Using a GCP VM like a ‘local’ Deep Learning machine

Everything from CLI

This is a short-but-complete guide to setting up not just Jupyter, but a whole Deep Learning environment on a Google Cloud Virtual Machine.

Outline

These are the steps that this involves, and leaves the final machine pretty quick to start/stop:

  • Create a preemptible GPU-enabled GCP machine
  • Set up ssh so that it can be used easily
    • Include ‘tunnels’ that enable jupyter or tensorboard to be accessed securely through ssh
  • Install software on the GCP machine:
    • the Nvidia drivers with cuda
    • frameworks like tensorflow and pytorch that match the cuda version
    • other ‘essentials’ like jupyter
  • Run things like jupyter
    • As a test
    • In a screen for longer-lived sessions
  • Mount the GCP machine’s files as if they were local
    • This allows us to use an IDE seamlessly

All-in-all, the below enables us to use GCP as a replacement for a local Deep Learning machine - and this may be:

  • a better economic choice (given current GPU pricing, plus the cost of electricity where you live)
  • more convenient/reliable than Colab

… for more serious Deep Learning development.

Create a suitable Google Cloud VM

( To see how to do this for a machine that will also auto-launch services and servers, see my Building a reusable Deep Learning VM on Google Cloud post. The two posts are ‘compatible’, and can be combined into one production + experimentation machine. ).

export PROJECT_NAME="my-special-project"
gcloud config set project ${PROJECT_NAME}
export INSTANCE_NAME="deep-learning-vm1"

And then run this to actually create the instance:

export ZONE="asia-southeast1-b"  # This is good for where I am located (Singapore)
export GCP_USER=`whoami`

gcloud beta compute --project=${PROJECT_NAME} instances create ${INSTANCE_NAME} \
  --zone=${ZONE} \
  --machine-type=n1-standard-8 \
  --subnet=default --network-tier=PREMIUM \
  --no-restart-on-failure --maintenance-policy=TERMINATE \
  --preemptible \
  --accelerator=type=nvidia-tesla-t4,count=1 \
  --image=ubuntu-2004-focal-v20210623 --image-project=ubuntu-os-cloud \
  --boot-disk-size=50GB --boot-disk-type=pd-balanced --boot-disk-device-name=${INSTANCE_NAME} \
  --no-shielded-secure-boot --shielded-vtpm --shielded-integrity-monitoring --reservation-affinity=any \
  --metadata=startup-script="#! /bin/bash
      su --login --command=/home/${GCP_USER}/3-on-start.bash ${GCP_USER}"

NB: The ./3-on-start.bash bit adds some extensibility for auto-running processes later. There’s no need to have a script actually at that location if it’s not required.

FWIW, from the GCP VM pricing estimate (though the numbers below are only when it’s switched ON, apart from the persistent disk) is :

$154.71 monthly estimate - That's about $0.212 hourly
  8 vCPUs + 30 GB memory          $68.91/month
  1 NVIDIA Tesla T4 GPU           $80.30/month
  50 GB balanced persistent disk   $5.50/month

The --accelerator=type=nvidia-tesla-t4,count=1 choice is clearly one that depends on your requirements - at the T4 level, it just makes the instance like a more reliable colab, but with the option for tensorboard, persistence and local disk mounting (to give some key advantages).

The --machine-type=N1-standard-8 choice may be a bit of an overkill for a Deep Learning instance, though (compared to scaling up the GPU side) it’s relatively low incremental cost for the additional room / cores provided.

Important Commands

Important command pallete (for quick reference).
Execute these on your local machine, once the gcloud config set project ${PROJECT_NAME} and INSTANCE_NAME things are set as above:

# Start the instance
gcloud compute instances start ${INSTANCE_NAME}

# stop the instance :: REMEMBER TO DO THIS WHEN YOU'RE DONE FOR THE DAY...
gcloud compute instances stop ${INSTANCE_NAME}

Configure ssh so we can make some tunnels…

In order to get sshfs running, we need to get ssh running more seamlessly. On the local machine, use the following (this has to be done while the VM is running):

gcloud compute config-ssh

# You should now be able to use ssh/scp with your instances.
# For example, try running:
# $ ssh deep-learning-vm1.asia-southeast1-b.my-special-project

This adds a Host deep-learning-vm1.asia-southeast1-b.my-special-project section to ~/.ssh/config.

Let’s store the required address into a variable:

# Sadly, these don't provide suitably formatted data...
#gcloud compute instances list --filter="name=${INSTANCE_NAME}"
#gcloud compute instances describe ${INSTANCE_NAME}

GCP_ADDR=`grep "Host ${INSTANCE_NAME}" ~/.ssh/config | tail --bytes=+6`
echo ${GCP_ADDR}
# deep-learning-vm1.asia-southeast1-b.my-special-project

Not only does this allow ‘real’ ssh access (rather than gcloud compute ssh ${INSTANCE_NAME} which is a wrapper), but by enabling direct system usage, we get rsync and sshfs compatibility for free.

Set up an ssh tunnel on the local machine

Having captured the right GCP hostname, we can now set up an ssh session, including a number of ‘tunnels’ that allow us to use the machine as if it were local to us.

On the local machine, use the following to tunnel ports (like the one we’re using for jupyter) through ssh:

## No need for these : We've got ssh set up properly...
#gcloud compute ssh  -- -N -p 8585 -D localhost:8585
#gcloud compute ssh  -- -N -L 8585:localhost:8585 -L 8586:localhost:8586 # ... etc

ssh ${GCP_ADDR} -L 8585:localhost:8585 -L 8586:localhost:8586 # ... etc

Included above is the 8585 port for jupyter, and an extra 8586 one for tensorboard (by way of example).

Install Nvidia drivers and cuda

Use ssh ${GCP_ADDR} to get into the GCP machine, and run :

sudo apt install linux-headers-$(uname -r)
#linux-headers-5.8.0-1035-gcp is already the newest version (5.8.0-1035.37~20.04.1).

curl -O https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/cuda-ubuntu2004.pin
sudo mv cuda-ubuntu2004.pin /etc/apt/preferences.d/cuda-repository-pin-600
sudo apt-key adv --fetch-keys https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/7fa2af80.pub
# gpg: key F60F4B3D7FA2AF80: public key "cudatools <cudatools@nvidia.com>" imported
sudo add-apt-repository "deb https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/ /"
sudo apt update
sudo apt -y install cuda
# This step takes ~16mins
#   and installs a whole heap of wierd stuff...  (fonts, etc)

# The output states that machine needs a reboot... (though it seems to be properly installed already)

sudo nvidia-smi
# | NVIDIA-SMI 470.42.01    Driver Version: 470.42.01    CUDA Version: 11.4     |
# |   0  Tesla T4            Off  | 00000000:00:04.0 Off |                    0 |
# | N/A   76C    P0    34W /  70W |      0MiB / 15109MiB |      0%      Default |

NB: Since the hard disk we’ve chosen is Persistent, all of this installation only needs to be done once.

Create a local venv for python

This is good python hygiene :

sudo apt install -y python3.8-venv
python3.8 -m venv env38 

. env38/bin/activate

pip install --upgrade pip

Install Deep Learning frameworks

Once you have a venv installed (assumed to be named as above):

. env38/bin/activate

pip install tensorflow tensorboard

# Agrees with the CUDA version installed above:
pip install torch==1.9.0+cu111 torchvision==0.10.0+cu111 torchaudio==0.9.0 -f https://download.pytorch.org/whl/torch_stable.html

# An example of a pytorch module with specific CUDA code included:
#pip install pytorch3d -f https://dl.fbaipublicfiles.com/pytorch3d/packaging/wheels/py38_cu111_pyt190/download.html

Install jupyter, etc

Once you have a venv installed (assumed to be named as above):

. env38/bin/activate

pip install jupyter
jupyter notebook --generate-config
#  `/home/USERNAME/.jupyter/jupyter_notebook_config.py`

Update the jupyter defaults

Add default configuration to ~/.jupyter/jupyter_notebook_config.py:

USER=`whoami` && echo "
c.NotebookApp.ip = 'localhost'
c.NotebookApp.open_browser = False
c.NotebookApp.port = 8585
c.NotebookApp.notebook_dir = '.'
c.NotebookApp.token=''
c.NotebookApp.password=''
   " >> /home/${USER}/.jupyter/jupyter_notebook_config.py

NB: This is very unsafe if you open port 8585 to the internet (here it is set to bind to localhost, rather than ‘*’, as a precaution)! . But we’re not going to do that: We’re only going to access it via an ssh tunnel to the server’s port 8585. And that’s all secure within what is essentially our own little VPN.

Launch Jupyter on the server

Since we’ve set up the default notebook-dir and other command-line options in the configuration, jupyter should work from whereever you launch it:

. env38/bin/activate

jupyter notebook

Launch Jupyter in the browser

Now you can get to the running instance :

  • Browse to http://localhost:8585/
    • There’s no need for a token, or a key, since the ssh tunnel makes this entirely private
      • i.e. nothing has been exposed for public internet access

This was a proof-of-concept. We can deal with tensorboard (and any other services you want to run) in the same way.

Launch Multiple services in the backend

We can use screen to run several different services on the server via this one ssh session (you will probably want to update the paths).

The following can be run on the GCP server using the ssh session created above :

NL="$(printf \\r)"
ACTIVATE_ENV=". ~/env38/bin/activate"

# this changes the the right base directory
pushd ./

SCREEN=jupyter
screen -dmS $SCREEN
screen -S $SCREEN -p 0 -X stuff "${ACTIVATE_ENV}${NL}"
#screen -S $SCREEN -p 0 -X stuff "cd ../whereever${NL}"
screen -S $SCREEN -p 0 -X stuff "jupyter notebook${NL}"

SCREEN=tensorboard
screen -dmS $SCREEN
screen -S $SCREEN -p 0 -X stuff "${ACTIVATE_ENV}${NL}"
#screen -S $SCREEN -p 0 -X stuff "cd ../whereever${NL}"
screen -S $SCREEN -p 0 -X stuff "tensorboard serve --port=8586 --logdir=./lightning_logs${NL}"

popd 

You can also make the above into a script so that it’s easy to run…

This should enable:

  • http://localhost:8585 for jupyter
  • http://localhost:8586 for tensorboard

Use sshfs to remotely access the VM filesystem

… so that it appears to be local storage

First, ensure that you have the fuse filesystem sshfs installed on your local machine:

  • Fedora details
    • Fedora local installation requires dnf install fuse-sshfs

The following mounts the GCP home directory on gcp_base locally - of course the remote path (and name of the local mount point) are totally up to you:

mkdir -p gcp_base

sshfs ${GCP_ADDR}:. gcp_base -o reconnect

Once finished :

fusermount -u ./gcp_base

Note that if you have files open within the mount point, fusermount will refuse to unmount the directory. This can also be caused by anything that might watch within the directory too (e.g. a source control manager looking for changed files, or a web server auto-refresh). Just stop these things running locally, and fusermount will then dismount properly (you may have to find out which process has open files within the mountpoint using lsof ./gcp).

Normal Operation!

Open two tabs locally, and in each, run :

export PROJECT_NAME="my-special-project"
gcloud config set project ${PROJECT_NAME}
export INSTANCE_NAME="deep-learning-vm1"
export GCP_ADDR=`grep "Host ${INSTANCE_NAME}" ~/.ssh/config | tail --bytes=+6`
echo ${GCP_ADDR}  # Just to check that it's found

Tab 1

This tab will run the server side of the interface:

gcloud compute instances start ${INSTANCE_NAME}
gcloud compute config-ssh
ssh ${GCP_ADDR} -L 8585:localhost:8585 -L 8586:localhost:8586 # ... etc

# Inside the `ssh` session, run the 'screen script' above
#   And leave the ssh session running ...
#     (even if you don't need the terminal, you may want the tunnels open)

Tab 2

This tab will run the local side of the interface:

sshfs ${GCP_ADDR}:. gcp_base -o reconnect

Load editors, etc, to operate on these files ‘locally’.

Terminate the GCP VM when done…

# if needed:
fusermount -u ./gcp_base

gcloud compute instances stop ${INSTANCE_NAME}

End

All done!



Martin Andrews

{Finance, Software, AI} entrepreneur, living in Singapore with my family.



blog comments powered by Disqus