Mudir

Your Machine Learning laboratory's workspace management tool

Mudir is a GPU workspace management tool, user-friendly, to protect your teams' research

Features

• You can use it with GPUs/CPUs servers
• Allows virtualization at low cost
• Jupyter environment support
• Two clicks to run your personal workspace
• A safe setup for experiments and datasets

System requirements

• Ubuntu18.04 on host server(s)
• MySQL 8.0.24 server
• NVIDIA GPU with installed CUDA drivers >= 10.x
• Docker runtime 19.03.12
• nvidia-container-toolkit
• nodejs 12.20.1
• ** Optional FTP server
• ** Optional docker-compose 1.29.0

Setup

 git clone git@github.com:moa-mudir/mudir.git
 cd mudir
 docker build -t <your-build-tag> .
 npm install
 cp default.env .env
 

• modify .env according to your preferences, you can change the host, password and image name you've just tagged to start

.env file

Plugin	Description	values (default)
DATABASE_HOST	<ip> of mysql database	localhost
DATABASE_USER	<username> of mysql database	root
DATABASE_PASSWORD	<password> of mysql database	super2009
DATABASE_PORT	<port> of database	3306
DATABASE_NAME	<name> of database	containers
BASE_DOCKER_IMAGE	<tag> of image	jupyterlab:1.0
DATASET_URI	root folder of your dataset	/mnt/dataset
TEMP_URI	tmp folder of your experiment	/mnt/exp
SHARED_GPU	Shared memory by GPU	4
UID	linux host user id	1000
GID	linux host group id	1000
DATA_VOLUME	(Optional) mounting database folder	/mnt/db

After configuring the system, you can test it by running node src/app.js

if it successfully runs you can set it as a service to run on system up:

• edit gpu-manager.service modify the EnvironmentFile WorkingDirectory to point to the .env and root folder of the app correspondingly
• sudo cp gpu-manager.service /etc/systemd/system/
• sudo systemctl daemon-reload
• sudo systemctl start gpu-manager.service
• sudo systemct status gpu-manager.service to check the status
• ** If you want to host the mysql on the same machine, you can run docker-compose up to spin a database server

Usage

• After successfully running the service, from your browser open <host-ip>:3000.

• Login with admin/admin credentials

• Navigate to workspaces to create/delete a workspace of a given name with description

• A workspace will have 2 ports, 1 for http jupyter and 1 for custom usage

• The name given to the workspace will create a directory on the filesystem (maybe mounted fileserver) to host the assets of experiment. Though allowing you to mount it over FTP and have persistent protected assets

• Navigate to Edit users to add/delete a user

• Clicking a workspace button when it is green it will redirect you to a jupyter environment

Contributing

• Use Commit standard
• When you work on issue, branch out a branch name as follows:
  • feature/MUD-XX-description
  • bugfix/MUD-XX-description