fix: kubelet 1.20 device checkpoint support

[aisensiy]

The checkpoint file /var/lib/kubelet/device-plugins/kubelet_internal_checkpoint format changed in kubelet 1.20 (https://github.com/kubernetes/kubernetes/blob/d72c056260e771e8cd01220203582de0c0015786/pkg/kubelet/cm/devicemanager/checkpoint/checkpoint.go#L33-L43) which will make the current code crash.

The format of file changes from original:

{
  "Data": {
    "PodDeviceEntries": [
      {
        "PodUID": "...",
        "ContainerName": "...",
        "ResourceName": "...",
        "DeviceIDs": [ # <-------- KEY PART
          "GPU-xxxx"
        ],
        "AllocResp": "..."
      }
    ],
    "RegisteredDevices": {
      "nvidia.com/gpu": [
        ...
      ]
    }
  }
}

To:

{
  "Data": {
    "PodDeviceEntries": [
      {
        "PodUID": "...",
        "ContainerName": "...",
        "ResourceName": "...",
        "DeviceIDs": { # <-------- KEY PART
          "0": [
            "GPU-xxxx"
          ]
        },
        "AllocResp": "..."
      }
    ],
    "RegisteredDevices": {
      "nvidia.com/gpu": [
        ...
      ]
    }
  }
}

[mYmNeo]

Thank you for your commit, but you misunderstanding the checkpoint file of gpu-manager, gpu-manager use this file to recover deviceID, it doesn't care about other fields.

[aisensiy]

Thank you for your commit, but you misunderstanding the checkpoint file of gpu-manager, gpu-manager use this file to recover deviceID, it doesn't care about other fields.

I am not trying to change the structure of /etc/gpu-manager/checkpoint/gpumanager_internal_checkpoint. I am trying to fix the parse of /var/lib/kubelet/device-plugins/kubelet_internal_checkpoint.

[mYmNeo]

Thank you for your commit, but you misunderstanding the checkpoint file of gpu-manager, gpu-manager use this file to recover deviceID, it doesn't care about other fields.

I am not trying to change the structure of /etc/gpu-manager/checkpoint/gpumanager_internal_checkpoint. I am trying to fix the parse of /var/lib/kubelet/device-plugins/kubelet_internal_checkpoint.

I get your point.