Is it possible to define the devices id myself or to get the devices id using k8s command?

[dwqdqcadsc]

I found that the device id in this plugin is generated by SHA. It is not the same as the real device id. How can I easily get these generated ids?

[squat]

Hi @dwqdqcadsc I'm not sure what you mean. What device are you looking at? Please share the plug-in configuration as well as the desired vs real device paths are loaded into your container. That should help debug.

[dwqdqcadsc]

Hello, thank you for answering my question.
func (gp *GenericPlugin) discoverPath() ([]device, error) {
var devices []device
var mountPath string
for _, group := range gp.ds.Groups {
paths := make([][]string, len(group.Paths))
var length int
var limitLength int
// Discover all the devices matching each pattern in the Paths group.
for i, path := range group.Paths {
matches, err := fs.Glob(gp.fs, path.Path)
if err != nil {
return nil, err
}
sort.Strings(matches)
for j := uint(0); j < path.Limit; j++ {
paths[i] = append(paths[i], matches...)
}
// Keep track of the shortest reusable length in the group.
if i == 0 || len(paths[i]) < limitLength {
limitLength = len(paths[i])
}
// Keep track of the greatest natural length in the group.
if len(matches) > length {
length = len(matches)
}
}
// Cap the length at the maximum reusable length.
if length > limitLength {
length = limitLength
}
for i := 0; i < length; i++ {
for j := uint(0); j < group.Count; j++ {
h := sha1.New()
h.Write([]byte(strconv.FormatUint(uint64(j), 10)))
d := device{
Device: &v1beta1.Device{
Health: v1beta1.Healthy,
},
}
for k, path := range group.Paths {
mountPath = path.MountPath
if mountPath == "" {
mountPath = paths[k][i]
}
switch path.Type {
case DevicePathType:
d.deviceSpecs = append(d.deviceSpecs, &v1beta1.DeviceSpec{
HostPath: paths[k][i],
ContainerPath: mountPath,
Permissions: path.Permissions,
})
case MountPathType:
d.mounts = append(d.mounts, &v1beta1.Mount{
HostPath: paths[k][i],
ContainerPath: mountPath,
ReadOnly: path.ReadOnly,
})
}
h.Write([]byte(paths[k][i]))
}
d.ID = fmt.Sprintf("%x", h.Sum(nil))
devices = append(devices, d)
}
}
}
return devices, nil
}
this is the code about discovering the devices. please see this line :d.ID = fmt.Sprintf("%x", h.Sum(nil)). I need to get this d.ID in a easy way or i want to define it myself. In my work, I will provide k8s many devices made by myself. I need to know which device is exactly used in which pod. That is why i need to get this d.ID or i want to define it myself. If you know a easy way to get this , please let me know . Thank you so much~

[squat]

@dwqdqcadsc The device ID you are describing is only for the used to provide a stable ID to the Kubelet for purposes of tracking when devices appear or disappear and tracking which devices are currently occupied. It has nothing to do with the device ID that is ultimately exposed on the container's file system under the /dev hierarchy.

If you still need this device ID, then you can generate it yourself using the same logic you shared. However, this not supposed to be part of any public interface for the device plugin so we can't guarantee that the way this ID is calculated won't change in the future.

One reason why this device ID cannot simply be the same as the device ID in the kernel is that the generic device plugin (GDP) allows administrators to specify that a device should be mountable multiple times, e.g. allowing the FUSE device to be mounted unlimited times. Each FUSE device would require a unique ID and so we can't just use the device ID.

I am certainly open to other ways of creating unique and consistent device IDs that are more predictable, e.g. a predictable device path followed by a GDP-specific suffix that tools could ignore. Do you have anything in mind?

[dwqdqcadsc]

Thank you for explaining this. Though there are some devices can be mounted unlimited times, device like gpu seems can be only used mounted one time right? It seems that nvidia gpu will make its device id as environment variable in pod. In my scenario, I simulated Nvidia graphics cards, created simulation devices, and registered them in k8s through gdp. For real Nvidia graphics card devices, I can obtain their device IDs in the pod to monitor which specific graphics card is used by the large models I deploy in the pod. However, when I use gdp, I cannot achieve the above logic.

[squat]

Is it possible to ... get the devices id using k8s command?

You should always be able to find some information about the device by doing a kubectl exec -it <pod-name> -- nvidia-smi (or the analogous command for you GPU maker) and process the output accordingly. Does something like this work for you?