How uGrapher Boosts GNN Performance 3.5× with a Unified Graph Operator Abstraction

Alibaba Cloud Machine Learning Platform PAI and Shanghai Jiao Tong University’s team have had their paper “uGrapher: High‑Performance Graph Operator Computation via Unified Abstraction for Graph Neural Networks” accepted to ASPLOS 2023.

ASPLOS is a top‑tier systems conference that bridges architecture, programming languages, and operating systems, and has historically driven core innovations such as multicore processors, virtualization, RAID, RISC, and deep‑learning accelerators.

The acceptance highlights PAI’s advanced capabilities in high‑performance GNN computation and showcases Chinese machine‑learning system innovation on the global stage.

Graph neural networks introduce irregular graph operators whose variability and complexity increase with algorithmic evolution, making static hand‑written kernels inefficient across diverse models and datasets.

uGrapher addresses this by providing a unified high‑performance interface that abstracts all graph operators into a mixed sparse‑dense intermediate representation, decoupling computation from scheduling and enabling adaptive loop‑transform strategies on GPUs. Compared with DGL, PyG, and GNNAdvisor, uGrapher achieves an average 3.5× speedup.

The key designs of uGrapher are currently being integrated into PAI’s large‑scale GNN framework GraphLearn to accelerate industrial‑grade graph applications.

Paper details: Title : uGrapher: High‑Performance Graph Operator Computation via Unified Abstraction for Graph Neural Networks; Authors : Zhou Yangjie, Leng Jingwen, Song Yaoxu, Lu Shuwen, Wang Mian, Li Chao, Guo Minyi, Shen Wenting, Li Yong, Lin Wei, et al.; PDF : https://dl.acm.org/doi/10.1145/3575693.3575723.