GLIST: Towards In-Storage Graph Learning

Authors: 

Cangyuan Li, Ying Wang, Cheng Liu, and Shengwen Liang, SKLCA, Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China; University of Chinese Academy of Sciences, Beijing, China; Huawei Li, SKLCA, Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China; University of Chinese Academy of Sciences, Beijing, China; Peng Cheng Laboratory, Shenzhen, China; Xiaowei Li, SKLCA, Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China; University of Chinese Academy of Sciences, Beijing, China

Abstract: 

Graph learning is an emerging technique widely used in diverse applications such as recommender system and medicine design. Real-world graph learning applications typically operate on large attributed graphs with rich information, which do not fit in the memory. Consequently, the graph learning requests have to go across the deep I/O stack and move massive data from storage to host memory, which incurs considerable latency and power consumption. To address this problem, we developed GLIST, an efficient in-storage graph learning system, to process graph learning requests inside SSDs. It has a customized graph learning accelerator implemented in the storage and enables the storage to directly respond to the graph learning requests. Thus, GLIST greatly reduces the data movement overhead in contrast to conventional GPGPU based systems. In addition, GLIST offers a set of high-level graph learning APIs and allows developers to deploy their graph learning service conveniently. Experimental results on an FPGA-based prototype show that GLIST achieves 13.2× and 10.1× average speedup and reduces the power consumption by up to 98.7% and 98.0% respectively on a series of graph learning tasks when compared to CPU and GPU based solutions.

Open Access Media

USENIX is committed to Open Access to the research presented at our events. Papers and proceedings are freely available to everyone once the event begins. Any video, audio, and/or slides that are posted after the event are also free and open to everyone. Support USENIX and our commitment to Open Access.

BibTeX
@inproceedings {273912,
author = {Cangyuan Li and Ying Wang and Cheng Liu and Shengwen Liang and Huawei Li and Xiaowei Li},
title = {{GLIST}: Towards In-Storage Graph Learning},
booktitle = {2021 {USENIX} Annual Technical Conference ({USENIX} {ATC} 21)},
year = {2021},
isbn = {978-1-939133-23-6},
pages = {225--238},
url = {https://www.usenix.org/conference/atc21/presentation/li-cangyuan},
publisher = {{USENIX} Association},
month = jul,
}