cryo: Allow GPU nodes to spawn across AZs

docs/howto/features/gpu.md

@@ -114,6 +114,9 @@ AWS, and we can configure a node group there to provide us GPUs.
         tags+: {
             "k8s.io/cluster-autoscaler/node-template/resources/nvidia.com/gpu": "1"
         },
+        // Allow provisioning GPUs across all AZs, to prevent situation where all
+        // GPUs in a single AZ are in use and no new nodes can be spawned
+        availabilityZones: masterAzs,
     }
    ```
 
@@ -122,6 +125,10 @@ AWS, and we can configure a node group there to provide us GPUs.
    1 GPU per node. If you're using a different machine type with
    more GPUs, adjust this definition accordingly.
 
+   We use a prior variable, `masterAzs`, to allow for GPU nodes to spawn in all
+   AZ in the region, rather than just a specific one. This is helpful as a single
+   zone may run out of GPUs rather fast.
+
 2. Render the `.jsonnet` file into a `.yaml` file that `eksctl` can use
 
    ```bash

eksctl/nasa-cryo.jsonnet

@@ -33,6 +33,9 @@ local notebookNodes = [
         tags+: {
             "k8s.io/cluster-autoscaler/node-template/resources/nvidia.com/gpu": "1"
         },
+        // Allow provisioning GPUs across all AZs, to prevent situation where all
+        // GPUs in a single AZ are in use and no new nodes can be spawned
+        availabilityZones: masterAzs,
     },
 ];
 

terraform/aws/efs.tf

@@ -1,17 +1,15 @@
-// Find out which subnet and security group our EFS mount target should be in
+// Find out which subnets and security group our EFS mount target should be in
 // It needs to be in the public subnet where our nodes are, as the nodes will be
 // doing the mounting operation. It should be in a security group shared by all
-// the nodes.
-data "aws_subnet" "cluster_node_subnet" {
+// the nodes. We create a mount target in each subnet, even if we primarily put
+// all our nodes in one - this allows for GPU nodes to be spread out across
+// AZ when needed
+data "aws_subnets" "cluster_node_subnets" {
 
   filter {
     name   = "vpc-id"
     values = [data.aws_eks_cluster.cluster.vpc_config[0]["vpc_id"]]
   }
-  filter {
-    name   = "availability-zone"
-    values = [var.cluster_nodes_location]
-  }
 
   filter {
     name   = "tag:aws:cloudformation:logical-id"
@@ -70,8 +68,10 @@ resource "aws_efs_file_system" "homedirs" {
 }
 
 resource "aws_efs_mount_target" "homedirs" {
+  for_each = toset(data.aws_subnets.cluster_node_subnets.ids)
+
   file_system_id  = aws_efs_file_system.homedirs.id
-  subnet_id       = data.aws_subnet.cluster_node_subnet.id
+  subnet_id       = each.key
   security_groups = [data.aws_security_group.cluster_nodes_shared_security_group.id]
 }