ParamSpMM_artifacts

The artifact for ParamSpMM: Empowering Graph Neural Networks with High-Performance Parameterized Sparse Matrix-Matrix Multiplication on GPUs.

environment requirements

OS: Ubuntu 20.04.5 LTS (x86_64)
GCC version: (Ubuntu 9.4.0-1ubuntu1~20.04.1) 9.4.0
CUDA runtime version: 11.6.124

environment setup

# install libraries for compiling rabbit reordering
apt-get update 
apt-get install -y libboost-all-dev libgoogle-perftools-dev
# create ParamSpMM conda enviroment
conda create -n param python=3.9
# install pytorch 1.13
conda activate param
conda install pytorch==1.13.0 torchvision==0.14.0 torchaudio==0.13.0 pytorch-cuda=11.6 -c pytorch -c nvidia
# install DGL
conda install -c dglteam dgl-cuda11.6==0.9.1post1
# install neccessary libraries
pip install ssgetpy
pip install pandas
pip install ogb
pip install scikit-learn==1.1.3
pip install ipykernel
pip install seaborn

All the following scripts are executed under Anaconda environment param

Use ParamSpMM as a tool

• In directory /script: run bash build_param.sh to build and bind ParamSpMM and ParamSDDMM into pytorch. The ML models for SpMM-decider are already available in directory /model.

• To run ParamSpMM with different column dimensions:

     # path: the path of *.mtx file
     # mtx_name: the name of the sparse matrix
     # dim_list: the column dimensions of dense input matrices 
     python decider.py --path ./toy_dataset/ --mtx_name pubmed --dim_list 16

• To run GCN:

     # num_classes: dimension of output classes
     # num_features: dimension of features
     # num_epoches: training epoches
     # hidden_list: hidden layer size 
     python GNN.py --path ./toy_dataset/ --mtx_name pubmed --num_classes 16 \
        --num_features 16 --num_epoches 200 --hidden_list 16 --num_layers 5

• To run GIN:

     python GNN.py --net GIN --path ./toy_dataset/ --mtx_name pubmed --num_classes 16 \
       --num_features 16 --hidden_list 32 --num_layers 5 

• To run AGNN embedded with ParamSpMM and ParamSDDMM:

     python GNN.py --net AGNN --path ./toy_dataset/ --mtx_name pubmed --num_classes 16 \
       --num_features 16 --hidden_list 32 --num_layers 5 

Rebuild SpMM-decider from scratch

Despite we provide the trained ML models of SpMM-decider in directory /models, you can still train your own SpMM-decider based on the following steps:

step 0: prepare graph datasets

For easy reproduction, we provide the processed matrix datasets Here. The matrices used for ParamSpMM evaluation and SpMM-decieder training are in ./all-mtx. The rabbit-reordered matrices are in ./all-mtx/rabbit. The names of all the 202 matrices and their features are in mtx_info.csv. The 6 large graphs for GNN training are in ./OGB. The rabbit-reordered matrices are in ./OGB/rabbit. After finishing downloading, try to unzip the file with:

bash 0_unzip.sh

step 1: construct the training set for SpMM-decider

In this step, we construct the training set for SpMM-decider by collecting the optimal configurations of ParamSpMM across all the matrices. Baseline SpMM libraries including cuSPARSE and GE-SpMM are also evaluated in this part.

In directory /script:

• run bash 2_construct_training_set.sh As a result, the performances of ParamSpMM, cuSPARSE and GE-SpMM are written to ParamSpMM-log. For example,
  • dim16.csv stores the throughputs of ParamSpMM under different configurations from reduced selection space.
  • dim16_OP_SpMM.csv contains the matrices information and their optimal ParamSpMM configurations, which are used as the training set for SpMM-decider when dim=16.
  • dim16_basline.csv contains the throughput of cuSPARSE and GE-SpMM when dim=16.

The throughputs when $dim \in {16, 32, 48, 64, 80, 96, 112, 128, 144, 160, 176, 192, 208, 224, 240, 256}$ are evaluated.

step 2: Model training, the performance of SpMM-decider

We choose random forest models for predicting the optimal configuration of ParamSpMM in SpMM-decider. The training process in rnd_tree.ipynb. The trained random forests models are then serialized pickle files and stored in ./models, which can be used repeatedly.