Home

Awesome

<div align=center> <img src="./img/hetu.png" width="300" /> </div>

HETU

<!--- [![license](https://img.shields.io/github/license/apache/zookeeper?color=282661)](LICENSE) --->

Documentation | Examples

Hetu is a high-performance distributed deep learning system targeting trillions of parameters DL model training, developed by <a href="https://cuibinpku.github.io" target="_blank" rel="nofollow">DAIR Lab</a> at Peking University. It takes account of both high availability in industry and innovation in academia, which has a number of advanced characteristics:

We welcome everyone interested in machine learning or graph computing to contribute codes, create issues or pull requests. Please refer to Contribution Guide for more details.

Key Features

<div align=center> <img src="./img/features.png" width="800" /> </div>

Installation

  1. Clone the repository.

  2. Prepare the environment. We use Anaconda to manage packages. The following command create the conda environment to be used:conda env create -f environment.yml. Please prepare Cuda toolkit and CuDNN in advance.

  3. We use CMake to compile Hetu. Please copy the example configuration for compilation by cp cmake/config.example.cmake cmake/config.cmake. Users can modify the configuration file to enable/disable the compilation of each module. For advanced users (who not using the provided conda environment), the prerequisites for different modules in Hetu is listed in appendix.

# modify paths and configurations in cmake/config.cmake

# generate Makefile
mkdir build && cd build && cmake ..

# compile
# make all
make -j 8
# make hetu, version is specified in cmake/config.cmake
make hetu -j 8
# make allreduce module
make allreduce -j 8
# make ps module
make ps -j 8
# make geometric module
make geometric -j 8
# make hetu-cache module
make hetu_cache -j 8
  1. Prepare environment for running. Edit the hetu.exp file and set the environment path for python and the path for executable mpirun if necessary (for advanced users not using the provided conda environment). Then execute the command source hetu.exp .

Basic Usage

Train logistic regression on gpu:

bash examples/cnn/scripts/hetu_1gpu.sh logreg MNIST

Train a 3-layer mlp on gpu:

bash examples/cnn/scripts/hetu_1gpu.sh mlp CIFAR10

Train a 3-layer cnn with gpu:

bash examples/cnn/scripts/hetu_1gpu.sh cnn_3_layers MNIST

Train a 3-layer mlp with allreduce on 8 gpus (use mpirun):

bash examples/cnn/scripts/hetu_8gpu.sh mlp CIFAR10

Train a 3-layer mlp with PS on 1 server and 2 workers:

# in the script we launch the scheduler and server, and two workers
bash examples/cnn/scripts/hetu_2gpu_ps.sh mlp CIFAR10

More Examples

Please refer to examples directory, which contains CNN, NLP, CTR, MoE, GNN training scripts. If you want more, please join our community and contact with us!

Community

Enterprise Users

If you are enterprise users and find Hetu is useful in your work, please let us know, and we are glad to add your company logo here.

<img src="./img/tencent.png" width = "200"/> <img src="./img/alibabacloud.png" width = "200"/> <img src="./img/kuaishou.png" width = "200"/>

License

The entire codebase is under license

Papers

  1. Xupeng Miao, Hailin Zhang, Yining Shi, Xiaonan Nie, Zhi Yang, Yangyu Tao, Bin Cui. HET: Scaling out Huge Embedding Model Training via Cache-enabled Distributed Framework. VLDB 2022 (Best Scalable Data Science Paper), ChinaSys 2021 Winter
  2. Xupeng Miao, Yining Shi, Hailin Zhang, Xin Zhang, Xiaonan Nie, Zhi Yang, Bin Cui. HET-GMP: a Graph-based System Approach to Scaling Large Embedding Model Training. SIGMOD 2022, ChinaSys 2022
  3. Xupeng Miao, Yujie Wang, Youhe Jiang, Chunan Shi and Xiaonan Nie, Hailin Zhang, Bin Cui. Galvatron: Efficient Transformer Training over Multiple GPUs Using Automatic Parallelism. VLDB 2023
  4. Xupeng Miao, Xiaonan Nie, Yingxia Shao, Zhi Yang, Jiawei Jiang, Lingxiao Ma, Bin Cui. Heterogeneity-Aware Distributed Machine Learning Training via Partial Reduce. SIGMOD 2021
  5. Xupeng Miao, Lingxiao Ma, Zhi Yang, Yingxia Shao, Bin Cui, Lele Yu, Jiawei Jiang. CuWide: Towards Efficient Flow-based Training for Sparse Wide Models on GPUs. TKDE 2021, ICDE 2021.
  6. Xiaonan Nie, Xupeng Miao, Zhi Yang, Bin Cui. TSplit: Fine-grained GPU Memory Management for Efficient DNN Training via Tensor Splitting. ICDE 2022.
  7. Sicong Dong, Xupeng Miao, Pengkai Liu, Xin Wang, Bin Cui, Jianxin Li. HET-KG: Communication-Efficient Knowledge Graph Embedding Training via Hotness-Aware Cache. ICDE 2022.
  8. Xupeng Miao, Yujie Wang, Jia Shen, Yingxia Shao, Bin Cui. Graph Neural Network Training Acceleration over Multi-GPUs. Journal of Software (Chinese).
  9. Xiaonan Nie, Shijie Cao, Xupeng Miao, Lingxiao Ma, Jilong Xue, Youshan Miao, Zichao Yang, Zhi Yang, Bin Cui. EvoMoE: An Evolutional Mixture-of-Experts Training Framework via Dense-To-Sparse Gate. arXiv 2021.
  10. coming soon

Cite

If you use Hetu in a scientific publication, we would appreciate citations to the following paper:

 @article{miao2021het,
   title={HET: Scaling out Huge Embedding Model Training via Cache-enabled Distributed Framework},
   author={Miao, Xupeng and Zhang, Hailin and Shi, Yining and Nie, Xiaonan and Yang, Zhi and Tao, Yangyu and Cui, Bin},
   journal = {Proc. {VLDB} Endow.},
   volume = {15},
   number = {2},
   pages = {312--320},
   year = {2022},
   publisher = {VLDB Endowment}
 }

Acknowledgements

We learned and borrowed insights from a few open source projects including TinyFlow, autodist, tf.distribute, FlexFlow and Angel.

Appendix

The prerequisites for different modules in Hetu is listed as follows:

"*" means you should prepare by yourself, while others support auto-download

Hetu: OpenMP(*), CMake(*)
Hetu (version mkl): MKL 1.6.1
Hetu (version gpu): CUDA 10.1(*), CUDNN 7.5(*), CUB 1.12.1(*)
Hetu (version all): both

Hetu-AllReduce: MPI 3.1, NCCL 2.8(*), this module needs GPU version

Hetu-PS: Protobuf(*), ZeroMQ 4.3.2

Hetu-Geometric: Pybind11(*), Metis(*)

Hetu-Cache: Pybind11(*), this module needs PS module

##################################################################
Tips for preparing the prerequisites

Preparing CUDA, CUDNN, CUB, NCCL(NCCl is already in conda environment):
1. download from https://developer.nvidia.com 
2. download CUB from https://github.com/NVIDIA/cub/releases/tag/1.12.1
3. install
4. modify paths in cmake/config.cmake if necessary

Preparing OpenMP:
Your just need to ensure your compiler support openmp.

Preparing CMake, Protobuf, Pybind11, Metis:
Install by anaconda: 
conda install cmake=3.18 libprotobuf pybind11=2.6.0 metis

Preparing OpenMPI (not necessary):
install by anaconda: `conda install -c conda-forge openmpi=4.0.3`
or
1. download from https://download.open-mpi.org/release/open-mpi/v4.0/openmpi-4.0.3.tar.gz
2. build openmpi by `./configure /path/to/build && make -j8 && make install`
3. modify MPI_HOME to /path/to/build in cmake/config.cmake

Preparing MKL (not necessary):
install by anaconda: `conda install -c conda-forge onednn`
or
1. download from https://github.com/intel/mkl-dnn/archive/v1.6.1.tar.gz
2. build mkl by `mkdir /path/to/build && cd /path/to/build && cmake /path/to/root && make -j8` 
3. modify MKL_ROOT to /path/to/root and MKL_BUILD to /path/to/build in cmake/config.cmake 

Preparing ZeroMQ (not necessary):
install by anaconda: `conda install -c anaconda zeromq=4.3.2`
or
1. download from https://github.com/zeromq/libzmq/releases/download/v4.3.2/zeromq-4.3.2.zip
2. build zeromq by 'mkdir /path/to/build && cd /path/to/build && cmake /path/to/root && make -j8`
3. modify ZMQ_ROOT to /path/to/build in cmake/config.cmake