HyperGef

This is a repo of Mlsys'23 paper[to apppear] HyperGef, a novel system of designing efficient fusion kernel for Hypergraph Nerual Network. Here we focus on accelerating HGNNConv and UniGNNConv, which are classic models of hypergraph GNN. By kernel fusion and workload balance, our backend can bring significant speedup compared to previous SpMM backend. We compare our implementation to the cuSPARSE baseline in the kernel setting, and compare to the PyG and DGL framework in the end2end setting.

Requirement

cuda >= 11.6
torch >= 1.11
dgl >= 0.9
pyg >= 2.1
pandas
sklearn
GPUtil

Note that currently pyg don't support torch 2.0+, so make sure you use pytorch 1.x version. Also, please download the optional dependencies of pyg like torch_scatter and torch_sparse libraries.

Setup

Please first config your $CUDA_HOME environment. It probably locates at /usr/local/cuda or something like that. Then config the ncu_report environment: export PYTHONPATH="${CUDA_HOME}/nsight-compute-xxxx.x.x/extras/python"

Build the project

bash build.sh

Download dataset

cd HyperGsys/data
bash prepare.sh

How to Run Model

Run HGNN with our backend

cd HyperGsys

python ugsys.py --backend hgsys --model-name HGNN

Compared with other backend (DGL, PyG)

python ugsys.py --backend dgl --model-name HGNN
python ugsys.py --backend pyg --model-name HGNN

Run C++ Hyper Aggregation

cd HyperGsys/source
make
./aggr_proto ../data/mtx_data/cora.mtx [feature_length,e.g: 32]

To verify functionality

cd test
pytest

To run Artifact Evaluation

cd experiment
make
python fig6.py
python fig7.py
python fig8.py
python fig9.py
python fig10.py

The result of fig8 will appear in profile/tables directory. Other result will be presented in figX.csv file.