Near-Optimal Sparse Allreduce for Distributed Deep Learning (published in PPoPP'22)

Ok-Topk is a scheme for distributed training with sparse gradients. Ok-Topk integrates a novel sparse allreduce algorithm (less than 6k communication volume which is asymptotically optimal) with the decentralized parallel Stochastic Gradient Descent (SGD) optimizer, and its convergence is proved theoretically and empirically. All baselines, such as TopkDSA, gTopk, and Guassiank, are already integrated in the repo.

Setup the environment

To install the required Python modules:

conda create --name py38_oktopk python=3.8

conda activate py38_oktopk

pip3 install pip==20.2.4

pip install -r requirements.txt

MPICC="cc -shared" pip install --no-binary=mpi4py mpi4py

git clone https://github.com/NVIDIA/apex

cd apex

pip install -v --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./

Prepare Datasets

Cifar-10 for VGG

cd ./VGG/vgg_data

wget https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz

tar -zxvf cifar-10-python.tar.gz

AN4 for LSTM

cd ./LSTM/audio_data

wget https://www.dropbox.com/s/l5w4up20u5pfjxf/an4.zip

unzip an4.zip

Wikipedia for BERT

cd ./BERT/bert/bert_data/

Prepare the dataset according to the README file.

Run jobs

We run experiments on GPU clusters with SLURM job scheduler. To evaluate the performance of Ok-Topk, Gaussiank, gtopk, topkA, topkDSA, and dense, run the jobs as follows.

To run VGG jobs

cd ./VGG

./sbatch_vgg_jobs.sh

To run LSTM jobs

cd ./LSTM

./sbatch_lstm_jobs.sh

To run BERT jobs

cd ./BERT/bert/

./sbatch_bert_jobs.sh

Publication

The work of Ok-Topk is pulished in PPoPP'22. DOI

License

See LICENSE.