A first step example of creating an environment that allows for the implementation of deep learning in a GPU supported Ubuntu Packet server using Terraform.
This guide was created for users who may not have Linux machines with GPU supported drivers, or want to engage in machine learning research with the use of bare metal servers. We will be using a GPU accelerated Packet server to implement a machine learning environment, which will require a Packet account to be created, and the use of this server will have an hourly billing fee.
Please keep in mind that the device created will have Ubuntu 16.04 as the operating system and several packages and installations will be installed automatically. It will be helpful to have prior knowledge in Linux commands and the following installations, but you may install other applications to aid you in your machine learning research as desired:
- Anaconda
- Docker CE
- Nvidia Docker
- Cuda 10.1
- Tensorflow-GPU
- Jupyter
- Getting Started
- Clone The Repository
- Set Necessary Variables
- Initialize Terraform
- Deploying the Packet Server
- Enter Your Server
- Check The Script
- Create a Jupyter Notebook Without a Domain Name
- Create a Jupyter Notebook With a SSL-Enabled Domain Name
- Acknowledgments
Begin by installing Terraform on your machine. (A Mac was used in the making of this guide, but you may use any machine as long as Terraform supports installation for your operating system).
Please ensure that you download the file with the correct operating system and architecture of the machine you will be using Terraform on, or use a package manager (like Homebrew if you have a Mac or Linux machine, or Chocolatey if you have Windows) to automate the installation process.
Verify that your Terraform installation has been performed successfully and that your PATH has been set correctly by running the terraform command.
You should get the following output:
Usage: terraform [-version] [-help] <command> [args]
(...)
For further help on installing Terraform on your operating system, please visit Terraform's Getting Started guide.
Clone this repository to your machine to have all the necessary files needed to easily deploy a server with GPU support.
This can be done by entering the following:
$ git clone https://github.com/tchan2/packet-gpu
Now you should have all the necessary files for deploying a Packet server with GPU acceleration.
After cloning the repository, enter the packet-gpu directory and go into the variables.tf file.
You will see that variables auth_token and project_id have been left blank.
Please enter your API key in the quotes indicated in the default field. If you do not, Terraform will give an error stating that you do not have the correct authentication token needed to create this server.
variable "auth_token" {
(...)
default = "<YOUR_API_TOKEN_HERE>"
}
If you have already created a project and would like to use it, please skip to the section below.
As a default, it has been assumed that a project has not been created, so the following block of code in packet-gpu.tf creates the packet_project resource:
resource "packet_project" "<your_project_name>" {
name = "Your Project Name"
}
Your project name has been defaulted to tf_project and the description name has been defaulted to New Project, but you may customize it accordingly.
If you have changed the project name, please also adjust the following code to allow Terraform to use the created project's project_id. Otherwise, leave it alone:
module "packet-gpu" {
(...)
project_id = "${packet_project.<your_project_name>.id}"
}
Please visit the section on packet_project on the Terraform Packet Provider page for more information of other fields you can add to customize your Packet project further.
You may now skip to the Initialize Terraform section of this page.
If you have already created a project in your Packet account and would like to use it, first enter variables.tf.
Enter your project ID in default field in your variable project_id. Your project ID can be found by going into your preferred project, and taking the portion of text after https://app.packet.net/projects/<YOUR_PROJECT_ID> in your browser bar.
variable "project_id" {
(...)
default = "<YOUR_PROJECT_ID_HERE>"
}
Now, switch to file packet-gpu.tf.
Comment out the block of code that creates a packet_project resource:
# resource "packet_project" "your_project_name" {
# name = "Your Project Name"
# }
WARNING: Please make sure you comment this resource out, or else it will create a new project even if you will not be using it, which will incur extra fees.
Then, comment out project_id = "${packet_project.tf_project.id}", and uncomment project_id = "${var.project_id}" to use the project ID you have set in your variables.tf file:
module "packet-gpu" {
(...)
project_id = "${var.project_id}"
# project_id = "${packet_project.tf_project.id}"
(...)
After you are done, please continue to the section below.
In order to deploy the Packet server, we must initialize Terraform for deployment. We do this by entering the following:
$ terraform init
Now, we can view the plan for deployment by running:
$ terraform plan
This will show what resources will be added, changed, or destroyed. Please look this plan over to ensure that it is to your preferences.
If everything looks good, we can start deploying our Packet server. Run the following:
$ terraform apply
Check the plan again, and answer yes if everything is to your liking.
If all the steps have been completed successfully, a Packet server will have been deployed!
If you would like to create new devices and reuse your projects without having to continuously create and/or delete them, you can use modules!
Click to expand
To add modules, please create a new module in packet-gpu.tf like so:
module "module-name" {
source = "./modules/packet-gpu"
project_id = "${var.project_id}"
hostname = "${var.hostname}"
plan = "${var.plan}"
facility = "${var.facility}"
os = "${var.os}"
billing_cycle = "${var.billing_cycle}"
}
As an example, I have used
module-nameas the new module name. You can set it to your liking, but please remember that if you create a new module, you must set it to a different name than the ones you have already created.
Here, you can choose to manually set your variables here instead of using the default, and your set variables will override the default ones.
This will create a new device completely separate from the one(s) you already have, and make it possible for you to create a new device and reuse the project you already have or already created.
If you would like to apply changes to your module(s) without affecting the one(s) already created, run terraform apply with your target module like so:
$ terraform apply -target=module.module-name
This will update only the targetted module, and not touch the rest.
For more information, please visit Terraform's help documents on modules.
Open up a terminal, and SSH into your device. You may do this by entering:
ssh -L 8888:localhost:8888 user@PUBLIC_IPv4_ADDRESS
The IP address of your server can be found by logging into your Packet account and checking there, or by searching the terraform.tfstate file that has just been created after running terraform apply.
Once there, please create a passphrase to secure your server.
In order to use the packages and installations in this server, it will be required that the script running in the server's user data has been completed.
To check the progress of this script, please run:
$ cat /var/log/cloud-init-output.log
You may run this commmand as many times as you would like as the log will update as the script runs in the background.
Once you see
(...)
Completed. Script finished.
at the bottom of the cloud-init-output log, the script has been completed.
To check your installations, install necessary packages, and create an insecure Jupyter notebook without a domain name in a Conda environment, run these commands:
$ wget -O http-jupyter.sh https://raw.githubusercontent.com/tchan2/packet-gpu/master/scripts/without-domain.sh
$ chmod +x http-jupyter.sh
$ ./http-jupyter.sh
If you have left the script as the defaults, this will test your installations and make sure that they have installed correctly, create and activate an GPU-supported Anaconda environment called jupyter_env, and create a Jupyter notebook!
Once this runs, you will see a link like the following that you can use to access your Jupyter notebook:
[C 16:21:28.754 NotebookApp]
(...)
Or copy and paste one of these URLs:
http://localhost:8888/?token=<YOUR-PERSONALIZED-TOKEN>
Just copy and paste this link into your preferred browser, and begin coding!
To check your installations, install necessary packages, and run a SSL-enabled Jupyter Notebook with a domain name, please continue below.
Note: The following guide will focus on using No-IP to create a domain name.
If you do not wish to use No-IP and want to use another site or an already created domain name, please make sure your IP address has been set correctly, and then continue to section Run Script below.
No-IP Guide
Please go on no-ip.com to create a free domain name!
Sign up using your preferred email, and fill out any necessary fields.
In order to create a hostname, please refer to the sidebar to the left.
Click on My Services, and then click on DNS Records. You will see a button that says Add a Hostname.
(Note: If you would like to pay for a domain with no-ip.com, you may click on Domain Registration.)
Please set your hostname, and your preferred domain.
Then, for the Hostname Type, please set it to DNS Hostname (A).
After you have created your domain name, please insert your public IPv4 address into the IP Address field.
Click on Add Hostname and you have successfully created your domain name!
Please allow at least 5-10 minutes for this change to take place, and continue to section Run Script below.
Run the following to get my script onto your Packet server to be able to enable SSL wrapping for your domain, and to add that domain name to your Jupyter notebook.
$ wget -O https-jupyter.sh https://raw.githubusercontent.com/tchan2/packet-gpu/master/scripts/with-domain.sh
$ chmod +x https-jupyter.sh
As we will be using Certbot to obtain a SSL certificate, we will need to provide the domain name and email in order to obtain it.
Please pass your domain name and then your email as the arguments to successfully run your script:
$ ./https-jupyter.sh your.domain.name your@email.com
Now, your Jupyter notebook should have started running! You should see this:
(...)
[I 21:16:06.647 NotebookApp] The Jupyter Notebook is running at:
[I 21:16:06.647 NotebookApp] https://your.domain.name/
(...)
Please follow the link provided, enter your password, and start coding!
Special thanks to My, Zak and, Joseph for all the help on this product!