Using single GPU with multiple containers

[ndesh26]

I was able to find some unofficial support for this in case of the Nvidia plugin. Is there similar support for this plugin?

I'm interested in this feature and would be willing to implement it. I want to know if there are other people who would be interested in such a feature. And also if the maintainers of this repo would like to add such a feature.

[thomas-riccardi]

Porting gpushare-device-plugin (and gpushare-scheduler-extender if it needs adaptation) to GKE would be much useful to us!
(This is the next step after our simple GPU sharing fork of container-engine-accelerators).

[danisla]

There is a simple way to do GPU sharing on GKE by creating symlinks on the host filesystem to the /dev/nvidia0 device. The plugin will pickup these new devices and register them with the node. There is no advanced scheduling, but it does let you attach the same GPU to multiple containers.

Here is an example I use in a daemonset to add the symlinks to the node after the NVIDIA driver has been installed:

# Create symlinks to NVIDIA device to support GPU sharing.
NVIDIA_0_SHARE=16

OLD_DEV=/dev/nvidia0
for i in $(seq 1 $(($NVIDIA_0_SHARE - 1))); do
  NEW_DEV=/dev/nvidia${i}
  echo "Linking $OLD_DEV -> $NEW_DEV"
  ln -sf $OLD_DEV $NEW_DEV
done

[ndesh26]

@danisla That's a nice, simple solution for a use case with basic scheduling.

[HenriTEL]

Another trick would be to make your node RAM match the GPU memory.
Assuming your pods restrict their GPU memory limit themself you can schedule using resources.requests.memory.

[ide8]

@danisla, could you please clarify where exactly this bash script should be applied?
It doesn't work in case when I run it strictly on the worker node.

[danisla]

@ide8 here is a DaemonSet that I apply to create the symlinks on new nodes.

[ide8]

@danisla , thank you but seems that it doesn't work with EKS
Trying to figure it out

[ndesh26]

@HenriTEL That would work when we just have a single node. But in case of multiple nodes having GPUs with different memories, it will become increasingly difficult to ensure proper scheduling and that we are not overcommitting our resources.

@ALL I will try to create a PoC for it since there are enough people interested in this. I will create a pull here when I'm done.

[ide8]

@ndesh26 , in our case it doesn't work even with a single node, with and without nvidia device plugin, but that's maybe specific to EKS.

[ndesh26]

@ide8 I was able to successfully run @danisla's solution on our system, your problem might be specific to EKS.

[HenriTEL]

@ide8 Note that the nvidia-device-plugin deployed by gke does not use nvidia-docker, it mounts devices in the containers directly.
I apply this DaemonSet to install nvidia-docker and set nvidia as default runtime.