GPU Provider

What is GPU Provider

• Tool for running gpu workloads deployed as standard pods in significantly cheaper cloud provider like VastAI and others
• GPU Provider is the extension of virtual kubelet. Deploying it into cluster creates virtual node. Pods scheduled there will spawn its containers in selected cloud provider (VastAI)
• Provider exposes Gateway with Wireguard endpoint to which remote containers connect. This allows for full communication between containers and rest of the cluster
• Scheduling will be done by default scheduler. No changes to existing infra needs to be done

Limitations

• only 1 container per pod applicable for: VastAI

• no image command default (must be explicitly specified) applicable for: all

• kubectl commands acting on node (logs|exec) are not available applicable for: managed clusters without access to root CA

• user-space limited networking model applicable for: VastAI

  needs to connect over HTTP Proxy or configure TLS tunnel

Unimplemented Features (as of v0.2.0)

• service accounts