tekdogan / gsuite

GNN-Inference Benchmark Suite. Experiment files (codes, datasets, etc.) of the IISWC'22 paper.

gSuite

• gSuite is a flexible and framework independent benchmark suite for Graph Neural Networks at Inference level on GPUs.
• It allows researchers and engineers to build any desired GNN pipeline by just passing a few parameters to program.
• Choice can include using either a GNN framework (such as PyG and DGL) or our implementations, which are not dependent to any other ML libraries (such as PyTorch or Tensorflow).

Further information can be found in the paper:

• IEEE Xplore link
• arXiv link

Software Architecture

⚠️ This repository may contain out of date code chunks, which may lead compile to fail.

Usage

Just execute the main script by passing a few parameters (see UI Parameters below).
Example usage:
python3 main.py --config "conf.json" --model "gcn" --dataset "cora"

UI Parameters

--config: (mandatory) local configuration file which includes default parameters
--gpu-id : (optional) device id
--model : (optional) GNN model
--dataset : (optional) dataset to be processed

Before Compiling

volta lab

export PATH=/okyanus/progs/cuda/cuda-10.1/bin:$PATH
export CPATH=/okyanus/progs/cuda/cuda-10.1/include:$CPATH
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH/okyanus/progs/cuda/cuda-10.1/lib64/

pascal lab

export PATH=/usr/local/cuda-8.0/bin:$PATH
export CPATH=/usr/local/cuda-8.0/include:$CPATH
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH/usr/local/cuda-8.0/lib64/

Profile via NVPROF

nvprof -f --analysis-metrics -o cudaDataLoader.nvprof ./cudaDataLoader.o --benchmark

lib compilation

nvcc -lcublas -std=c++11 -c -arch=compute_61 cuBlasUtil.cu Data_Util.cu CU_GCN_MP.cu CU_SpMM_GCN.cu cudaDataLoader.cu --compiler-options -fPIC
nvcc --shared -o libCU_SpMM_GCN.so cudaDataLoader.o cuBlasUtil.o Data_Util.o CU_SpMM_GCN.o --compiler-options -fPIC -std=c++11

tests

indexSelect:
nvcc -g --cudart shared -lcublas_static -lculibos -ldl -lpthread -lcudart -lcudadevrt -std=c++11 -gencode arch=compute_61,code=compute_61 index_select.cu Data_Util.cu test/indexSelect_test.cu -o test/indexSelect_test.o

linear:
nvcc -g --cudart shared -lcublas_static -lculibos -ldl -lpthread -lcudart -lcudadevrt -std=c++11 -gencode arch=compute_61,code=compute_61 linear.cu cuBlasUtil.cu Data_Util.cu test/linear_test.cu -o test/linear_test.o

scatter (aggregation):
nvcc -g --cudart shared -lcublas_static -lculibos -ldl -lpthread -lcudart -lcudadevrt -std=c++11 -gencode arch=compute_61,code=compute_61 scatter_cuda.cu Data_Util.cu test/scatter_test1.cu -o test/scatter_test1.o

scatter (node degrees):
nvcc -g --cudart shared -lcublas_static -lculibos -ldl -lpthread -lcudart -lcudadevrt -std=c++11 -gencode arch=compute_61,code=compute_61 scatter_cuda.cu Data_Util.cu test/scatter_test2.cu -o test/scatter_test2.o