If you use our cluster (see access.md), just activate the prepared environment.

source /data/eurosysae/.venv/cxgnn/bin/activate

Else, you need to setup the environment and prepare the data:

Activate virtual environment:

mkdir ~/.venv
python3 -m venv ~/.venv/cxgnn
source ~/.venv/cxgnn/bin/activate

Install requirements. Make sure CxGNN-DL and the modified triton are cloned with --recurse-submodules and put aside with CxGNN-Compute.

After it, the directory tree should be like

.
|-- CxGNN-Compute
|-- CxGNN-DL
`-- triton

cd CxGNN-Compute
bash install.sh # this will install the prerequisites (e.g., CxGNN-DL) and CxGNN-Compute

All datasets are from OGB. We have pre-processed them for faster read. You can get access to them:

cd CxGNN-Compute
bash download.sh

After it, the directory tree should be like

.
|-- CxGNN-Compute
|-- CxGNN-DL
|-- triton
`-- data

Scripts and READMEs for experiments are put in test/ae/

If you meet any problem, please contact us through email (hkz20@mails.tsinghua.edu.cn) or HotCRP.

• Q: The program blocks when running overall test. A: Check the overall test readme to fix the performance bug in PyG.
• Q: There are CUDA OOM errors in overall test. A: Some baseline test will suffer from OOM, their number will not be displayed in the result file.
• Q: I can only run arxiv in the accuracy test. A: The node feature data of the other two datasets are too large and not uploaded to the cloud drive. If you are interested in them, please contact me for the full datasets.
• Q: Failed compilation. A: Make sure you have CUDA, cmake, and GCC in your path; And use --recurse-submodules when git clone