cloud

Cloud Instance Setup Instructions

Google has made $50.00 USD Compute Engine credits available to all of us. There is some extra budget, but do not count on it. Please remember to shut down your instances when you are done working.

Accounts, Projects and Billing

• Make sure the only Google account you are currently logged in as is your @berkeley.edu account. (Not the @ischool.berkeley.edu one!)
• Visit cloud.google.com
• Click "Console" in the upper right corner.

• From the "Project" drop down menu, pick "Create project" and fill in the resulting form.

• Create a new browser tab.
• In the new tab, visit https://google.secure.force.com/GCPEDU/?cid=x5YDDxzK2St64XRhk8aG7eA2p6VTRGjX%2F9joDW5FZOjTkB1FMqJkrVMhUwXVlG%2FE/ request a coupon, and follow the instructions.
• Switch back to the original tab.
• From the "hamburger" menu in the upper left corner, select Billing.

• Verify the project is associated with your Compute Engine credits. You should see the course name in the Promotion ID field.

Start the Compute Engine instance

• From the hamburger menu again, select "Compute Engine"

• (You may see a dialog box for a minute or two as Compute Engine initializes for your project.)

• Click "Create Instance"

• Fill in the instance creation form as shown below. Don't worry about the cost; the rate shown assumes always-on, so you shouldn't come close to the limit if you shut down when not in use.

  • Name: w266
  • Zone: (pick one close to you)
  • Machine Type: 1 vCPUs, 3.75 GB
  • Boot Disk: Ubuntu 16.04 LTS (modify boot disk size to 100 GB)
  • (Optionally) allow HTTP(s) traffic

• Before clicking "Create", be absolutely sure that you have increased the disk space to at least 100 GB and have selected Ubuntu as your OS (the setup script that you'll run in a later step may not work otherwise).
• Click "Create" at the bottom of the form. You may need to wait a minute or two for your instance to boot.

The settings above (1 vCPU, 3.75 GB memory) should be fine for most parts of most assignments. For some assignments (particularly, training neural networks) and for your final project, you may need more juice. Fortunately, it's very easy to resize an instance. Go to the VM Instances page, shut down your instance, and select "Edit" to select the number of CPUs and amount of RAM you need. When you boot up the instance and log in, your configuration will be exactly the same, just with more power.

Connecting to your instance

Google Cloud provides a handy in-browser SSH client. Click "SSH" on the VM Instances page, and it'll automatically handle setting up SSH keys. If you want to use your own terminal, see below.

Setting up your instance

In your terminal on the remote instance, type:

git clone https://github.com/datasci-w266/2017-fall-main.git ~/w266
./w266/assignment/a0/cloud/setup.sh
source ~/.bashrc

setup.sh installs Anaconda and TensorFlow. There are a handful of times you'll need to accept the license agreement or confirm an install path (just accept the defaults, and allow it to modify your .bashrc file).

Note: if you're not using a Cloud instance, but are running Ubuntu on your laptop or a VM, you can also use setup.sh or cherry-pick commands from it to install the necessary packages.

Run a notebook and connect to it from your local machine

In order to use Jupyter notebooks, we need to set up an SSH tunnel. This will allow the browser on your local machine to connect directly to the Jupyter server and Python backend running on your cloud instance. All the below commands should be run on your laptop, unless specified otherwise.

• First, install the Google Cloud SDK for your operating system. Follow the instructions at: https://cloud.google.com/sdk/

• If it gives you the option to run an ./install.sh script or similar, do so.

• Run gcloud init. When prompted, select the W266 project you created. When it prompts you to set up Google Compute Engine, say Y, and enter the zone of your instance as the default (this would be us-central1-b in our example).

• Run gcloud compute ssh <instance-name> to SSH to your instance. If it's your first time doing so, you'll be prompted to set up an SSH key. Enter a passphrase you won't forget!

  • Note: If your local machine username is different from the VM instance username, you will want to instead run gcloud compute ssh <vm-username>@<instance-name> to ensure you login as the user where w266/ is located.

• If you're able to log in successfully, hit Ctrl+D or type logout to exit.

• Run the following command from your laptop, which will SSH to your server and tunnel ports 8888 and 6006:

gcloud compute ssh --ssh-flag="-L 8888:127.0.0.1:8888" --ssh-flag="-L 6006:127.0.0.1:6006" <vm-username>@<instance-name>

Note: port forwarding may not work properly using the web SSH client on the Google Cloud Console. We recommend installing the Cloud SDK and using it from the command line.

That's a lot to type every time, so you might want to make a shell alias for it. Port 8888 is the default for Jupyter, and we'll be using port 6006 to access some monitoring interfaces for TensorFlow.

• If you're successfully logged in, start a Jupyter notebook server on your remote instance with jupyter notebook. You should see a message like "The Jupyter Notebook is running at: http://localhost:8888":

• Open a browser on your laptop and go to http://localhost:8888. You should see the Jupyter notebook home:

You might want to use tmux (or screen) to keep the notebook server running, even if you get disconnected or want to leave something running while you close your terminal. On your instance, run:

tmux new-session -s notebook
# should see a new terminal prompt
jupyter notebook

• You can hit Ctrl+b then d to detach from the session, and type tmux attach -t notebook to re-attach.
• More on tmux: A tmux Primer

Shutting down your instance

Don't forget to shut down your instance, or you will quickly use up all your free credits.

• The easiest and safest way to shut down your instance is to type the following in the ssh session:

sudo shutdown now

Alternatively, you can force your instance down by:

• Return to the "VM instances" page
• Tick the box next to your instance
• Click "STOP" in the header

Restarting your instance

Restarting your instance is as simple as:

• Return to the "VM instances" page
• Tick the box next to your instance
• Click "START" in the header

Using your own SSH client (advanced)

If you don't want to use gcloud compute ssh, you can do the following to access your instance like any other server:

• Once it boots, click on the three dots next to "ssh" and pick "Use another SSH client".
• Follow the instructions on the linked page to create ssh keys and install them on your instance via the "Metadata" section.

• Finish following the instructions to ssh to your instance. Note for this step, you'll need your machine's External IP Address which can be found on the main VM Instances screen.

Note: you won't have to set up ssh keys every time you connect (you'll simply reuse the ones you just set up).

Now you can SSH normally. To SSH with port forwarding for Jupyter and TensorBoard, use:

ssh -L 8888:127.0.0.1:8888 -L 6006:127.0.0.1:6006 <instance_external_ip>