Need to assign specific gpu or vgpu?

[jiangxiaobin96]

How to select which gpu or vgpu to allocate?

for _, str := range req.DevicesIDs {
	vGpuID, err := readVgpuIDFromFile(vGpuBasePath, str, "mdev_type/name")
	if err != nil || vGpuID != dpi.deviceName {
		log.Println("Could not get vGPU type identifier for device ", str)
		continue
	}

	key := fmt.Sprintf("%s_%s", vgpuPrefix, dpi.deviceName)
	if _, exists := envList[key]; !exists {
		envList[key] = []string{}
	}
	envList[key] = append(envList[key], str)
}

It seems that request YAML will assign specific vgpuID(str in code) and device plugin will use str to find specific vgpu in list to allocate.

[rthallisey]

Are you trying to assign workloads to a specific physical gpu? Or is your question about how to assign a vgpu a profile and expose it on a Kubernetes Node?

[jiangxiaobin96]

If a Kubernetes Node has multi gpu or vgpu, how to select which gpu or vgpu to use.

[rthallisey]

Let's say I have these gpus and vgpus exposed on my node:

    nvidia.com/1e37: "2"
    nvidia.com/GeForce_RTX_T10-8: "2"
    nvidia.com/GeForce_RTX_T10x-2: "8"
    nvidia.com/GeForce_RTX_T10x-4: "4"

I can't say, "I want gpu 1e37 with pci id 01:00.00" or "I want vgpu RTX_10-8 on gpu with pci id 01:00.00".
I can say, "I want a 1e37 gpu" or "I want an RTX_10-8" or "I want an RTX_10-8 and a 1e37".

There isn't an exposed interface to select gpus by pci id, however you can control where workloads are scheduled in k8s (https://kubernetes.io/docs/concepts/overview/working-with-objects/labels/), which you should use if you need that level of control.

[rthallisey]

Closing. Please reopen if you are still having an issue.