Using GPU Virtualization

This section describes how to use the GPU virtualization capability to isolate the computing power from the GPU memory and efficiently use GPU device resources.

Prerequisites

You have prepared GPU virtualization resources.
If you want to create a cluster using commands, use kubectl to connect to the cluster. For details, see Connecting to a Cluster Using kubectl.

Constraints

A single GPU can be virtualized into a maximum of 20 xGPU devices.
After GPU virtualization is used, init containers are not supported.
GPU virtualization supports two isolation modes: GPU memory isolation and isolation between GPU memory and computing power. A single GPU can schedule only workloads in the same isolation mode.
Autoscaler cannot be used to automatically scale in or out GPU nodes.
xGPU isolation does not allow you to request for GPU memory by calling CUDA API cudaMallocManaged(), which is also known as using UVM. For more information, see NVIDIA official documents. Use other methods to request for GPU memory, for example, by calling cudaMalloc().
When a containerized application is initializing, the real-time compute monitored by the nvidia-smi may exceed the upper limit of the available compute of the container.

Creating a GPU Virtualization Application

Using the CCE Console

Log in to the CCE console.
Click the cluster name to go to the cluster console, choose Workloads in the navigation pane, and click the Create Workload in the upper right corner.
Set basic information about the workload.

Choose Container Settings > Basic Info and configure the GPU quota.
- Video memory: The unit is MiB. The value must be a positive integer that is a multiple of 128. If the value exceeds the memory of a single GPU, scheduling cannot be performed.
- Computing power: The unit is %. The value must be a multiple of 5 and cannot exceed 100.
- If the GPU memory is set to the capacity upper limit of a single GPU or the computing power is set to 100%, the entire GPU will be used.
- When GPU virtualization is used, the workload scheduler defaults to Volcano and cannot be changed.
This section describes how to use GPU virtualization. For details about other parameters, see Workloads.

After completing the setting, click Create.

After a workload is created, you can try to verify the isolation capability of GPU virtualization.

kubectl exec -it gpu-app -- nvidia-smi

Expected output:

Wed Apr 12 07:54:59 2023       
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 470.141.03   Driver Version: 470.141.03   CUDA Version: 11.4     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  Tesla V100-SXM2...  Off  | 00000000:21:01.0 Off |                    0 |
| N/A   27C    P0    37W / 300W |   4912MiB /  5120MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
+-----------------------------------------------------------------------------+

The expected output indicates that the total GPU memory allocated to the pod is 5120 MiB, and 4912 MiB is used.

Run the following command on the node to check the isolation status of the GPU memory:

nvidia-smi

Expected output:

Wed Apr 12 09:31:10 2023        
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 470.141.03   Driver Version: 470.141.03   CUDA Version: 11.4     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  Tesla V100-SXM2...  Off  | 00000000:21:01.0 Off |                    0 |
| N/A   27C    P0    37W / 300W |   4957MiB / 16160MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|    0   N/A  N/A    760445      C   python                           4835MiB |
+-----------------------------------------------------------------------------+

The expected output indicates that the total GPU memory on the node is 16160 MiB, and the example pod uses 4957 MiB.

Using kubectl

Use kubectl to connect to the cluster.

Create an application that uses GPU virtualization.

The GPU memory isolation and isolation between GPU memory and computing power are supported. The computing power cannot be isolated only. volcano.sh and gpu-core.percentage cannot be set separately.

Create a gpu-app.yaml file with the following content.

Isolate GPU memory only:

apiVersion: apps/v1
kind: Deployment
metadata:
  name: gpu-app
  labels:
    app: gpu-app
spec:
  replicas: 1
  selector:
    matchLabels:
      app: gpu-app
  template: 
    metadata:
      labels:
        app: gpu-app
    spec:
      containers:
      - name: container-1
        image: <your_image_address>     # Replace it with your image address.
        resources:
          limits:
            volcano.sh/gpu-mem.128Mi: 40  # GPU memory allocated to the pod. The value is a multiple of 128 MiB (40 x 128 = 5120 MiB).
      imagePullSecrets:
        - name: default-secret
      schedulerName: volcano

Isolate the GPU memory from computing power:

apiVersion: apps/v1
kind: Deployment
metadata:
  name: gpu-app
  labels:
    app: gpu-app
spec:
  replicas: 1
  selector:
    matchLabels:
      app: gpu-app
  template: 
    metadata:
      labels:
        app: gpu-app
    spec:
      containers:
      - name: container-1
        image: <your_image_address>     # Replace it with your image address.
        resources:
          limits:
            volcano.sh/gpu-mem.128Mi: 40  # GPU memory allocated to the pod. The value is a multiple of 128 MiB (40 x 128 = 5120 MiB).
            volcano.sh/gpu-core.percentage: 25    # Computing power allocated to the pod
      imagePullSecrets:
        - name: default-secret
      schedulerName: volcano

**Table 1** Key parameters
Parameter	Mandatory	Description
volcano.sh/gpu-mem.128Mi	No	The value is a positive integer that is a multiple of 128 in the unit of MiB. If the value exceeds the memory of a single GPU, scheduling cannot be performed.
volcano.sh/gpu-core.percentage	No	The unit of the computing power is %. The value must be a multiple of 5 and cannot exceed 100.

If the GPU memory is set to the capacity upper limit of a single GPU or the computing power is set to 100%, the entire GPU will be used.
When GPU virtualization is used, the workload scheduler defaults to Volcano and cannot be changed.

Run the following command to create an application:
```
kubectl apply -f gpu-app.yaml
```

Verify the isolation capability of GPU virtualization.

kubectl exec -it gpu-app -- nvidia-smi

Expected output:

Wed Apr 12 07:54:59 2023       
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 470.141.03   Driver Version: 470.141.03   CUDA Version: 11.4     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  Tesla V100-SXM2...  Off  | 00000000:21:01.0 Off |                    0 |
| N/A   27C    P0    37W / 300W |   4912MiB /  5120MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
+-----------------------------------------------------------------------------+

The expected output indicates that the total GPU memory allocated to the pod is 5120 MiB, and 4912 MiB is used.

Run the following command on the node to check the isolation status of the GPU memory:

nvidia-smi

Expected output:

Wed Apr 12 09:31:10 2023        
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 470.141.03   Driver Version: 470.141.03   CUDA Version: 11.4     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  Tesla V100-SXM2...  Off  | 00000000:21:01.0 Off |                    0 |
| N/A   27C    P0    37W / 300W |   4957MiB / 16160MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|    0   N/A  N/A    760445      C   python                           4835MiB |
+-----------------------------------------------------------------------------+

The expected output indicates that the total GPU memory on the node is 16160 MiB, and the example pod uses 4957 MiB.

Parent topic: GPU Virtualization

Previous topic: Preparing xGPU Resources

Next topic: Supporting Kubernetes' Default GPU Scheduling