k8s-device-plugin VS metallb

https://github.com/NVIDIA/k8s-device-plugin/issues/332#issue...

K8s-device-plugin

This framework allows the use of external devices (e.g., NVIDIA GPUs, AMD GPUS, SR-IOV NICs) without modifying core Kubernetes components.

Not talking about MIG. NVIDIA device plugin. https://github.com/NVIDIA/k8s-device-plugin

GPU Operator Device plugin

If you ever tried to use GPU-based instances with AWS ECS, or on EKS using the default Nvidia plugin, you would know that it's not possible to make a task/pod shared the same GPU on an instance. If you want to add more replicas to your service (for redundancy or load balancing), you would need one GPU for each replica.

Some notes on Plex/Emby/Kodi and transcoding. If you want true transcoding with GPU acceleration, you have to have Nvidia GPU or be a k8s device plugin genius. The whole idea of mounting elastic devices in k8s is fairly new and rather complex. In the mean time transcoding is best done on a beefy device with a proper CPU (eg i7) or specifically Nvidia GPU because there are numerous pre-made plugins. I just run Plex and Emby on an old ATX gaming machine without GPU acceleration and it works totally fine. They were barely usable for just me when running on the RPis, wouldn't recommend it unless you can figure out how to mount the correct devices in the pod using a custom raspberry pi device plugin . . . lol good luck! - Arm labs device manager: https://community.arm.com/developer/research/b/articles/posts/a-smarter-device-manager-for-kubernetes-on-the-edge - Deis labs Akri device manager: https://github.com/deislabs/akri - Nvidia GPU plugin: https://github.com/NVIDIA/k8s-device-plugin

Hey guys, I want to share a guide I’m pretty proud of which is talking about setting up kubernetes which leverages https://kubespray.io/#/ and https://metallb.universe.tf/ so you can host this yourself most people when spinning up kubernetes opt for k3s or get stuck with all the options or unable to setup the external ips for their services so these tools will eliminate the problem.

Have a closer look at MetallLB.

To avoid that, you can use a combination of haproxy and keepalived, an enterprise grade load balancer like the one from F5 or Citrix. Besides that you can also work with https://kube-vip.io or https://metallb.universe.tf.

Not sure if klipper is usable in a cluster with multiple nodes, as it binds to one port only. You may want to use MetalLB instead: https://metallb.universe.tf/

Then deploy MetalLB https://metallb.universe.tf/

What issue do you see with the load balancer? For self hosted clusters, one can use MetalLB for example to have such single outfacing IP which will failover to another node keeping the same IP if a node dies.

Quoting from their docs:

metalLB to manage bare-metal LoadBalancer services - WIP - Only L2 configuration can be set-up via playbook.