Running multi-instance GPUs

This page provides instructions on how to increase utilization and reduce costs by running multi-instance GPUs. With this configuration, you partition an NVIDIA A100 or H100 graphics processing unit (GPU) to share a single GPU across multiple containers on Google Kubernetes Engine (GKE).

Before reading this page, ensure that you're familiar with Kubernetes concepts such as Pods, nodes, deployments, and namespaces and are familiar with GKE concepts such as node pools, autoscaling, and auto-provisioning.

Introduction

Kubernetes allocates one full GPU per container even if the container only needs a fraction of the GPU for its workload, which might lead to wasted resources and cost overrun, especially if you are using the latest generation of powerful GPUs. To improve GPU utilization, multi-instance GPUs allow you to partition a single supported GPU in up to seven slices. Each slice can be allocated to one container on the node independently, for a maximum of seven containers per GPU. Multi-instance GPUs provide hardware isolation between the workloads, and consistent and predictable QoS for all containers running on the GPU.

For CUDA® applications, multi-instance GPUs are largely transparent. Each GPU partition appears as a regular GPU resource, and the programming model remains unchanged.

For more information on multi-instance GPUs, refer to the NVIDIA multi-instance GPU user guide.

Supported GPUs

The following GPU types support multi-instance GPUs:

• NVIDIA A100 (40GB)
• NVIDIA A100 (80GB)

• NVIDIA H100 (80GB)

Multi-instance GPU partitions

The A100 GPU and H100 GPU consist of seven compute units and eight memory units, which can be partitioned into GPU instances of varying sizes. The GPU partition sizes use the following syntax: [compute]g.[memory]gb. For example, a GPU partition size of 1g.5gb refers to a GPU instance with one compute unit (1/7th of streaming multiprocessors on the GPU), and one memory unit (5 GB). The partition size for the GPUs can be specified when you deploy an Autopilot workload or when you create a Standard cluster.

The partitioning table in the NVIDIA multi-instance GPU user guide lists all the different GPU partition sizes, along with the amount of compute and memory resources available on each GPU partition. The table also shows the number of GPU instances for each partition size that can be created on the GPU.

The following table lists the partition sizes that GKE supports:

Each GPU on each node within a node pool is partitioned the same way. For example, consider a node pool with two nodes, four GPUs on each node, and a partition size of 1g.5gb. GKE creates seven partitions of size 1g.5gb on each GPU. Since there are four GPUs on each node, there are 28 1g.5gb GPU partitions available on each node. Since there are two nodes in the node pool, a total of 56 1g.5gb GPU partitions are available in the entire node pool.

To create a GKE Standard cluster with more than one type of GPU partition, you must create multiple node pools. For example, if you want nodes with 1g.5gb and 3g.20gb GPU partitions in a cluster, you must create two node pools: one with the GPU partition size set to 1g.5gb, and the other with 3g.20gb.

A GKE Autopilot cluster automatically creates nodes with the correct partition configuration when you create workloads that require different partition sizes.

Each node is labeled with the size of GPU partitions that are available on the node. This labeling allows workloads to target nodes with the needed GPU partition size. For example, on a node with 1g.5gb GPU instances, the node is labeled as:

cloud.google.com/gke-gpu-partition-size=1g.5gb

How it works

To use multi-instance GPUs, you perform the following tasks:

1. Create a cluster with multi-instance GPUs enabled.
2. Manually install drivers.
3. Verify how many GPU resources are on the node.
4. Deploy containers using multi-instance GPUs.

Pricing

Multi-instance GPUs are exclusive to A100 GPUs and H100 GPUs and are subject to the corresponding GPU pricing in addition to any other products used to run your workloads. You can only attach whole GPUs to nodes in your cluster for partitioning. For GPU pricing information, refer to the GPUs pricing page.

Limitations

• Using multi-instance GPU partitions with GKE is not recommended for untrusted workloads.
• Autoscaling and auto-provisioning GPU partitions are fully supported on GKE version 1.20.7-gke.400 or later. In earlier versions only node pools with at least one node can be auto-scaled based on demand for specific GPU partition sizes from workloads.
• GPU utilization metrics (for example, duty_cycle) are not available for multi-instance GPUs.
• Multi-instance splits a physical GPU into discrete instances, each of which is isolated from the others at the hardware level. A container that uses a multi-instance GPU instance can only access the CPU and memory resources available to that instance.
• A pod can only consume up to one multi-instance GPU instance.

Before you begin

Before you start, make sure you have performed the following tasks:

• Enable the Google Kubernetes Engine API.
Enable Google Kubernetes Engine API
• If you want to use the Google Cloud CLI for this task, install and then initialize the gcloud CLI. If you previously installed the gcloud CLI, get the latest version by running gcloud components update.

• In Autopilot, multi-instance GPUs are supported in GKE version 1.29.3-gke.1093000 and later.
• You must have sufficient NVIDIA A100 GPU quota. See Requesting an increase in quota.
• If you want to use multi-instance GPUs with Autopilot, you can learn more about using GPUs with Autopilot at Deploy GPU workloads in Autopilot.
• GKE assigns the Accelerator compute class to all multi-instance GPU workloads in Autopilot clusters.

Create a cluster with multi-instance GPUs enabled

If you use GKE Standard, you must enable multi-instance GPUs in the cluster. Autopilot clusters that run version 1.29.3-gke.1093000 and later enable multi-instance GPUs by default. To use multi-instance GPUs in Autopilot, see the Deploy containers using multi-instance GPU section of this page.

When you create a Standard cluster with multi-instance GPUs, you must specify gpuPartitionSize along with acceleratorType and acceleratorCount. The acceleratorType must be nvidia-tesla-a100, nvidia-a100-80gb, or nvidia-h100-80gb.

The following example shows how to create a GKE cluster with one node, and seven GPU partitions of size 1g.5gb on the node. The other steps in this page use a GPU partition size of 1g.5gb, which creates seven partitions on each GPU. You can also use any of the supported GPU partition sizes mentioned earlier.

You can use the Google Cloud CLI or Terraform.

gcloud container clusters create CLUSTER_NAME  \
    --project=PROJECT_ID  \
    --zone ZONE  \
    --cluster-version=CLUSTER_VERSION  \
    --accelerator type=nvidia-tesla-a100,count=1,gpu-partition-size=1g.5gb,gpu-driver-version=DRIVER_VERSION  \
    --machine-type=a2-highgpu-1g  \
    --num-nodes=1

resource "google_container_cluster" "default" {
  name               = "gke-standard-zonal-gpu"
  location           = "us-central1-a"
  initial_node_count = 1

  node_config {
    guest_accelerator {
      type  = "nvidia-tesla-t4"
      count = 1
      gpu_driver_installation_config {
        gpu_driver_version = "LATEST"
      }
    }
    machine_type = "n1-standard-2"
  }

  # Set `deletion_protection` to `true` will ensure that one cannot
  # accidentally delete this instance by use of Terraform.
  deletion_protection = false
}

Connect to the cluster

Configure kubectl to connect to the newly created cluster:

gcloud container clusters get-credentials CLUSTER_NAME

Install drivers

If you chose to disable automatic driver installation when creating the cluster, or if you're running a GKE version earlier than 1.27.2-gke.1200, you must manually install a compatible NVIDIA driver after creation completes. Multi-instance GPUs require an NVIDIA driver version 450.80.02 or later.

After the driver is installed, multi-instance GPU mode is enabled. If you automatically installed drivers, your nodes reboot when the GPU device plugin starts to create GPU partitions. If you manually installed drivers, your nodes reboot when driver installation completes. The reboot might take a few minutes to complete.

Verify how many GPU resources are on the node

Run the following command to verify that the capacity and allocatable count of nvidia.com/gpu resources is 7:

kubectl describe nodes

Here's the output from the command:

...
Capacity:
  ...
  nvidia.com/gpu:             7
Allocatable:
  ...
  nvidia.com/gpu:             7

Deploy containers using multi-instance GPU

You can deploy up to one container per multi-instance GPU device on the node. In this example, with a partition size of 1g.5gb, there are seven multi-instance GPU partitions available on the node. As a result, you can deploy up to seven containers that request GPUs on this node.

1. Here's an example that starts the cuda:11.0.3-base-ubi7 container and runs nvidia-smi to print the UUID of the GPU within the container. In this example, there are seven containers, and each container receives one GPU partition. This example also sets the cloud.google.com/gke-gpu-partition-size node selector to target nodes with 1g.5gb GPU partitions.

   Autopilot

         cat &lt&ltEOF | kubectl apply -f -
         apiVersion: apps/v1
         kind: Deployment
         metadata:
           name: cuda-simple
         spec:
           replicas: 7
           selector:
             matchLabels:
               app: cuda-simple
           template:
             metadata:
               labels:
                 app: cuda-simple
             spec:
               nodeSelector:
                 cloud.google.com/gke-gpu-partition-size: 1g.5gb
                 cloud.google.com/gke-accelerator: nvidia-tesla-a100
                 cloud.google.com/gke-accelerator-count: "1"
               containers:
               - name: cuda-simple
                 image: nvidia/cuda:11.0.3-base-ubi7
                 command:
                 - bash
                 - -c
                 - |
                   /usr/local/nvidia/bin/nvidia-smi -L; sleep 300
                 resources:
                   limits:
                     nvidia.com/gpu: 1
         EOF
         

   This manifest does the following:

   • Requests the nvidia-tesla-a100 GPU type by setting the cloud.google.com/gke-accelerator node selector.
   • Splits the GPU into the 1g.5gb partition size.
   • Attaches a single GPU to the node by setting the cloud.google.com/gke-accelerator-count node selector.

   Standard

         cat &lt&ltEOF | kubectl apply -f -
         apiVersion: apps/v1
         kind: Deployment
         metadata:
           name: cuda-simple
         spec:
           replicas: 7
           selector:
             matchLabels:
               app: cuda-simple
           template:
             metadata:
               labels:
                 app: cuda-simple
             spec:
               nodeSelector:
                 cloud.google.com/gke-gpu-partition-size: 1g.5gb
               containers:
               - name: cuda-simple
                 image: nvidia/cuda:11.0.3-base-ubi7
                 command:
                 - bash
                 - -c
                 - |
                   /usr/local/nvidia/bin/nvidia-smi -L; sleep 300
                 resources:
                   limits:
                     nvidia.com/gpu: 1
         EOF
         

   This manifest does the following:

   • Requests a single GPU with partition size 1g.5gb.

2. Verify that all seven Pods are running:

   kubectl get pods
   

   Here's the output from the command:

   NAME                           READY   STATUS    RESTARTS   AGE
   cuda-simple-849c47f6f6-4twr2   1/1     Running   0          7s
   cuda-simple-849c47f6f6-8cjrb   1/1     Running   0          7s
   cuda-simple-849c47f6f6-cfp2s   1/1     Running   0          7s
   cuda-simple-849c47f6f6-dts6g   1/1     Running   0          7s
   cuda-simple-849c47f6f6-fk2bs   1/1     Running   0          7s
   cuda-simple-849c47f6f6-kcv52   1/1     Running   0          7s
   cuda-simple-849c47f6f6-pjljc   1/1     Running   0          7s
   

3. View the logs to see the GPU UUID, using the name of any Pod from the previous command:

   kubectl logs cuda-simple-849c47f6f6-4twr2
   

   Here's the output from the command:

   GPU 0: A100-SXM4-40GB (UUID: GPU-45eafa61-be49-c331-f8a2-282736687ab1)
     MIG 1g.5gb Device 0: (UUID: MIG-GPU-45eafa61-be49-c331-f8a2-282736687ab1/11/0)
   

What's next

• Learn more about GPUs.
• Learn how to configure time-sharing on GPUs.
• Learn more about cluster multi-tenancy.
• Learn more about best practices for enterprise multi-tenancy.