Brief History
In March of 2017, I wrote about Opaque Integer Resources whereby specific hardware capabilities could be used in Kubernetes. Alpha in 1.5, it allowed for the potential to enable resource like Last Level Cache, GPUs and Many Integrated Core devices etc etc.
In Kubernetes 1.8, Opaque Integer Resources were replaced with Extended Resources. This was a great move as it migrated from a kubernetes.io/uri model to allow resources to be assigned to any domain outside kubernetes.io and thus simply extend the API with API aggregation.
Extended Resources are a phenomenal start to vastly expand opportunities around Kubernetes workloads but it still had the potential to require modifications to Kubernetes core in order to actually use a new resource. And this is where Device Plugins come in.
Requirements:
Kubernetes 1.8
DevicePlugins enabled in Kubelet
Device Plugins
Device Plugins is a common framework by which hardware devices for specific vendors can be plugged into Kubernetes.
Think of it this way:
Extended Resources = how to use a new resource Device Plugins = how vendors can advertise to and hook into Kubernetes without modifying Core
One of the first examples of Device Plugins in use is with Nvidia k8s-device-plugin. Which makes complete sense because Nvidia is leading an entire industry is various hardware arenas, GPU being just one of them.
How Device Plugins work
Device Plugins are/should be containers running in Kubernetes that provide access to a vendor (or enterprise) specific resource. The container advertises said resource to Kubelet via gRPC. Because this is hardware specific, it must be done on a per node basis. However a Daemonset can be deployed to cover a multitude of nodes for the same resource across multiple machines.
The Device Plugin has three parts:
Registration – plugin advertises itself to Kubelet over gPRC
ListandWatch – provides list of devices and/or modifies existing state of device on change including failures
Allocate – device specific instructions for Kubelet to make available to a container
At first glance this may seem rather simple but it should be noted that prior to Device Plugins, Kubelet specifically handled each device. Which is where hardware vendors had to contribute back to Kubernetes Core to provide net new hardware resources. With device plugin manager, this will be abstracted out and responsibility will lay on the vendor. Kubelet will then keep a socket open to ListandWatch for any changes in device state or list of devices.
Use of new devices through Extended Resources
Once a new Device Plugin is advertised to the cluster. It will be quite simple to use.
Now lets imagine we are using Nvidia’s GPU device plugin at nvidia.com/gpu
Here is how we allocate a gpu resource to a container.
apiVersion: v1
kind: Pod
metadata:
name: need-some-gpu-pod
spec:
containers:
- name: my-container-needing-gpu
image: myimage
resources:
requests:
cpu: 2
nvidia.com/gpu: 1
Gotchas
(At the time of this post)
Integers Only – this is common in Kubernetes but worth noting. 1 for gpu above can not be 0.5.
No Overallocation – Unlike Memory and CPU, devices can not be over allocated. So if Requests and Limits are specified, they must equal each other.
Resource Naming – I can’t confirm this but playing around with nvidia gpu I was unable to create multiple of the same device across multiple nodes.
Example:
I had difficulty advertising nvidia.com/gpu on node 2 once it was advertised on node one.
If correct, this would mean I would need to add nvidia.com/gpu-<node_name> or something of that measure to add the gpu device for multiple servers in a cluster. And also call out that specific device when assigning to the container requiring the resource. Now keep in mind, this is alpha so I would expect it to get modified rapidly but it is currently a limitation.
More info on Device Plugins
For a deeper review of the Device Plugin Manager
More on Extended Resources and Opaque Integer Resources
@devoperandi