MiCRO is abbreviation of 'Minimizing compression ratio error on-the-fly'. MiCRO sparsifies gradients in partitioned gradient vectors, and estimates accurate threshold for sparsification condition by minimizing compression ratio error. Benchmarks include: 1) image classification using CNNs, 2) language modelling using LSTMs cnn-lstm, and 3) recommendation using NCF ncf.

To install the necessary dependencies, create Conda environment using environment.yml by running the following commands. Note: check compatibility of each package version such as cudatoolkit and cudnn with your device, e.g., NVIDIA Tesla V100 GPU is compatible.

$ conda env create --file environment.yml
$ conda activate micro
$ python -m spacy download en
$ conda deactivate micro

The scripts to run code are written for SLURM workload manager. The source code supports distributed training with multi-node and multi-GPU. In run.sh, you can specify model, dataset, reducer, and world_size.

• If you use SLURM, use pararun and modify it for your configuration. The script pararun executes run.sh in parallel. The script run.sh includes setup for distributed training.
• If you do not use SLURM, you do not need to use pararun. Instead, run run.sh on your nodes, then rendezvous of pytorch allows processes are connected.

• If you use SLURM, use following command.

$ sbatch pararun

• If you do not use SLURM, use following command on each node.

$ hostip=<ip> port=<port> mpirun -np <world_size> run.sh

• To download dataset, use following command.

$ ./prepare_dataset.sh

• If you use SLURM, use following command.

$ sbatch pararun

• If you do not use SLURM, use following command on each node.

$ hostip=<ip> port=<port> mpirun -np <world_size> run.sh

If you use this code, please cite the following [Paper]:

• MiCRO: Near-Zero Cost Gradient Sparsification for Scaling and Accelerating Distributed DNN Training. Daegun Yoon, Sangyoon Oh. HiPC 2023, Dec. 2023.

@article{yoon2023micro,
  title={MiCRO: Near-Zero Cost Gradient Sparsification for Scaling and Accelerating Distributed DNN Training},
  author={Yoon, Daegun and Oh, Sangyoon},
  journal={arXiv preprint arXiv:2310.00967},
  year={2023}
}

If you have any questions about this project, contact me by one of the followings:

• slashxp@naver.com
• kljp@ajou.ac.kr