A3 clusters may be created through either GKE or a MIG via the modules found here. Due to the recency of A3's release, features are limited in each control plane, and those limitations are listed below.
Feature \ Module | gke |
mig-cos |
---|---|---|
VM Image | COS-Cloud | COS-Cloud |
Compact placement policy | Yes | Yes |
Kubernetes support | Yes | No |
An A3 cluster of eight nodes (two node pools with four nodes each) booting with a COS-Cloud image can be created via GKE by running the following two commands:
cat >./terraform.tfvars <<EOF
project_id = "my-project"
region = "us-central1"
resource_prefix = "my-cluster"
node_pools = [
{
zone = "us-central1-c"
node_count = 4
},
{
zone = "us-central1-c"
node_count = 4
},
]
EOF
docker run --rm -v "${PWD}:/root/aiinfra/input" \
us-docker.pkg.dev/gce-ai-infra/cluster-provision-dev/cluster-provision-image:latest \
create a3 gke
A deeper dive into how to use this tool can be found at the top-level README.
An A3 cluster of eight nodes (two instance groups with four instances each) booting with a COS-Cloud image can be created via a managed instance group by running the following two commands:
cat >./terraform.tfvars <<EOF
instance_groups = [
{
target_size = 4
zone = "us-central1-c"
},
{
target_size = 4
zone = "us-central1-c"
},
]
project_id = "my-project"
region = "us-central1"
resource_prefix = "my-cluster"
EOF
docker run --rm -v "${PWD}:/root/aiinfra/input" \
us-docker.pkg.dev/gce-ai-infra/cluster-provision-dev/cluster-provision-image:latest \
create a3 mig-cos
A deeper dive into how to use this tool can be found at the top-level README.