This repo is for the SOSP'23 artifact evaluation of paper "PIT: Optimization of Dynamic Sparse Deep Learning Models via Permutation Invariant Transformation".

• Artifacts Available: The source code of Sparta is available at: https://github.com/microsoft/SparTA/tree/pit_artifact

• Artifacts Functional: Documentation: the following document includes detailed guidelines on how to build, install, test PIT and the experiments to compare with other baselines.

• Results Reproduced: To reproduce the main results presented in our paper, we provide a Docker image containing all the environments and baseline softwares. We also provide detailed guideline to help reproduce the results step by step.

1. V100 experiments:

• 8x V100 32GB GPUs with more than 700GB disk space

2. A100 experiments

• 1x A100 80GB GPU with more than 200GB disk space

First, git clone the source code.

git clone https://github.com/microsoft/SparTA
cd SparTA && git checkout pit_artifact

To make the reproducing easier, we provide a docker image that contains all dependencies and baselines. Build the docker image:

cd image
// for the experiments on V100
sudo docker build . -f Dockerfile -t artifact:v100
// for the experiments on A100
sudo docker build . -f Dockerfile.a100 -t artifact:a100

The docker image compilation takes approximately an hour. To save time, we also provide pre-compiled images on the DockerHub: zhengningxin/pit_artifact:v100, zhengningxin/pit_artifact:a100.

If you meet out of disk space error, you can mount an external disk to the /data/ directory inside the docker container. For example:

docker run -it --gpus all --shm-size=150gb -v /tmp/data_docker:/data --rm -d --name pit_artifact -h docker zhengningxin/pit_artifact:a100

docker run -it --gpus all --shm-size=32gb --rm -d --name pit_artifact -h docker zhengningxin/pit_artifact:v100
docker exec -it pit_artifact bash
mkdir -p workspace && cd workspace
git clone https://github.com/microsoft/SparTA && cd SparTA && git checkout pit_artifact && git lfs pull
bash script/init_env_v100.sh
bash script/run_v100.sh

docker run -it --gpus all --shm-size=32gb --rm -d --name pit_artifact -h docker zhengningxin/pit_artifact:a100
docker exec -it pit_artifact bash
mkdir -p workspace && cd workspace
git clone https://github.com/microsoft/SparTA && cd SparTA && git checkout pit_artifact
bash script/init_env_a100.sh
bash script/run_a100.sh

You can get the results in the results directory.

You can find the video and the generated figures on OneDrive.

You can find the video and the generated figures on OneDrive.