Create an AI-optimized MIG with A4X

This document describes how to create a managed instance group (MIG) that uses A4X machine type. To learn more about the machine type, see A4X series in the Compute Engine documentation.

Create a MIG if you want to manage multiple virtual machines (VMs) as a single entity. To learn more about MIGs, see Managed instance groups in the Compute Engine documentation.

To learn about VM and cluster creation options, see Deployment options overview page.

Limitations

When you create a MIG with A4X VMs, the following limitations apply:

  • If you create a regional MIG, then the MIG can only create VMs in the zone that contains your VPC network's profile.

  • You can't configure instance flexibility in the MIG.

  • If you apply a workload policy to a MIG, you cannot change the policy in the MIG when the group has VMs in it. To change the policy in a MIG that has VMs, you must first resize the MIG to zero.

  • You can't use an instance template that specifies a placement policy to create a MIG that uses a workload policy.

  • You must use the reservation-bound provisioning model. Other provisioning models are not supported.

  • You can't use a resize request to add VMs to the MIG. You must set the target size of the MIG to add VMs.

Before you begin

Before creating a MIG, if you haven't already done so, complete the following steps:

  1. Choose a consumption option: the option that you pick determines how you want to get and use GPU resources.

    2. To learn more, see Choose a consumption option.

  2. Obtain capacity: to learn how to obtain capacity for your consumption option.

    3. To learn more, see Capacity overview.

Required roles

To get the permissions that you need to create a MIG, ask your administrator to grant you the Compute Instance Admin (v1) (roles/compute.instanceAdmin.v1) IAM role on the project. For more information about granting roles, see Manage access to projects, folders, and organizations.

This predefined role contains the permissions required to create a MIG. To see the exact permissions that are required, expand the Required permissions section:

Required permissions

The following permissions are required to create a MIG:

  • To create a MIG: compute.instanceGroupManagers.create on the project

You might also be able to get these permissions with custom roles or other predefined roles.

A4X fundamentals

An A4X cluster is organized into a hierarchy of blocks and subblocks to facilitate large-scale, non-blocking network performance. Understanding this topology is key when reserving capacity and deploying workloads.

A4X instance
An A4X instance is a single A4X machine type that has 4 GPUs attached.
NVLink domain or subblock
An NVLink domain, also referred to as a subblock, is the fundamental unit of A4X capacity. An NVLink domain consists of 18 A4X instances (72 GPUs) connected using a multi-node NVLink system. You create an A4X NVLink domain or a subblock by applying a workload policy that specifies a 1x72 topology.
Block
An A4X block is composed of 25 subblocks (NVLink domains), totalling up to 450 A4X instances (1,800 GPUs). The subblocks are rail-aligned for efficient scaling. Each subblock requires a MIG. Therefore, for a single A4X block, you can create 25 MIGs.

The following table shows the supported topology options for A4X instances:

Topology (acceleratorTopology) Number of GPUs Number of instances
1x72 72 18

Overview

Creating a MIG with A4X machine type includes the following steps:

  1. Create VPC networks
  2. Create a workload policy
  3. Create an instance template
  4. Create a MIG

Create VPC networks

To set up the network for A4X machine type, create three VPC networks for the following network interfaces:

  • 2 VPC networks for the gVNIC network interfaces (NIC). These are used for host to host communication.
  • 1 VPC network with the RDMA network profile is required for the CX-7 NICs when creating multiple A4X subblocks. This VPC network needs to have 4 subnets, one subnet for each CX-7 NIC. These NICs use RDMA over Converged Ethernet (RoCE), providing the high-bandwidth, low-latency communication that's essential for scaling out to multiple A4X subblocks. For a single A4X subblock, you can skip this VPC network because within a single subblock, direct GPU to GPU communication is handled by the multi-node NVLink.

For more information about NIC arrangement, see Review network bandwidth and NIC arrangement.

Create the networks either manually by following the instruction guides or automatically by using the provided script.

Instruction guides

To create the networks, you can use the following instructions:

For these VPC networks, we recommend setting the maximum transmission unit (MTU) to a larger value. For A4X machine type, the recommended MTU is 8896 bytes. To review the recommended MTU settings for other GPU machine types, see MTU settings for GPU machine types.

Script

To create the networks, follow these steps.

For these VPC networks, we recommend setting the maximum transmission unit (MTU) to a larger value. For A4X machine type, the recommended MTU is 8896 bytes. To review the recommended MTU settings for other GPU machine types, see MTU settings for GPU machine types.

  1. Use the following script to create VPC networks for the gVNICs.

      
    #!/bin/bash

    # Create standard VPCs (network and subnets) for the gVNICs
    for N in $(seq 0 1); do
      gcloud compute networks create GVNIC_NAME_PREFIX-net-$N \
        --subnet-mode=custom \
        --mtu=8896

      gcloud compute networks subnets create GVNIC_NAME_PREFIX-sub-$N \
        --network=GVNIC_NAME_PREFIX-net-$N \
        --region=REGION \
        --range=192.168.$N.0/24

      gcloud compute firewall-rules create GVNIC_NAME_PREFIX-internal-$N \
        --network=GVNIC_NAME_PREFIX-net-$N \
        --action=ALLOW \
        --rules=tcp:0-65535,udp:0-65535,icmp \
        --source-ranges=192.168.0.0/16
    done

    # Create SSH firewall rules
    gcloud compute firewall-rules create GVNIC_NAME_PREFIX-ssh \
      --network=GVNIC_NAME_PREFIX-net-0 \
      --action=ALLOW \
      --rules=tcp:22 \
      --source-ranges=IP_RANGE

    # Assumes that an external IP is only created for vNIC 0
    gcloud compute firewall-rules create GVNIC_NAME_PREFIX-allow-ping-net-0 \
      --network=GVNIC_NAME_PREFIX-net-0 \
      --action=ALLOW \
      --rules=icmp \
      --source-ranges=IP_RANGE

  2. If you require multiple A4X subblocks, use the following script to create the VPC network and subnets for the four CX-7 NICs on each A4X instance.

      
    # List and make sure network profiles exist in the machine type's zone
    gcloud compute network-profiles list --filter "location.name=ZONE"

    # Create network for CX-7
    gcloud compute networks create RDMA_NAME_PREFIX-mrdma \
      --network-profile=ZONE-vpc-roce \
      --subnet-mode custom \
      --mtu=8896

    # Create subnets
    for N in $(seq 0 3); do
      gcloud compute networks subnets create RDMA_NAME_PREFIX-mrdma-sub-$N \
        --network=RDMA_NAME_PREFIX-mrdma \
        --region=REGION \
        --range=192.168.$((N+2)).0/24 & # offset to avoid overlap with gVNICs
    done

    Replace the following:

    • GVNIC_NAME_PREFIX: the custom name prefix to use for the standard VPC networks and subnets for the gVNICs.
    • RDMA_NAME_PREFIX: the custom name prefix to use for the VPC network and subnets with the RDMA network profile for the CX-7 NICs.
    • ZONE: specify a zone in which the machine type that you want to use is available, such as us-central1-a. For information about regions, see GPU availability by regions and zones.
    • REGION: the region where you want to create the subnets. This region must correspond to the zone specified. For example, if your zone is us-central1-a, then your region is us-central1.
    • IP_RANGE: the IP range to use for the SSH firewall rules.
  3. Optional: To verify that the VPC network resources are created successfully, check the network settings in the Google Cloud console:
    1. In the Google Cloud console, go to the VPC networks page.

      Go to VPC networks

    2. Search the list for the networks that you created in the previous step.
    3. To view the subnets, firewall rules, and other network settings, click the name of the network.

Create a workload policy

You cannot update a workload policy after you create it. To make changes in a workload policy, you must create a new one.

To create a workload policy, select one of the following options:

gcloud

To create a workload policy, use the gcloud compute resource-policies create workload-policy command.

For inter-accelerator network configuration, specify the --accelerator-topology flag in the command.

Use the following command for A4X machine type. The accelerator topology of 1x72 indicates 72 GPUs connected via a NVLink within a block.

gcloud compute resource-policies create workload-policy WORKLOAD_POLICY_NAME \
    --type=high-throughput \
    --accelerator-topology=1x72 \
    --region=REGION

Replace the following:

  • WORKLOAD_POLICY_NAME: the name of the workload policy.
  • REGION: the region where you want to create the workload policy. Specify a region in which you want to create the MIG and the machine type that you want to use is available. For information about regions, see GPU availability by regions and zones.

REST

To create a workload policy, make a POST request to the resourcePolicies.insert method.

For inter-accelerator network configuration, specify the acceleratorTopology field in the request.

Make the following request for A4X machine type. The accelerator topology of 1x72 indicates 72 GPUs connected via a NVLink within a block.

POST https://compute.googleapis.com/compute/v1/projects/PROJECT_ID/regions/REGION/resourcePolicies
  {
    "name": "WORKLOAD_POLICY_NAME"
    "workloadPolicy": {
      "type": "HIGH_THROUGHPUT",
      "acceleratorTopology": "1x72"
    }
  }

Replace the following:

  • PROJECT_ID: your project ID
  • REGION: the region where you want to create the workload policy. Specify a region in which you want to create the MIG and the machine type that you want to use is available. For information about regions, see GPU availability by regions and zones.
  • WORKLOAD_POLICY_NAME: the name of the workload policy.

Create an instance template

Specify the VM properties for a MIG by creating an instance template.

To create an instance template, select one of the following options.

The following commands also set the access scope for your instances. To simplify permissions management, Google recommends that you set the access scope on an instance to cloud-platform access and then use IAM roles to define what services the instance can access. For more information, see Scopes best practice.

gcloud

To create a regional instance template, use the gcloud compute instance-templates create command.

gcloud compute instance-templates create INSTANCE_TEMPLATE_NAME \
    --machine-type=a4x-highgpu-4g \
    --image-family=IMAGE_FAMILY \
    --image-project=IMAGE_PROJECT \
    --instance-template-region=REGION \
    --boot-disk-type=hyperdisk-balanced \
    --boot-disk-size=DISK_SIZE \
    --scopes=cloud-platform \
    --network-interface=nic-type=GVNIC,network=GVNIC_NAME_PREFIX-net-0,subnet=GVNIC_NAME_PREFIX-sub-0 \
    --network-interface=nic-type=GVNIC,network=GVNIC_NAME_PREFIX-net-1,subnet=GVNIC_NAME_PREFIX-sub-1,no-address \
    --network-interface=nic-type=MRDMA,network=RDMA_NAME_PREFIX-mrdma,subnet=RDMA_NAME_PREFIX-mrdma-sub-0,no-address \
    --network-interface=nic-type=MRDMA,network=RDMA_NAME_PREFIX-mrdma,subnet=RDMA_NAME_PREFIX-mrdma-sub-1,no-address \
    --network-interface=nic-type=MRDMA,network=RDMA_NAME_PREFIX-mrdma,subnet=RDMA_NAME_PREFIX-mrdma-sub-2,no-address \
    --network-interface=nic-type=MRDMA,network=RDMA_NAME_PREFIX-mrdma,subnet=RDMA_NAME_PREFIX-mrdma-sub-3,no-address \
    --reservation-affinity=specific \
    --reservation=RESERVATION \
    --provisioning-model=RESERVATION_BOUND \
    --instance-termination-action=DELETE \
    --maintenance-policy=TERMINATE

Replace the following:

  • INSTANCE_TEMPLATE_NAME: the name of the instance template.
  • IMAGE_FAMILY: the image family of the OS image that you want to use. For a list of supported operating systems, see Supported operating systems.
  • IMAGE_PROJECT: the project ID of the OS image.
  • REGION: the region where you want to create the instance template. Specify a region in which the machine type that you want to use is available. For information about regions, see GPU availability by regions and zones.
  • DISK_SIZE: the size of the boot disk in GB.
  • GVNIC_NAME_PREFIX: the name prefix that you specified when creating the standard VPC networks and subnets that use gVNIC NICs.
  • RDMA_NAME_PREFIX: the name prefix that you specified when creating the VPC networks and subnets that use RDMA NICs.
  • RESERVATION: the reservation name, a block, or a subblock within a reservation. To get the reservation name or the available blocks, see View reserved capacity. Based on your requirements for instance placement, choose one of the following:
    • To create the instances on any single block: 
          projects/RESERVATION_OWNER_PROJECT_ID/reservations/RESERVATION_NAME
    • To create the instances on a specific block: 
          projects/RESERVATION_OWNER_PROJECT_ID/reservations/RESERVATION_NAME/reservationBlocks/RESERVATION_BLOCK_NAME
    • To create the instances in a specific subblock: 
          projects/RESERVATION_OWNER_PROJECT_ID/reservations/RESERVATION_NAME/reservationBlocks/RESERVATION_BLOCK_NAME/subblock/RESERVATION_SUBBLOCK_NAME

REST

To create a regional instance template, make a POST request to the regionInstanceTemplates.insert method.

POST https://compute.googleapis.com/compute/v1/projects/PROJECT_ID/regions/REGION/instanceTemplates
{
  "name":"INSTANCE_TEMPLATE_NAME",
  "properties":{
    "machineType":"a4x-highgpu-4g",
    "disks":[
      {
        "boot":true,
        "initializeParams":{
          "diskSizeGb":"DISK_SIZE",
          "diskType":"hyperdisk-balanced",
          "sourceImage":"projects/IMAGE_PROJECT/global/images/family/IMAGE_FAMILY"
        },
        "mode":"READ_WRITE",
        "type":"PERSISTENT"
      }
    ],
    "serviceAccounts": [
      {
        "email": "default",
        "scopes": [
          "https://www.googleapis.com/auth/cloud-platform"
        ]
      }
    ],
    "networkInterfaces": [
      {
        "accessConfigs": [
          {
            "name": "external-nat",
            "type": "ONE_TO_ONE_NAT"
          }
        ],
        "network": "projects/NETWORK_PROJECT_ID/global/networks/GVNIC_NAME_PREFIX-net-0",
        "nicType": "GVNIC",
        "subnetwork": "projects/NETWORK_PROJECT_ID/region/REGION/subnetworks/GVNIC_NAME_PREFIX-sub-0"
      },
      {
        "network": "projects/NETWORK_PROJECT_ID/global/networks/GVNIC_NAME_PREFIX-net-1",
        "nicType": "GVNIC",
        "subnetwork": "projects/NETWORK_PROJECT_ID/region/REGION/subnetworks/GVNIC_NAME_PREFIX-sub-1"
      },
      {
        "network": "projects/NETWORK_PROJECT_ID/global/networks/RDMA_NAME_PREFIX-mrdma",
        "nicType": "MRDMA",
        "subnetwork": "projects/NETWORK_PROJECT_ID/region/REGION/subnetworks/RDMA_NAME_PREFIX-mrdma-sub-0"
      },
      {
        "network": "projects/NETWORK_PROJECT_ID/global/networks/RDMA_NAME_PREFIX-mrdma",
        "nicType": "MRDMA",
        "subnetwork": "projects/NETWORK_PROJECT_ID/region/REGION/subnetworks/RDMA_NAME_PREFIX-mrdma-sub-1"
      },
      {
        "network": "projects/NETWORK_PROJECT_ID/global/networks/RDMA_NAME_PREFIX-mrdma",
        "nicType": "MRDMA",
        "subnetwork": "projects/NETWORK_PROJECT_ID/region/REGION/subnetworks/RDMA_NAME_PREFIX-mrdma-sub-2"
      },
      {
        "network": "projects/NETWORK_PROJECT_ID/global/networks/RDMA_NAME_PREFIX-mrdma",
        "nicType": "MRDMA",
        "subnetwork": "projects/NETWORK_PROJECT_ID/region/REGION/subnetworks/RDMA_NAME_PREFIX-mrdma-sub-3"
      }
    ],
    "reservationAffinity":{
        "consumeReservationType":"SPECIFIC_RESERVATION",
        "key":"compute.googleapis.com/reservation-name",
        "values":[
          "RESERVATION"
        ]
      },
    "scheduling":{
        "provisioningModel":"RESERVATION_BOUND",
        "instanceTerminationAction":"DELETE",
        "onHostMaintenance": "TERMINATE",
        "automaticRestart":true
      }
  }
}

Replace the following:

  • INSTANCE_TEMPLATE_NAME: the name of the instance template.
  • IMAGE_FAMILY: the image family of the OS image that you want to use. For a list of supported operating systems, see Supported operating systems.
  • IMAGE_PROJECT: the project ID of the OS image.
  • REGION: the region where you want to create the instance template. Specify a region in which the machine type that you want to use is available. For information about regions, see GPU availability by regions and zones.
  • DISK_SIZE: the size of the boot disk in GB.
  • NETWORK_PROJECT_ID: the project ID of the network.
  • GVNIC_NAME_PREFIX: the name prefix that you specified when creating the standard VPC networks and subnets that use gVNIC NICs.
  • REGION: the region of the subnetwork.
  • RDMA_NAME_PREFIX: the name prefix that you specified when creating the VPC networks and subnets that use RDMA NICs.
  • RESERVATION: the reservation name, a block, or a subblock within a reservation. To get the reservation name or the available blocks, see View reserved capacity. Based on your requirements for instance placement, choose one of the following:
    • To create the instances on any single block: 
          projects/RESERVATION_OWNER_PROJECT_ID/reservations/RESERVATION_NAME
    • To create the instances on a specific block: 
          projects/RESERVATION_OWNER_PROJECT_ID/reservations/RESERVATION_NAME/reservationBlocks/RESERVATION_BLOCK_NAME
    • To create the instances in a specific subblock: 
          projects/RESERVATION_OWNER_PROJECT_ID/reservations/RESERVATION_NAME/reservationBlocks/RESERVATION_BLOCK_NAME/subblock/RESERVATION_SUBBLOCK_NAME

Create a MIG

When you use an A4X machine type, you must create a MIG by specifying its target size.

To obtain a GPU topology of 1x72, create a MIG with 18 instances. When you create the MIG, apply the workload policy that specifies the acceleratorTopology field. Applying the policy ensures that Compute Engine creates all 18 instances in one subblock to use an NVLink domain. If a subblock lacks capacity for all 18 instances, then any managed instance that can't be created immediately will remain in CREATING state until capacity becomes available. These managed instances represent the VMs that the MIG creates when there is capacity.

When you apply a workload policy with the acceleratorTopology field set to 1x72, you cannot create more than 18 instances in a MIG. Specifying more than 18 instances causes the MIG creation to fail. To create multiple NVLink domains, create a separate MIG for each domain and apply the same workload policy to each MIG.

To create a MIG, select one of the following options:

gcloud

To create a MIG with a specified target size, use the instance-groups managed create command.

Create a zonal or regional MIG as follows:

  • To create a zonal MIG, use the following command: 
    gcloud compute instance-groups managed create MIG_NAME \
  --template=INSTANCE_TEMPLATE_URL \
  --size=TARGET_SIZE \
  --workload-policy=WORKLOAD_POLICY_URL \
  --zone=ZONE
  • To create a regional MIG, use the following command: 
    gcloud compute instance-groups managed create MIG_NAME \
    --template=INSTANCE_TEMPLATE_URL \
    --size=TARGET_SIZE \
    --workload-policy=WORKLOAD_POLICY_URL \
    --region=REGION
Replace the following:
  • MIG_NAME: the name of the MIG.
  • INSTANCE_TEMPLATE_URL: the URL of the instance template that you want to use to create VMs in the MIG. The URL can contain either the ID or name of the instance template. Specify one of the following values:
    • For a regional instance template: projects/PROJECT_ID/regions/REGION/instanceTemplates/INSTANCE_TEMPLATE_ID
    • For a global instance template: INSTANCE_TEMPLATE_ID
  • TARGET_SIZE: the number of VMs that you want in the MIG. For the accelerator topology of 1x72, set the target size to 18.
  • WORKLOAD_POLICY_URL: the URL of the workload policy. For example, projects/example-project/regions/us-central1/resourcePolicies/example-workload-policy.
  • ZONE: the zone in which you want to create the MIG. Specify a zone within the workload policy's region.
  • REGION: the region in which you want to create the MIG. Specify the same region as that of the workload policy. For a regional MIG, instead of a region, you can specify the zones in that region by using the --zones flag.

REST

To create a MIG with a specified target size, make a POST request.

Create a zonal or regional MIG as follows:

  • To create a zonal MIG, make a POST request to the instanceGroupManagers.insert method. 
    POST https://compute.googleapis.com/compute/v1/projects/PROJECT_ID/zones/ZONE/instanceGroupManagers
{
  "versions": [
    {
      "instanceTemplate": "INSTANCE_TEMPLATE_URL"
    }
  ],
  "name": "MIG_NAME",
  "instanceTemplate": "INSTANCE_TEMPLATE_URL",
  "targetSize": "TARGET_SIZE",
  "resourcePolicies": {
    "workloadPolicy": WORKLOAD_POLICY_URL
  }
}
  • To create a regional MIG, make a POST request to the regionInstanceGroupManagers.insert method. 
      POST https://compute.googleapis.com/compute/v1/projects/PROJECT_ID/regions/REGION/instanceGroupManagers
  {
    "versions": [
      {
        "instanceTemplate": "INSTANCE_TEMPLATE_URL"
      }
    ],
    "name": "MIG_NAME",
    "instanceTemplate": "INSTANCE_TEMPLATE_URL",
    "targetSize": "TARGET_SIZE",
    "resourcePolicies": {
      "workloadPolicy": WORKLOAD_POLICY_URL
    }
  }
Replace the following:
  • PROJECT_ID: the project ID.
  • ZONE: the zone in which you want to create the MIG. Specify a zone within the workload policy's region.
  • REGION: the region in which you want to create the MIG. Specify the same region as that of the workload policy.
  • INSTANCE_TEMPLATE_URL: the URL of the instance template that you want to use to create VMs in the MIG. The URL can contain either the ID or name of the instance template. Specify one of the following values:
    • For a regional instance template: projects/PROJECT_ID/regions/REGION/instanceTemplates/INSTANCE_TEMPLATE_ID
    • For a global instance template: INSTANCE_TEMPLATE_ID
  • MIG_NAME: the name of the MIG.
  • TARGET_SIZE: the number of VMs that you want in the MIG. For the accelerator topology of 1x72, set the target size to 18.
  • WORKLOAD_POLICY_URL: the URL of the workload policy. For example, projects/example-project/regions/us-central1/resourcePolicies/example-workload-policy.

What's next?