This repo is for the paper Speeding up Deep Learning with Transient Servers presented at ICAC 2019. The paper was also covered by a Google for Education case study blog.
The paper explores the opportunity of conducting distributed training of deep neural networks with transient cloud resources, especially the GPU servers. The transient resources are cheaper than on-demand ones, but with the caveat of being revoked by the cloud provider at any given time. By using transient cloud servers, we achieved the potential of up to 62.9% monetary savings and consequently 7.7X speed up due to the spare budget to deploy more servers in training. We also identified several opportunities for both cloud infrastructure and deep learning frameworks to provide better support for transient distributed training.
The repo contains code, found in the code folder, to reproduce the experiments mentioned in the paper. The code will request cloud resources that the user specifies, propagate training scripts, and set up distributed training jobs on the servers. The repo also provides the experiment data shown in the paper, located in the data folder. For details on the data, please see the readme inside the data folder.
-
Our data driven approach shows that by using transient cloud servers, we can achieve the potential of up to 62.9% monetary savings and 7.7X speed up compared to a single server baseline.
-
By launching large scale transient cloud servers, we are able to gather data on the impact transient revocation has on training performance, in terms of training time, cost, and converged accuracy and trained models.
-
We demonstrate several benefits and limitations of using heterogeneous servers in distributed training. In particular, our findings suggest a number of plausible transient-aware designs for deep learning frameworks, including the ability to train with dynamic cluster sizes, to better exploit these cheap transient servers.
Fig 1. Training performance and cost of using cluster with 4 transient K80 GPU servers
Fig 2. Training performance and cost of using cluster with 8 transient K80 GPU servers
Fig 3. Training cost of using different configurations of heterogeneous clusters with 4 GPU servers
The code used custom cloud images for both GPU and CPU servers, and the images are currently not public. Thus in order to run it, you need to create two custom images first.
We ran the code on cloud servers with Ubuntu 18.04 LTS, 4-8 vCPU cores and 24-51 GB memory, with 100 GB disk space. Ubuntu 16.04 LTS might work, but with some unexpected behavior.
First create a VM and ssh into it.
Then update apt-get and install the dependencies.
sudo apt-get update
sudo apt-get install -y openjdk-8-jdk git python-dev python3-dev python-numpy python3-numpy build-essential python-pip python3-pip python-virtualenv swig python-wheel libcurl3-dev curl g++ freeglut3-dev libx11-dev libxmu-dev libxi-dev libglu1-mesa libglu1-mesa-dev parallelInstall nvidia drivers; the code is based on CUDA 9.0. Notice: all the CUDA and CUDNN related dependencies are not required for the CPU image.
### Install NVIDIA driver
sudo apt install nvidia-384 nvidia-384-dev
### CUDA 9.0 requires gcc 6.0
sudo apt install gcc-6
sudo apt install g++-6
### Get CUDA 9.0 files and install
wget https://developer.nvidia.com/compute/cuda/9.0/Prod/local_installers/cuda_9.0.176_384.81_linux-run
chmod +x cuda_9.0.176_384.81_linux.run
sudo ./cuda_9.0.176_384.81_linux.run --overrideAfter rebooting the VM, check if CUDA is installed properly.
sudo reboot
nvidia-smiAn operational GPU would return something like:
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 396.26 Driver Version: 396.26 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
|===============================+======================+======================|
| 0 Tesla K80 Off | 00000000:00:04.0 Off | 0 |
| N/A 35C P8 27W / 149W | 15MiB / 11441MiB | 0% Default |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: GPU Memory |
| GPU PID Type Process name Usage |
|=============================================================================|
| 0 1658 G /usr/lib/xorg/Xorg 14MiB |
+-----------------------------------------------------------------------------+
Install CUDNN 7.5; you need to go to the Nvidia website and register, then download the tar file and install it.
Edit cuda path to ~/.bashrc and reload it.
echo 'export PATH=/usr/local/cuda-9.0/bin:$PATH' >> ~/.bashrc
echo 'export LD_LIBRARY_PATH=/usr/local/cuda-9.0/lib64:$LD_LIBRARY_PATH' >> ~/.bashrc
source ~/.bashrcDon't forget to move CUDNN to the CUDA folder.
sudo cp -P cuda/include/cudnn.h /usr/local/cuda-9.0/include
sudo cp -P cuda/lib64/libcudnn* /usr/local/cuda-9.0/lib64/
sudo chmod a+r /usr/local/cuda-9.0/lib64/libcudnn*
The last step would be to install TensorFlow 1.10 and modified Tensor2Tensor. Tensor2Tensor can be found in the code folder.
sudo pip install tensorflow-gpu==1.10
## for cpu servers install tensorflow==1.10 instead
pip install -e ~/code/tensor2tensorAfter the dependency installation, make two images, one for workers and one for parameter servers. Example command as below (to create the instance-gpu and instance-cpu images that we used in the code):
gcloud compute instances set-disk-auto-delete instance-gpu \
--disk instance-gpu --no-auto-delete
gcloud compute instances set-disk-auto-delete instance-cpu \
--disk instance-cpu --no-auto-delete
gcloud compute images create gpu-ubuntu18 \
--source-disk instance-gpu
gcloud compute images create cpu-ubuntu18 \
--source-disk instance-cpuThe code supports training models implemented in the Tensor2Tensor library. For the paper we mainly used ResNet models. The code currently supports Google Compute Engine.
To run the code, simply input the following command. It will set up a cluster with 1 parameter server and 4 workers equipped with K80 GPU, and train the CIFAR-10 dataset on ResNet-32 for 64k steps. The trained model will be generated in the specified cloud bucket.
python main.py --proj-name=YOUR_PROJ_NAME --cred-path=YOUR_GCE_CREDENTIAL_PATH --job-name=res32 --num-ps=1 --ps-core-num=4 --num-worker=4 --num-shard=1 --bucket-dir=gs://YOUR_BUCKET/ --model=resnet --hparam-set=resnet_cifar_32 --problem=image_cifar10 --train-steps=64000 --ckpt-frequency=100000 --automation-test=0 --setSlot=1 --maxWorker=8 --zone=us-west1-b --gpu=k80 --hetero=0Explanations of the other parameters can be found below; they are experimental and not the core focus of the paper:
-
ps-core-num: determines the number of vCPU cores for parameter servers.
-
num-shard: how many shards to partition the parameter set.
-
ckpt-frequency: how frequent to checkpoint during training.
-
automation-test: only used in combination with a monitor, currently not supported.
-
setSlot: part of test for dynamic learning rate.
-
maxWorker: part of test for dynamic learning rate.
Alternatively, if you want to test out heterogeneous cluster config, for example 4 workers and 1 parameter: parameter server in us-west1-b, 2 K80 GPU servers in us-west1-b, 1 P100 server in us-central1-a, and 1 V100 server in us-east1-a, you can use the following command:
python main.py --proj-name=YOUR_PROJ_NAME --cred-path=YOUR_GCE_CREDENTIAL_PATH --job-name=res32 --num-ps=1 --ps-core-num=4 --num-worker=4 --num-shard=1 --bucket-dir=gs://YOUR_BUCKET/ --model=resnet --hparam-set=resnet_cifar_32 --problem=image_cifar10 --train-steps=64000 --ckpt-frequency=100000 --automation-test=0 --setSlot=1 --maxWorker=8 --hetero=1 --gpu_array=k80 k80 p100 v100 --zone_array=us-west1-b us-west1-b us-west1-b us-central1-a us-east1-bIf you would like to cite the paper, please cite it as:
@inproceedings{2019icac:speedup,
author = {Li, Shijian and Walls, Robert J. and Xu, Lijie and Guo, Tian},
title = {"Speeding up Deep Learning with Transient
Servers"},
booktitle = {Proceedings of the 16th IEEE International Conference on Autonomic Computing (ICAC'19) },
year = {2019},
}This work is supported in part by National Science Foundation grants #1755659 and #1815619, Google Cloud Platform Research credits, the National Natural Science Foundation of China (61802377), and the Youth Innovation Promotion Association at CAS.
More project information can be found in our lab's project site.
-
Shijian Li [email protected]
-
Robert Walls [email protected]
-
Tian Guo [email protected]