For scheduler and worker

1. Install conda virtual environment.
2. Run conda env create --prefix cs230 -f configuration.yaml under worker directory.
3. Activate the environment.
4. Install common library with pip install -e . command under common directory.
5. Install torch and torchvision with pip.

For FTP server and RabbitMQ broker

Deploy with docker from docker hub:

rabbitmq:3.12-management

garethflowers/ftp-server

For users

1. Install common library with pip install -e . command under common directory.

The RabbitMQ broker, FTP server, and the GPU capacity of each worker should be set up in the worker/config.json file.

    "broker" : {
        "broker_host": "18.119.97.104",
        "broker_port": "5673",
        "topics": {
            "broker_scheduling_topic": "node_1_scheduling"
        }
    },
    ...
    "ftp" : {
        "ftp_host": "169.234.56.23",
        "ftp_port": "21"
    },
    "workers": {
        "1": {
            "GPU": 8589934592
        },
        "2": {
            "GPU": 2147483648
        },
        "3": {
            "GPU": 0
        }
    },
    ...

Scheduler

There are three scheduling algorithm available:

python scheduler.py [next-available, round-robin, priority-based]

e.g.

python scheduler.py next-available

Worker

python daemon.py [worker_id]

e.g.

python daemon.py 1

Please look at tester.py to learn how a user should send a new task request to scheduler and retrieve the results.