XuLin: A Data-Centric Accelerator for High-Performance Hypergraph Processing

What's new?

XuLin is accepted to be MICRO 2022.

XuLin-F for HBM-Enabled FPGA platforms has been released.

Introduction

XuLin is a data-centric accelerator for high-performance hypergraph processing. We implement XuLin on a Xilinx Alveo U280 FPGA accelerator card. The FPGA version is referred to as XuLin-F. To our best knowledge and experiments, this is the fastest hypergraph graph processing accelerator.

Prerequisites

Hardware

This project works on Xilinx Alveo U280 Data Center Accelerator Card.

Operating System

Ubuntu 18.04 LTS

Software

Vitis Core Development Kit 2019.2

Alveo U280 Package File on Vitis 2019.2

and

Vitis Core Development Kit 2021.2

Alveo U280 Package File on Vitis 2021.2

Run the code

1. Install the software development environment according to Xilinx documentation. After installation, you can use commands

sudo /opt/xilinx/xrt/bin/xbmgmt flash --scan
/opt/xilinx/xrt/bin/xbutil validate

to make sure that the runtime environment and the alveo card is ready.

2. Add the xrt and deployment environment to your PATH. Typically you can run

source /opt/xilinx/xrt/setup.sh

3. You can clone and run the pre-build binary file as follows. The pre-build binary file is compiled by Vitis 2019.2.

$ git clone https://github.com/CGCL-codes/XuLin-F.git
$ cd XuLin-F/bin/
$ ./XuLin-F-PR ./XuLin-F-PR.xclbin <hypergraph dataset>

Datasets

The datasets come from the following sources:

• Stanford Large Network Dataset Collection (SNAP)
• Koblenz Network Collection (KONECT)

Files can be downloaded directly from SNAP and KONECT.

Here are the links to the corresponding files in SNAP or KONECT format:

• trec-wt(TW): KONECT
• com-Orkut(OK): SNAP
• LiveJournal(LJ): KONECT
• delicious-ut(DU): KONECT
• Orkut-group(OG): KONECT

Evaluation

Table1: Hypergraph Datasets

Datasets	Vertex	Hyperedge	Bipedge	Size
trec-wt(TW)	0.55M	1.17M	83.63M	1.1GB
com-Orkut(OK)	2.32M	15.30M	107.08M	1.8GB
LiveJournal(LJ)	3.20M	7.49M	112.31M	1.5GB
delicious-ut(DU)	0.83M	4.51M	301.19M	3.3GB
Orkut-group(OG)	2.78M	8.73M	327.03M	4.6GB

Table2: Performance comparison on PageRank(One Iteration) between Hygra and XuLin-F (16-threads/16-PEs configuration)

Datasets	Hygra on CPU(s)	XuLin-F on U280(s)	SpeedUp
trec-wt(TW)	0.053	0.0235	2.26
com-Orkut(OK)	0.159	0.0247	6.44
LiveJournal(LJ)	0.595	0.0296	20.10
delicious-ut(DU)	1.08	0.0951	11.36
Orkut-group(OG)	0.505	0.0759	6.65

Publications

[MICRO'22] Qinggang Wang, Long Zheng, Ao Hu, Yu Huang, Pengcheng Yao, Chuangyi Gui, Xiaofei Liao, Hai Jin, Jingling Xue, "A Data-Centric Accelerator for High-Performance Hypergraph Processing", in MICRO-55: 55th Annual IEEE/ACM International Symposium on Microarchitecture (MICRO'22), 2022.

Authors and Copyright

XuLin-F is developed in National Engineering Research Center for Big Data Technology and System, Cluster and Grid Computing Lab, Services Computing Technology and System Lab, School of Computer Science and Technology, Huazhong University of Science and Technology, Wuhan, China by Qinggang Wang (qgwang@hust.edu.cn), Ao Hu (ahu@hust.edu.cn), Long Zheng (longzh@hust.edu.cn).

Copyright (C) 2022, STCS & CGCL and Huazhong University of Science and Technology.