Js2Hou / Pytorch-DDP-Analysis

Pytorch DDP Analysis

Introduction

This project studies how to distributed train model in Pytorch and analyzes its efficiency. We run the same code on a GPU, two GPUs and four GPUs (two GPUs per node), and record their runtime.

Experiments show that training model on 2 GPUs can save 1 of time than that on single GPU, but when done on 4 GPUs (2 GPUs per node) it will cost more time. We think it is possible that our dataset (cifar100) is too small. Nonetheless, it suggests that we should run code on a single node instead of multiple nodes to achieve better performance.

In addition, We provide a template of Pytorch distributed training.

Structure

Tamplate/
|-- data/
|
|-- models/
|   |-- __init__.py
|   |-- models.py
|
|-- scripts/
|   |-- train_single_gpu.sh  # script for training on single gpu
|   |-- train_single_node.sh  # script for training on single node with multi gpus
|   |-- train_multi_nodes.sh  # script for training on multi nodes
|
|-- dataset.py
|-- main.py
|-- metrics.py
|-- engine.py
|-- utils.py
|-- requirements.txt
|-- README

• dataset.py: loading data and doing data augmentatioin
• models/models.py: implement your model
• scripts: scripts for starting training
  • train_single_gpu.sh: training model on single gpu
  • train_single_node.sh: trainging model on single node with multi gpus
  • train_multi_nodes.sh: training on multi nodes. Execute this script on each node to start distrubuted training.

Experiment on DDP

Setting

• device: TITAN RTX
• model: resnet18
• dataset: cifar100
• epochs: 10
• batch size: 128
• dataset augmentation:

  transform = create_transform(
      input_size=input_size,
      is_training=True,
      color_jitter=0.4,
      auto_augment='rand-m9-mstd0.5-inc1',
      interpolation='bicubic',
      re_prob=0.25,
      re_mode='pixel',
      re_count=1,
  )

Results

• single GPU: 555.0834 s
• 2 GPUs on single node (launched by torch.distributed.launch, will be deprecated) : 301.9986 s
• 2 GPUs on single node (launched by torchrun) : 324.2550 s
• 4 GPUs on two nodes (launched by torchrun) : 549.2544 s

Usage

First, clone the repository locally:

git clone https://github.com/js2hou/Pytorch-DDP-Analysis.git

Then, install requirements:

pip install -r requirements.txt

Loading your dataset in dataset.py and implement your models in models/models.py. Remember to modify the code for calling the model in main.py. Last,

• run ./script/train_single_gpu.sh for training on single gpu
• run ./script/train_single_node.sh for training on multi gpus (recommended)
• run ./script/train_multi_nodes.sh on each node for training on multi nodes

License

This repository is released under the Apache 2.0 license as found in the LICENSE file.