I’ve been looking for an inexpensive way to run and experiment with napari on a remote GPU vm.
I’m still experimenting with different approaches to see what works best. I’ll add more posts as I learn more.
In this post, I’ll document the steps that I took to set up a remote DigitalOcean GPU droplet running Ubuntu with xfce4 for a lightweight GUI and tightvncserver using an SSH tunnel to connect to it. This setup allows me to run napari, interact with it, and run the built-in napari examples with reasonable performance.
⚠️ Use at your own risk. This is not a production setup and is intended for experimentation purposes only.
💰 Costs: I destroy the droplet after each session to avoid incurring costs when the droplet is not in use.
Components
- DigitalOcean GPU droplet: I chose the smallest GPU droplet available.
- MacBook Air M2 2022 with 16GB RAM as the local client machine
- TightVNC server: A lightweight VNC server for Linux
- TigerVNC viewer: A lightweight VNC viewer for macOS
- SSH tunnel: To securely connect to the VNC server over SSH
Set up the remote server
There were several processes that I used to set up the GPU droplet as the remote server.
1. Get GPU Access
I needed to request access from Digital Ocean to create the GPU droplet. I did this by submitting a support ticket to DigitalOcean. They approved my request, and I was able to create a GPU droplet.
2. Create the GPU Droplet
After receiving access, I created a new GPU droplet using the DigitalOcean control panel. I selected the smallest GPU droplet available ubuntu-gpu-4000adax1-20gb-tor1:
- Region: Toronto 1
- GPU: NVIDIA RTX4000 ADA
- Size: 1 GPU - 20 GB VRAM - 8 vCPU - 32 GB RAM - 500 GB Boot Disk
Insert image here
I also set up SSH access by using my SSH key during the droplet creation process. This allows me to connect to the droplet securely without needing a password.
Take note of the IP address of the droplet, as you’ll need it to connect later.
3. Connect to the Droplet
I connected to the droplet using SSH from my local machine:
ssh root@<droplet-ip-address>4. Setup the server user and firewall
# set up a new user
usermod -aG sudo sammy
# configure the firewall
ufw app list
ufw allow OpenSSH
ufw enable
ufw status
# copy SSH key to from root to the new user
rsync --archive --chown=sammy:sammy ~/.ssh /home/sammyTest the SSH connection with the new user:
ssh sammy@<droplet-ip-address>5. Install xfce4 and TightVNC server
From the user account, I installed and configured the lightweight desktop environment (xfce4) and the VNC server (tightvncserver):
# Install packages
sudo apt update
sudo apt install xfce4 xfce4-goodies
sudo apt install tightvncserver
# Start the VNC server to create the initial configuration files
vncserver
# Stop the vnc server before changing the configuration
vncserver -kill :1
# Back up the default xstartup file
mv ~/.vnc/xstartup ~/.vnc/xstartup.bak
# Create a new xstartup file
nano ~/.vnc/xstartupAdd the following content to the xstartup file:
#!/bin/bash
xrdb $HOME/.Xresources
startxfce4 &Save and exit the file.
# Make the xstartup file executable
chmod +x ~/.vnc/xstartup
# Start the VNC server
vncserver -localhost
startx xfce
startxSet up the client machine
1. Install the TigerVNC Viewer
On my MacBook Air, I installed the TigerVNC viewer to connect to the VNC server running on the remote GPU droplet.
Install the TigerVNC viewer by:
- Visit https://tigervnc.org/
- Visit https://github.com/TigerVNC/tigervnc/releases to find the binary for macOS
- At the time of writing, visit the following SourceForge URL: https://sourceforge.net/projects/tigervnc/files/stable/1.15.0/
- Download the
TigerVNC-1.15.0.dmgfile. - Open the downloaded file and drag the TigerVNC Viewer app to your Applications folder.
2. Set up the SSH Tunnel
To securely connect to the VNC server over SSH, I set up an SSH tunnel from my local machine to the remote GPU droplet. This allows me to forward the VNC traffic through the SSH connection.
Open a terminal on your MacBook Air and run the following command:
ssh -L 59000:localhost:5901 -C -N -l sammy <droplet-ip-address>3. Run the TigerVNC Viewer
To run the TigerVNC viewer, double-click the TigerVNC Viewer app in your Applications folder. It opens a window where you will enter the connection location and press “Connect”.
Insert image here
4. Use the TigerVNC terminal
If everything is set up correctly, you should see the remote desktop. You can now interact with the remote desktop by opening a terminal in the GUI.
Insert image here
Run napari
Open a terminal in the remote desktop environment.
Install uv:
curl -LsSf https://astral.sh/uv/install.sh | shCreate a uv virtual environment:
mkdir napari-playground cd napari-playground uv venvActivate the virtual environment and install napari:
source .venv/bin/activate uv install napari[all] # Install pyqt5 from apt to avoid an error with a missing Qt platform plugin sudo apt-get install python3-pyqt5Run napari:
uv run napari
Play with napari and examples
TODO
Destroy the droplet
After finishing my session, I destroyed the droplet to avoid incurring costs when it is not in use. You can do this from the DigitalOcean control panel by selecting the droplet and clicking on “Destroy”.
Attribution
Much of the information in this post is adapted from the following resources:
- https://bobcares.com/blog/digitalocean-vnc-ubuntu/
- https://www.digitalocean.com/community/tutorials/initial-server-setup-with-ubuntu-20-04
Conclusions
Automating this setup process would be a good next step. Using something like Ansible, Terraform, or shell scripts would make it easier to spin up the server with minimal typing.
Overall, I’m happy with the performance of napari to use the examples and begin interacting with napari for its development and testing purposes.
In the future, I’m hoping to use NVidia DGX Spark, which was announced earlier this year, to do GPU computing without a cloud instance. Looking forward to trying it out when it becomes available.
Until then, I’ve found the DigitalOcean GPU droplet to be a cost-effective option for experimenting with napari on a remote GPU vm.