Compute Node

Consists of compute-controller and compute-worker, currently these two are combined.

Compute nodes require high-performance GPUs and an external IP address.

Requirement

standard

• CPU: 16 core

• Memory: 32GB

• GPU: NVIDIA GPU with 12GB memory

• Storage: 1TB nvme

• Network: public IP, 10mbps

minimal

• CPU: 8 core

• Memory: 16GB

• GPU: NVIDIA GPU with 8GB memory

• Storage: 100GB ssd

• Network: public IP, 5mbps

Prepare

GPU Driver

1. Install the GPU driver

https://ubuntu.com/server/docs/nvidia-drivers-installation

After installation, check the graphics card information

> nvidia-smi

2. Install cuda require: cuda11.7.1

https://developer.nvidia.com/cuda-downloads

Install Docker

https://yeasy.gitbook.io/docker_practice/install

• Add User

Create a docker group and add the user to this group

sudo groupadd docker
sudo usermod -aG docker $USER
sudo systemctl restart docker
docker info

• Install docker nvidia toolkit

https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/install-guide.html

• pull image

docker pull nvidia/cuda:11.7.1-cudnn8-runtime-ubuntu20.04

Run

Initialize, using the ~/.compute directory by default

./compute-edge init

Using port 8081 by default, port 8081 needs to be accessible from the external network

./compute-edge daemon run -e EXPOSE_URL

For example, if the external IP is 1.2.3.4

./compute-edge daemon run -b 0.0.0.0:8081 -e http://1.2.3.4:8081