khazit / broker

Simple scheduler to manage a team's access to limited computational resources.

⚠️ No longer maintained

Broker

Broker was created as a tool for a team of 9 ML practitioners to manage their access to a single GPU machine. The idea was to avoid running multiple jobs at once (which would cause OOM errors), and to have a dashboard to visualize the jobs and their respective states.

According to Wikipedia:

A broker's prime responsibility is to bring sellers and buyers together and thus a broker is the third-person facilitator between a buyer and a seller.

In the same way, the global architecture of the system was designed around the Scheduler, the API that connects users with runners.

Theoretically, the system could scale to multiple runners. But has yet to be tested.

Install

Scheduler

To install the Scheduler on the main host machine (API + Dashboard):

Use docker-compose to build the two images

$ docker-compose -f docker/docker-compose-dev.yml build

And run the app:

$ docker-compose -f docker/docker-compose-dev.yml up

You can either use the Web UI ...

... or the bash script broker for command line usage.

Runners

Runners are single file python scripts, that you can run directly on your python env or on a docker container using docker/runner-dev.Dockerfile. The only dependancy is requests.

$ cd runner
$ python runner.py --scheduler_ip [IP] --scheduler_port [PORT]

Caveats

This is my first time writing an API / doing web dev. This is a development version that has no security at all.

Aknowledgements

The dashboard uses this template https://github.com/BootstrapDash/StarAdmin-Free-Vue-Admin-Template