Skip to content

Latest commit

 

History

History
115 lines (99 loc) · 4.14 KB

GoogleCloudSetup.md

File metadata and controls

115 lines (99 loc) · 4.14 KB
Raw

Google Cloud setup

Here is a small tutorial on how to setup a remote dev machine on Google Cloud. If you haven't used GCE before, start here and complete some turial.

Make ssh key for your dev instances

https://cloud.google.com/compute/docs/connect/create-ssh-keys The USERNAME below is your

ssh-keygen -t rsa -f ~/.ssh/gce_key -C <USERNAME> -b 2048

To see the generated public key that you're going to add to GCE:

cat ~/.ssh/gce_key.pub

(optional) Add this ssh key globally

Create the cheapest test CPU VM to practice from scratch

Go to Google Compute Engine (GCE), turn on GCE API. You should see creating Create a VM with the following options:

  • default region is the cheapest
  • E2 is a good cheapest option, but you need at least 16GB of RAM if you'll use pip
  • "Availability policies", choose "Spot" to save money
  • Check "Enable display service"
  • Boot disk, "Change", Choose Debian Deep Learning with CUDA 11 for GPU support (or Ubuntu with at least 20gb disk size for CPU-only)
  • Firewall, check "Allow HTTP/HTTPS traffic"
  • Advanced options, Disks, Add new disk, pick "Standard" (cheapest)
  • Advanced options, Security, Manage Access, Add manually generated SSH keys, add the content of ~/.ssh/gce_key.pub

Now create the instance, you should see a green checkmark in "Status" column.

Add this to your ~/.ssh/config. Your USERNAME is what's before your @gmail.com. EXTERNAL_IP is what you see in "External IP" column of your running instance

Host gce
  HostName <EXTERNAL_IP>
  User <USERNAME>
  IdentityFile ~/.ssh/gce_key

You should be able to login ssh gce from your laptop. Call sudo passwd to change the password.

Format your empty disk or attach existing disk

Follow (this tutorial)[https://cloud.google.com/compute/docs/disks/format-mount-disk-linux]. In short, for empty disk:

sudo lsblk
# you should see your large disk size under sdb. Now create the filesystem
sudo mkfs.ext4 -m 0 -E lazy_itable_init=0,lazy_journal_init=0,discard /dev/sdb
# mount the disk
sudo mkdir -p /mnt/disks/disk-1
sudo mount -o discard,defaults /dev/sdb /mnt/disks/disk-1
# mount on boot
# copy UUID returned by the following command:
sudo blkid /dev/sdb
# then
sudo vim /etc/fstab , Shift+G, o
# add this:
UUID="<UUID_FROM_ABOVE>" /mnt/disks/disk-1 ext4 discard,defaults 0 2

If you attached an existing disk and you used a debian deep learning image, the disk is by default mounted to /home/jupyter

Move home folder onto new large drive:

If you used a debian deep learning image, see below. For a new mounted image:

cd /mnt/disks/disk-1
mkdir -p home/<USERNAME> 
sudo rsync -avz --progress /home/<USERNAME>/ /mnt/disks/disk-1/home/<USERNAME>/
sudo vim /etc/passwd
# find your username entry and change /home/<USERNAME> to /mnt/disks/disk-1/home/<USERNAME>
sudo chown -R <USERNAME>:<USERNAME> /mnt/disks/disk-1/home/<USERNAME> 
sudo reboot

If you used a debian deep learning image, the disk is by default mounted to /home/jupyter

sudo vim /etc/passwd
# find your username entry and change /home/<USERNAME> to /home/jupyter/home/<USERNAME>
sudo chown -R <USERNAME>:<USERNAME> /home/jupyter/home/<USERNAME>
sudo reboot

Now when you ssh again into ~ and call pwd you should see your new home location.

Setup github keys

git config --global user.email "[email protected]"
git config --global user.name "Your Name"

Add a ~/.ssh/gce_key to your GitHub account. Copy the keys from your laptop to the dev VM: scp ~/.ssh/gce_key* gce:~/.ssh/ Add the following at the end of ~/.bashrc

# If not running interactively, return early
# https://stackoverflow.com/questions/64790393/indicated-packet-length-too-large-error-when-using-remote-interpreter-in-pycha
[[ $- == *i* ]] || return
eval $(ssh-agent)
ssh-add ~/.ssh/gce_key

Now you should be able to clone your repo.

If you're running GCE deeplearning image, disable jupyter service:

sudo systemctl stop jupyter.service
sudo rm /etc/systemd/system/multi-user.target.wants/jupyter.service