Characterizing and Understanding GCNs on GPU

TLDR

GCNs achieve state‑of‑the‑art performance on graph‑structured data, and their training and inference are accelerated by GPUs, yet detailed characterization of GCN workloads on GPUs is lacking. This study characterizes GCN inference workloads on NVIDIA V100 GPUs and proposes guidelines to optimize their software and hardware execution. We analyze inference‑stage GCN workloads and benchmark GCN models on an NVIDIA V100 GPU. The study yields guidelines that improve software and hardware efficiency for GCN execution on GPUs.

Abstract

Graph convolutional neural networks (GCNs) have achieved state-of-the-art performance on graph-structured data analysis. Like traditional neural networks, training and inference of GCNs are accelerated with GPUs. Therefore, characterizing and understanding the execution pattern of GCNs on GPU is important for both software and hardware optimization. Unfortunately, to the best of our knowledge, there is no detailed characterization effort of GCN workloads on GPU. In this letter, we characterize GCN workloads at inference stage and explore GCN models on NVIDIA V100 GPU. Given the characterization and exploration, we propose several useful guidelines for both software optimization and hardware optimization for the efficient execution of GCNs on GPU.