nooblyh/tgs

Artifacts for our NSDI'23 paper TGS

TGS

1. Introduction

This repository contains one version of the source code for our NSDI'23 paper "Transparent GPU Sharing in Container Clouds for Deep Learning Workloads" [Paper]

2. Environment requirement

Please see requirement.txt and paper for more details.

3. Prerequisites

Run the following commands:

sudo apt install patchelf wget unzip make
pip3 install -r requirement.txt
docker pull bingyangwu2000/tf_torch
docker pull bingyangwu2000/pytorch_with_unified_memory
docker pull bingyangwu2000/antman
docker pull bingyangwu2000/espnet2

4. Build

Run the following commands

git clone --recursive https://github.com/BingyangWu/TGS.git
cd TGS
make rpc
./download.sh
cd hijack
./build.sh

4. Run example

TGS:

./scripts/test_tgs.sh

Co-execution:

./scripts/test_co_ex.sh

MPS:

./scripts/test_mps.sh

AntMan:

./script/test_antman.sh

MIG:

./script/test_mig.sh

When run experiments in Figure 5, please use image bingyangwu2000/pytorch_with_unified_memory for Co-execution, MPS and MIG.

When run experiments in Figure 9(a), please use image bingyangwu2000/espnet2.

When run experiments in Figure 12, please use image goldensea/megatron:v2.