AWS capacity blocks with OpenTofu/terraform
2025-01-04 - Some pitfalls to avoid
Tags: AWS OpenTofu terraform
Introduction
AWS capacity blocks for machine learning are a short term GPU instance reservation mechanism. It is somewhat recent and has some rough edges when used via OpenTofu/terraform because of the incomplete documentation. I had to figure out things the hard way a few months ago, here they are.
EC2 launch template
When you reserve a capacity block, you get a capacity reservation id. You need to feed this id to an EC2 launch template. The twist is that you also need to use a specific instance market option not specified in the AWS provider’s documentation for this to work:
resource "aws_launch_template" "main" {
capacity_reservation_specification {
capacity_reservation_target {
capacity_reservation_id = "cr-XXXXXX"
}
}
instance_market_options {
market_type = "capacity-block"
}
instance_type = "p4d.24xlarge"
# soc2: IMDSv2 for all ec2 instances
metadata_options {
http_endpoint = "enabled"
http_put_response_hop_limit = 1
http_tokens = "required"
instance_metadata_tags = "enabled"
}
name = "imdsv2-${var.name}"
}
EKS node group
In order to use a capacity block reservation for a kubernetes node group, you need to:
- set a specific capacity type, not specified in the AWS provider’s documentation
- use an AMI with GPU support
- disable the kubernetes cluster autoscaler if you are using it (and you should)
resource "aws_eks_node_group" "main" {
for_each = var.node_groups
ami_type = each.value.gpu ? "AL2_x86_64_GPU" : null
capacity_type = each.value.capacity_reservation != null ? "CAPACITY_BLOCK" : null
cluster_name = aws_eks_cluster.main.name
labels = {
adyxax-gpu-node = each.value.gpu
adyxax-node-group = each.key
}
launch_template {
name = aws_launch_template.imdsv2[each.key].name
version = aws_launch_template.imdsv2[each.key].latest_version
}
node_group_name = each.key
node_role_arn = aws_iam_role.nodes.arn
scaling_config {
desired_size = each.value.scaling.min
max_size = each.value.scaling.max
min_size = each.value.scaling.min
}
subnet_ids = local.subnet_ids
tags = {
"k8s.io/cluster-autoscaler/enabled" = each.value.capacity_reservation == null
}
update_config {
max_unavailable = 1
}
version = local.versions.aws-eks.nodes-version
depends_on = [
aws_iam_role_policy_attachment.AmazonEC2ContainerRegistryReadOnly,
aws_iam_role_policy_attachment.AmazonEKSCNIPolicy,
aws_iam_role_policy_attachment.AmazonEKSWorkerNodePolicy,
]
lifecycle {
create_before_destroy = true
ignore_changes = [scaling_config[0].desired_size]
}
}
Conclusion
There is a terraform resource to provision the capacity blocks themselves that might be of interest, but I did not attempt to use it seriously. Capacity blocks are never available right when you create them, you need to book them days (sometimes weeks) in advance. Though OpenTofu/terraform has some basic date and time handling functions I could use to work around this, my needs are too sparse to go through the hassle of automating this.