a speedup of up to 2.7X compared to existing state-of-the-art frameworks.