Running summarizing tools on a new article

Check the paper below, then start its synthesis.


HGL: Accelerating Heterogeneous GNN Training with Holistic Representation and Optimization

Yuntao Gui (Chinese University of Hong Kong), Yidi Wu (Chinese University of Hong Kong), Han Yang (Chinese University of Hong Kong), Tatiana Jin (Chinese University of Hong Kong), Boyang Li (Chinese University of Hong Kong), Qihui Zhou (Chinese University of Hong Kong), James Sheung-Chak Cheng (Chinese University of Hong Kong), Fan Yu (Huawei Technologies (China))

International Conference for High Performance Computing, Networking, Storage and Analysis (SC)
License: not stated (all rights reserved)

Abstract: Graph neural networks (GNNs) have shown to significantly improve graph analytics. Existing systems for GNN training are primarily designed for homogeneous graphs. In industry, however, most graphs are actually heterogeneous in nature (i.e., having multiple types of nodes and edges). Existing systems train a heterogeneous GNN (HetGNN) as a composition of homogeneous GNN (HomoGNN) and thus suffer from critical limitations such as lack of memory optimization and limited operator parallelism. To address these limitations, we propose HGL - a heterogeneity-aware system for GNN training. At the core of HGL is an intermediate representation, called HIR, which provides a holistic representation for GNNs and enables cross-relation optimization in HetGNN training. We devise tailored optimizations on HIR, including graph stitching, operator fusion and operator bundling. Compared with DGL and PyG, HGL achieves a speedup from 7 to 22 times for training HetGNNs.

Published on 01 Nov. 2022