acm-uiuc/gpu-cluster

Root repo for ACM GPU Cluster

gpu-cluster

Root repo for ACM GPU Cluster

Overview

The GPU Cluster Project breaks down into a couple parts.

Javascript

GPU Cluster Frontend - Frontend written in React, calls backend

Python

GPU Cluster Backend - Backend written in python, recieves requests and creates containers

nvdocker - A python package to manage interfacing with GPU enabled containers

[Nvidia] Docker

GPU Cluster Images - Standard Deep Learning Containers

Slurm

Cluster management software.

Getting Started

Make sure you have Node.js, Yarn and Python and Docker installed before starting

Install Node

Install Docker

Install nvidia-docker

There are a couple components of the GPU Cluster. To get all of the source run:

git clone --recursive [email protected]:acm-uiuc/gpu-cluster

You can also use repo if you like that better

1. Install repo - https://android.googlesource.com/tools/repo/

   Mac OS

   brew install repo 

   Ubuntu 14.04+

   sudo apt install repo
   

2. Make a directory to house your GPU Cluster work

   mkdir $GOPATH/src/github.com/acm-uiuc/gpu-cluster

3. Within this directory run the following command to start managing the projects

   repo init -u [email protected]:acm-uiuc/gpu-cluster

4. Run the following command to grab the latest of all the repos

   repo sync

If you are going to work on the Docker images, make sure to have an NVIDIA Graphics Card and have NVIDIA Docker installed in addition to Docker. https://github.com/NVIDIA/nvidia-docker

To start you can build the cluster with ./build_gpu_cluster. Then you can use ./gpu_cluster to start the interface on port 4000.

If you want to install the cluster as a daemon, place the repo in /usr/local/gpu-cluster or change gpu_cluster.ini to reflect the location of the ./gpu_cluster executable.

Then install circus with

pip install circus

Then create a systemd service following these instructions and pointing to gpu_cluster.ini

http://circus.readthedocs.io/en/latest/for-ops/deployment/

Deployed as a daemon using systemd and circus

gpu-cluster's Issues

Symlink '''data''' into '''common_data''' in everyone's workspace

Add COCO to the vault

Move documentation from gpu-cluster-images into root repo

The root repo has no documentation about how to use the system, and acm-uiuc/gpu-cluster-images has a link to https://www-s.acm.illinois.edu/confluence/display/PROJ/Using+the+ACM+GPU+Cluster.

This should be in this repo, and the wiki article should have more information about how to actually run code on the cluster/login to it. Currently, the wiki article is very high level, without specific details of how to use it.

How should we manage the list of installed containers

Presumably as the list of containers gets longer we need to have some amount of agreement between the frontend and the backend about which containers can be launched. Related to acm-uiuc/gpu-cluster-backend#9

Usage Regulation and Scheduling

How do we prevent users from mining bitcoin

Control both frontend and backend together

Figure out a way to control the frontend and backend together

Move into new org

We need a name and should move the project into a org

Rename project

Currently I am thinking of renaming this project to gpu-playground. We should discuss the rename here.

Allow for custom dockerfiles

We control volumes and networking

Create conda package (just like snap)

Token Authentication is blocking access

Two days ago it was still open to testing.

Interface for non Jupyter based containers

Things like job running or ssh

Create a gpu-cluster snap package

(easier than apt to figure out)