GNNAdvisor: An Efficient Runtime System for GNN Acceleration on GPUs

Overview

GNNAdvisor: An Efficient Runtime System for GNN Acceleration
on GPUs

@inproceedings{GNNAdvisor,
  title={GNNAdvisor: An Efficient Runtime System for GNN Acceleration on GPUs},
  author={Yuke Wang, Boyuan Feng, Gushu Li, Shuangchen Li, Lei Deng, Yuan Xie, and Yufei Ding},
  booktitle={USENIX Symposium on Operating Systems Design and Implementation (OSDI'21)},
  year={2021}
}

1. Getting Started Instructions.

  • Clone this project
git clone --recursive [email protected]:YukeWang96/OSDI21_AE.git
  • Hardware:
  • CPU x86_64 with host memory >= 32GB. (Tested on Intel Xeon Silver 4110 (8-core 16-thread) CPU with 64GB host memory).
  • NVIDIA GPU (arch>=sm_60) with devcie memory >= 16GB. (Support NVIDIA Quadro P6000 (sm_61), Tesla V100 (sm_70), and RTX3090 (sm_86). Note that upon creating this artifact, we mainly evaluate our design on RTX3090. The execution time may be different across different devices but the overall trend of performance (speedup) is similar.
  • OS & Compiler:
  • Ubuntu 16.04+
  • gcc >= 7.5
  • cmake >= 3.14
  • CUDA >= 11.0 and nvcc >= 11.0
  • Important Files/Directories
  • dgl_baseline/: contains latest DGL implementation and python benchmark and result analysis scripts.
  • pyg_baseline/: contains latest PyG implementation and python benchmark and result analysis scripts.
  • Gunrock/: contains latest Gunrock implementation of SpMM kernel for neighbor aggregation and python benchmark script.
  • Docker/: contains docker file for setting up the compilation and running environment.
  • cu102/: dockerfile for sm < 80, such as Quadro P6000 and Tesla V100.
  • cu110/: dockerfile for sm >= 80, such as RTX 3090.
  • rabbit_module/: contains the source of rabbit reordering and python binding.
  • GNNAdvisor/: the directory for GNNAdvisor and Python benchmark and result analysis scripts.
  • GNNConv/: the C++/CUDA source code (GNNAdvisor_kernel.cu) for GNN sparse computation kernel, python binding of kernels (GNNAdvisor.cpp) and python setup.py installation script.
  • gnn_conv.py: the Python script for defining the GNN convolution at high-level.
  • param.py: the Python script for defining the input-level properties and different rules for handling this properties to generate performance-related configuration, such as warpPerBlock.
  • dataset.py: the Python loader for datasets from either plain .txt edgeList files or binary .npy file.
  • ./s7-4_1_neighbor_partitioning.py, ./s7-4_2_dimension_partitiong.py, ./s7-4_3_node_renumbering.py and ./s7-5_1_hidden_dimension.py are for running additional studies in our paper.
  • unitest.py: the Python script for verifying our basic sparse kernel.
  • osdi-ae-graphs/ containts the .npy files for all three Types of datasets.
  • osdi-ae-graphs-mtx/ containts the plain .mtx files for the Type III datasets for Gunrock SpMM kernel evaluation.

Step-1: Environment Setup

There are two ways to setup the environment of GNNAdvisor and baselines.

+ Method 1: Setup the environment via Docker (Recommended).

  • Install Docker Engine with NVIDIA GPU Support Toturial. We use the following commands
curl https://get.docker.com | sh \
  && sudo systemctl --now enable docker

distribution=$(. /etc/os-release;echo $ID$VERSION_ID) \
   && curl -s -L https://nvidia.github.io/nvidia-docker/gpgkey | sudo apt-key add - \
   && curl -s -L https://nvidia.github.io/nvidia-docker/$distribution/nvidia-docker.list | sudo tee /etc/apt/sources.list.d/nvidia-docker.list

sudo apt-get update
sudo apt-get install -y nvidia-docker2
sudo systemctl restart docker
  • cd Docker then either goto cu102/ (for Quadro P6000 and Tesla V100) or cu110/ (for RTX3090).
  • Run ./build.sh, it may takes a while (around 10 minutes) for building the container.
  • Run ./launch.sh then it will bring up an new interactive command line interface.
  • if your enounter problem below,
docker: Error response from daemon: could not select device driver "" with capabilities: [[gpu]].

then you need to

sudo apt-get update && sudo apt-get install -y nvidia-container-toolkit
sudo systemctl restart docker
  • The defualt GPU is device:0. If you want to run on different deivce. Consider using this command in ./launch.sh, e.g., using device:1
docker run -it --rm --gpus device=1 -v $PWD/../../:/GNNA osdi-ae:latest /bin/bash
  • Run ./install_pkg.sh to install the GNNAdvisor and rabbit_module. **Note: Select the correct sm version before install the package.
  • install_pkg_sm86.py for RTX3090.
  • install_pkg_sm70.py for Tesla V100.
  • install_pkg_sm61.py for Quadro P6000.
  • To clean the building packages when exit docker, run ./clean_build.sh, root access premission may required.

+ Method 2: Setup via conda and pip

1) Install system packages for compiling rabbit reordering (root user required).

  • libboost: sudo apt-get install libboost-all-dev.
  • tcmalloc: sudo apt-get install libgoogle-perftools-dev.
  • cmake: sudo apt-get update && sudo apt-get -y install cmake protobuf-compiler.

2) Install Pytorch environment.

  • Install conda on system Toturial.
  • Create a conda environment:
conda create -n env_name python=3.6
  • Install Pytorch:
conda install pytorch torchvision torchaudio cudatoolkit=11.1 -c pytorch -c conda-forge

or using pip [Note that make sure the pip you use is the pip from current conda environment. You can check this by which pip]

pip install torch==1.8.0+cu111 torchvision==0.9.0+cu111 torchaudio==0.8.0 -f https://download.pytorch.org/whl/torch_stable.html
pip install tqdm
pip install scipy
conda install -c dglteam dgl-cuda11.0
pip install torch requests
pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.8.0+cu111.html
pip install torch-sparse -f https://pytorch-geometric.com/whl/torch-1.8.0+cu111.html
pip install torch-cluster -f https://pytorch-geometric.com/whl/torch-1.8.0+cu111.html
pip install torch-spline-conv -f https://pytorch-geometric.com/whl/torch-1.8.0+cu111.html
pip install torch-geometric
  • Install GNNAdvisor Pytorch Binding.
  • Go to GNNAdvisor/GNNConv, then python setup.py install to install the GNNAdvisor modules.
  • Go to rabbit_module/src, then python setup.py install to install the rabbit reordering modules.

Step-2: Download the graph datasets.

  • Our preprocessed graph datasets in .npy format can be downloaded via this [LINK] (filename: osdi-ae-graphs.tar.gz).
  • Unzip the graph datasets tar -zxvf osdi-ae-graphs.tar.gz at the project root directory.
  • Note that node inital embeeding is not included, and we generate an all 1s embeeding matrix according to users input dimension parameter at the runtime for just performance evaluation.

3. Detailed Instructions.

  • GNN Model Setting.
  • GCN (2-layer with 16 hidden dimension)
  • GIN (5-layer with 64 hidden dimension)
  • Datasets.
  • Type I: citeseer, cora, pubmed, ppi
  • Type II: PROTEINS_full, OVCAR-8H, Yeast, DD, TWITTER-Real-Graph-Partial, SW-620H
  • Type III: amazon0505, artist, com-amazon, soc-BlogCatalog, amazon0601
  • Running DGL baseline on GNN training (Figure 9).
  • Go to dgl_baseline/ directory
  • /0_run_gcn.sh and ./0_run_gin.sh to run DGL and generate .csv result for GCN and GIN, respectively. Or you can run seperate commands,
  • ./0_bench_dgl_gcn.py| tee run_dgl_gcn.log to run the script and the report 200 epoch runtime for all evaluated datasets.
  • ./1_log2csv.py to convert the run_dgl_gcn.log to run_dgl_gcn.csv for ease of visualization.
  • Running PyG baseline on GNN training (Figure 10).
  • Go to pyg_baseline/ directory;
  • /0_run_gcn.sh and ./0_run_gin.sh to run PyG and generate .csv result for GCN and GIN, respectively. Or you can run seperate commands,
  • ./0_bench_pyg_gcn.py| tee run_pyg_gcn.log to run the script and the report 200 epoch runtime for all evaluated datasets.
  • ./1_log2csv.py run_pyg_gcn.log to convert log result to run_pyg_gcn.csv for ease of analysis.
  • Running Gunrock for single SpMM (neighbor aggregation) kernel.
  • We measure the single SpMM kernel performance with Gunrock (Note that based on most reviewers' feedback directly end-to-end inference comparison with Gunrock on sampled GraphSAGE model is not fair, therfore, we decide to compare our single SpMM kernel with Gunrock SpMM kernel).
  • Go to Gunrock/ directory. In case you forget to use --recursive when call git clone at the beginning, you need to call git submodule init && git submodule update to pull Gunrock repo.
  • Download the .mtx dataset of Type III graphs for Gunrock from this [LINK], then uncompress the .tar.gz file using tar -zxvf osdi-ae-graphs-mtx.tar.gz.
  • Under Gunrock/ call ./build_spmm.sh to build the Gunrock spmm kernel. (it may take for a while for complete).
  • ./0_bench_Gunrock.py for profile spmm. The instruction to run single neighbor aggregation kernel for GNNAdvisor can be found below by specifying an command line option.
  • Note that Running Gunrock experiments does not require the Docker environment, and it require the system has CMake >= 3.14, and CMake==3.20.0 is recommended.
  • Running GNNAdvisor (Figure 9 and Figure 10).
  • Go to GNNAdvisor/ directory.
  • ./0_run_gcn.sh and ./0_run_gin.sh to run GNNAdvisor and generate .csv result for GCN and GIN, respectively. Or you can run seperate command with different configurations as
  • ./0_bench_GNNA_GCN.py| tee run_GNNA_GCN.log to run the script and the report 200 epoch runtime for all evaluated datasets. Note that there are also several options (such as enable_rabbit) for configuring a profiling.
  • ./1_log2csv.py to convert the run_GNNA_GCN.log to run_GNNA_GCN.csv for ease of result analysis.
  • ./3_single_spmm_bench.py to profile a single SpMM kernel to compare with Gunrock SpMM kernel discussed above.
  • Stand alone running with specified parameters.
  • --dataset: the name of the dataset.
  • --dim: the size of input embedding dimension, default: 96.
  • --hidden: the size of hidden dimension, default: 16.
  • --classes: the number of output classes, default: 22.
  • --partSize: the size of neighbor-group, default: 32.
  • --dimWorker: the number of worker threads (<=32), default: 32.
  • --warpPerBlock: the number of warp per block, default: 8, recommended: GCN: (8), GIN: (2 for citeseer, 8 for remaining datasets).
  • --sharedMem: the shared memory size for each Stream-Multiprocessor on NVIDIA GPUs. A reference for different GPU architecture and its shared memory size can be found at here, default 96KB for RTX3090.
  • --model: gcn or gin. The evaluated example GCN model has 2 layers with 16 hidden dimensions, while the example GIN model has 5 layers with 64 hidden dimensions.
  • --num_epoches: the number of epoches for training, default: 200.
  • --loadFromTxt: If this flag is True, it will load the graph TXT edge list, where each line is an s1 d1. default: False (load from .npz which is fast).
  • --enable_rabbit: If this flag is True, it will be possible to use the rabbit-reordering routine. Otherwise, it will skip rabbit reordering for both auto and manual mode.
  • --manual_mode: If this flag is True, it will use the value from the parameter partSize, dimWorker and dimWorker. Otherwise, it will determine these three performance-related parameters automatically by Decider. Note that Decider will generate two different sets of parameters for input and hidden layers based on a GNN model and the dataset input characters. In manual mode the value of partSize, dimWorker and dimWorker will be applied to both input and hidden layer.
  • --verbose_mode: If this flag is True, it will print out all the details of configuration for running the experiments.
  • --single_spmm: If this flag is True, it will only profile a single spmm for 200 rounds. with the provided --dim as the D in NxNxD, where N is the number of nodes in a graph. Run ./3_single_spmm_bench.py for profiling single neighbor aggregation (SpMM) kernel in comparison with Gunrock SpMM.
  • --verify_spmm: If this flag is True, it will check the correctness of our SpMM kernel against the CPU reference result. Run ./4_verifying.py for verifying our major kernel (neighbor aggregation) correctness against CPU reference result from torch_sparse.spmm.

Note

  • Accuracy evaluation are omitted for all implementations and each sparse kernels are tested via the unitest.py
  • We focus on the training evaluation of the GNNs, and the reported time per epoch only includes the GNN model forward and backward computation, excluding the data loading and some preprocessing.
  • Since the paper draft submission and the creation of this artifact, DGL has update several of its kernel library (from v0.52 to v0.60). In this comparion we focus on the latest DGL version (v0.60).
  • Based on our profiling on RTX3090 and Quadro P6000, our design would show minor speedup on the simple GCN model (2-layer and 16 hidden dimension), but show more evident speedup on more complicated GIN model (5-layer and 64 hidden dimension), which can still demonstrate the effectiveness of our optimizations.
  • Our observation is that on small Type I graphs, our frameworks achieve significant speedup for both GCN and GIN model on RTX3090 and Quadro P6000. On larger Type II and Type III datasets, our GIN model implementation would show more evident speedups.
  • Running GNNAdvisor-related Studies (Figure 11(a,b,c) and Figure 12(a))
  • ./s7-4_1_neighbor_partitioning.py(Figure 11a) for neighbor partitioning study in Section 7.4.
  • ./s7-4_2_dimension_partitiong.py (Figure 11b) for dimension partitioning study in Section 7.4.
  • ./s7-4_3_node_renumbering.py (Figure 11c) for node renumbering study in Section 7.4.
  • ./s7-5_1_hidden_dimension.py (Figure 12a) for hidden dimension study in Section 7.5.
  • (Recommended) You can run all studies by simply running ./2_run_study.sh, it will first output all runtime collected information (e.g., average training epoch time) as a *.log file, then it will automically call ./2_study2csv.py to generate the corresponding *.csv for ease of analysis. You expected to get several .csv files looks like these (open with the Edit CSV plugin for vscode)
  • For neighbor_partitioning. Neighbor Partitioning
  • For dimension_partitiong. Dimension Partitioning
  • For hidden_dimension. Dimension Partitioning
  • For node_renumbering.
    Dimension Partitioning

Reference

  • Deep Graph Library
    Wang, Minjie, et al. Deep graph library: A graph-centric, highly-performant package for graph neural networks.. The International Conference on Learning Representations (ICLR), 2019.
  • Pytorch Geometric
    Fey, Matthias, and Jan Eric Lenssen. Fast graph representation learning with PyTorch Geometric. The International Conference on Learning Representations (ICLR), 2019.
  • Gunrock
    Wang, Yangzihao, et al. Gunrock: A high-performance graph processing library on the GPU. Proceedings of the 21st ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming (PPoPP), 2016.
  • Rabbit Order
    J. Arai, H. Shiokawa, T. Yamamuro, M. Onizuka, and S. Iwamura. Rabbit Order: Just-in-time Parallel Reordering for Fast Graph Analysis. IEEE International Parallel and Distributed Processing Symposium (IPDPS), 2016.
  • GE-SpMM. Guyue Huang, Guohao Dai, Yu Wang and Huazhong Yang.
    GE-SpMM: General-purposed Sparse Matrix-Matrix Multiplication on GPUs for Graph Neural Networks The International Conference for High Performance Computing, Networking, Storage and Analysis, 2020.
Comments
  • Does Fast-Math Intrinsics Align with Baseline?

    Does Fast-Math Intrinsics Align with Baseline?

    Hi, I am very interested in your OSDI21 work. However, I noticed that you used __fmaf_rn in your repo. This is a fast-math intrinsics according to documentation https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#intrinsic-functions, while I observed that usually nvcc will heavily emit multiply-add instruction even though expressions are written in naive form and no such intrinsics are used. I am not sure if using this intrinsic aligns with the baseline and how this intrinsic could help you achieve your goal. Could you explain them to me? Thank you.

    opened by K-Wu 2
  • Fair comparison with DGL

    Fair comparison with DGL

    I notice that the baseline models (directly imported from DGL) include some operations such as BatchNorm, while your models in the codebase do not. So do your models produce the correct output as in the original GNN paper? Is that the reason for the shorter latency than DGL in the paper?

    opened by PeterSwiss 2
  • Errors raised for large datasets like Reddit

    Errors raised for large datasets like Reddit

    Thanks for your good work. But I find your code does not work for large datasets like Reddit with around 233k nodes and 115m edges. It raises the error 'CUDA error: an illegal memory access was encountered'.

    I perform further experiments. I find the code works fine if I keep only 30m edges. But the same error is raised when I keep 40m edges.

    I also test it on other large datasets like LiveJournal-d and com-orkut. They have similar performance.

    Best wishes

    opened by Yangxc13 1
  • Question about GCN/GIN forward propagation

    Question about GCN/GIN forward propagation

    Hi Yuke,I read your paper and code, of course excellent work and thanks for the repository. I have question about the cuda kernel. I can see you computing GCN forward propagation in CUDA like this: https://github.com/YukeWang96/OSDI21_AE/blob/5c4f561aa27228164fc8c35d2468ea2cd9dc29ea/GNNAdvisor/GNNConv/GNNAdvisor_kernel.cu#L389 https://github.com/YukeWang96/OSDI21_AE/blob/5c4f561aa27228164fc8c35d2468ea2cd9dc29ea/GNNAdvisor/GNNConv/GNNAdvisor_kernel.cu#L399-L406 In my opinion, 1/degree_norm_inv should be used according to GCN paper, am I right? And in GIN cuda kernel: https://github.com/YukeWang96/OSDI21_AE/blob/5c4f561aa27228164fc8c35d2468ea2cd9dc29ea/GNNAdvisor/GNNConv/GNNAdvisor_kernel.cu#L676-L689 I think src node embedding $h^{k-1}_v$ should be used in forward propagation according to GIN paper. However, I can only see you aggregate the neighbor $h^{k-1}_u$ Thank you!

    opened by yofufufufu 4
  • Question: about ngs/partSize choosing

    Question: about ngs/partSize choosing

    Hi Yuke, thanks for your excellent work!

    I have a question about the selection of ngs parameter. In the Section 6 of the paper, I only find some constrains for the ngswithout direct formula. In the code repo, I find that in auto model ngs = avgNodeDegree, is this the final criteria of ngs?

    https://github.com/YukeWang96/OSDI21_AE/blob/f129823ae49f3b557ef525aaa189fc5c703e5c59/GNNAdvisor/param.py#L73

    In the settings I of Fig 14, the black dot gives ngs=1024 for amazon0505, but the average degree is about 12 for this dataset. I am confused about this, could you please tell me where the problem is? Thank you!

    opened by initzhang 4
  • Question: warp-based thread alignment

    Question: warp-based thread alignment

    Hi,

    Your work - GNNAdvisor, is fantastic and many thanks for creating this repository to share the experiment source code of GNNAdvisor.

    I have one doubt on warp-based thread alignment. In section4.3 of GNNAdvisor OSDI's paper, you said, "warp-aligned thread mapping can merge memory requests from the same warp into one global memory transaction." How to achieve the warp-aligned actions in CUDA: changing the number of warps per block or other thread scheduling/orchestration? Could you please give some guidance on where the source code is and how it works? Thanks for your time and help.

    Best wishes,

    opened by sdkjksfd 5
Releases(v0.6)
Owner
YUKE WANG
https://wang-yuke.com
YUKE WANG
Theano is a Python library that allows you to define, optimize, and evaluate mathematical expressions involving multi-dimensional arrays efficiently. It can use GPUs and perform efficient symbolic differentiation.

============================================================================================================ `MILA will stop developing Theano <https:

null 9.6k Jan 6, 2023
Theano is a Python library that allows you to define, optimize, and evaluate mathematical expressions involving multi-dimensional arrays efficiently. It can use GPUs and perform efficient symbolic differentiation.

============================================================================================================ `MILA will stop developing Theano <https:

null 9.3k Feb 12, 2021
Lowest memory consumption and second shortest runtime in NTIRE 2022 challenge on Efficient Super-Resolution

FMEN Lowest memory consumption and second shortest runtime in NTIRE 2022 on Efficient Super-Resolution. Our paper: Fast and Memory-Efficient Network T

null 33 Dec 1, 2022
Tensors and Dynamic neural networks in Python with strong GPU acceleration

PyTorch is a Python package that provides two high-level features: Tensor computation (like NumPy) with strong GPU acceleration Deep neural networks b

null 61.4k Jan 4, 2023
Tensors and Dynamic neural networks in Python with strong GPU acceleration

PyTorch is a Python package that provides two high-level features: Tensor computation (like NumPy) with strong GPU acceleration Deep neural networks b

null 46.1k Feb 13, 2021
SLIDE : In Defense of Smart Algorithms over Hardware Acceleration for Large-Scale Deep Learning Systems

The SLIDE package contains the source code for reproducing the main experiments in this paper. Dataset The Datasets can be downloaded in Amazon-

Intel Labs 72 Dec 16, 2022
MASA-SR: Matching Acceleration and Spatial Adaptation for Reference-Based Image Super-Resolution (CVPR2021)

MASA-SR Official PyTorch implementation of our CVPR2021 paper MASA-SR: Matching Acceleration and Spatial Adaptation for Reference-Based Image Super-Re

DV Lab 126 Dec 20, 2022
DI-HPC is an acceleration operator component for general algorithm modules in reinforcement learning algorithms

DI-HPC: Decision Intelligence - High Performance Computation DI-HPC is an acceleration operator component for general algorithm modules in reinforceme

OpenDILab 185 Dec 29, 2022
Neural Fixed-Point Acceleration for Convex Optimization

Licensing The majority of neural-scs is licensed under the CC BY-NC 4.0 License, however, portions of the project are available under separate license

Facebook Research 27 Oct 6, 2022
Calculates JMA (Japan Meteorological Agency) seismic intensity (shindo) scale from acceleration data recorded in NumPy array

shindo.py Calculates JMA (Japan Meteorological Agency) seismic intensity (shindo) scale from acceleration data stored in NumPy array Introduction Japa

RR_Inyo 3 Sep 23, 2022
A python package simulating the quasi-2D pseudospin-1/2 Gross-Pitaevskii equation with NVIDIA GPU acceleration.

A python package simulating the quasi-2D pseudospin-1/2 Gross-Pitaevskii equation with NVIDIA GPU acceleration. Introduction spinor-gpe is high-level,

null 2 Sep 20, 2022
ThunderSVM: A Fast SVM Library on GPUs and CPUs

What's new We have recently released ThunderGBM, a fast GBDT and Random Forest library on GPUs. add scikit-learn interface, see here Overview The miss

Xtra Computing Group 1.4k Dec 22, 2022
ThunderGBM: Fast GBDTs and Random Forests on GPUs

Documentations | Installation | Parameters | Python (scikit-learn) interface What's new? ThunderGBM won 2019 Best Paper Award from IEEE Transactions o

Xtra Computing Group 647 Jan 4, 2023
🔮 Execution time predictions for deep neural network training iterations across different GPUs.

Habitat: A Runtime-Based Computational Performance Predictor for Deep Neural Network Training Habitat is a tool that predicts a deep neural network's

Geoffrey Yu 44 Dec 27, 2022
BERT model training impelmentation using 1024 A100 GPUs for MLPerf Training v1.1

Pre-trained checkpoint and bert config json file Location of checkpoint and bert config json file This MLCommons members Google Drive location contain

SAIT (Samsung Advanced Institute of Technology) 12 Apr 27, 2022
Implementation for the paper 'YOLO-ReT: Towards High Accuracy Real-time Object Detection on Edge GPUs'

YOLO-ReT This is the original implementation of the paper: YOLO-ReT: Towards High Accuracy Real-time Object Detection on Edge GPUs. Prakhar Ganesh, Ya

null 69 Oct 19, 2022
Build and run Docker containers leveraging NVIDIA GPUs

NVIDIA Container Toolkit Introduction The NVIDIA Container Toolkit allows users to build and run GPU accelerated Docker containers. The toolkit includ

NVIDIA Corporation 15.6k Jan 1, 2023
Lunar is a neural network aimbot that uses real-time object detection accelerated with CUDA on Nvidia GPUs.

Lunar Lunar is a neural network aimbot that uses real-time object detection accelerated with CUDA on Nvidia GPUs. About Lunar can be modified to work

Zeyad Mansour 276 Jan 7, 2023