Added support for NFS server in Soperator (#112)

* Added support for NFS server in Soperator

* Move nfs cloud-init into module folder

* Added variables for nfs server presets in soperator

* Customize mount path for nfs server

modules/nfs-server/main.tf

@@ -28,7 +28,7 @@ resource "nebius_compute_v1_instance" "nfs_server" {
     }
   ]
 
-  cloud_init_user_data = templatefile("../modules/cloud-init/nfs-cloud-init.tftpl", {
+  cloud_init_user_data = templatefile("${path.module}/files/nfs-cloud-init.tftpl", {
     ssh_user_name  = var.ssh_user_name,
     ssh_public_key = var.ssh_public_key,
     nfs_ip_range   = var.nfs_ip_range,

soperator/installations/example/main.tf

@@ -68,6 +68,24 @@ module "filestore" {
   }
 }
 
+module "nfs-server" {
+  count          = var.nfs.enabled ? 1 : 0
+  source         = "../../../modules/nfs-server"
+  parent_id      = data.nebius_iam_v1_project.this.id
+  subnet_id      = data.nebius_vpc_v1_subnet.this.id
+  ssh_user_name  = "soperator"
+  ssh_public_key = var.slurm_login_ssh_root_public_keys[0]
+  nfs_ip_range   = data.nebius_vpc_v1_subnet.this.ipv4_private_pools.pools[0].cidrs[0].cidr
+  nfs_size       = var.nfs.size_gibibytes * 1024 * 1024 * 1024
+  nfs_path       = "/mnt/nfs"
+  platform       = var.nfs.resource.platform
+  preset         = var.nfs.resource.preset
+
+  providers = {
+    nebius = nebius
+  }
+}
+
 module "k8s" {
   depends_on = [
     module.filestore,
@@ -249,6 +267,13 @@ module "slurm" {
     } : null
   }
 
+  nfs = {
+    enabled    = var.nfs.enabled
+    path       = var.nfs.enabled ? module.nfs-server[0].nfs_export_path : null
+    host       = var.nfs.enabled ? module.nfs-server[0].nfs_server_internal_ip : null
+    mount_path = var.nfs.enabled ? var.nfs.mount_path : null
+  }
+
   shared_memory_size_gibibytes = var.slurm_shared_memory_size_gibibytes
 
   nccl_topology_type           = var.slurm_nodeset_workers[0].resource.platform == "gpu-h100-sxm" ? "H100 GPU cluster" : "auto"

soperator/installations/example/terraform.tfvars

@@ -94,6 +94,20 @@ filestore_accounting = {
 
 # endregion Storage
 
+# region nfs-server
+
+# nfs = {
+#   enabled = true
+#   size_gibibytes = 93
+#   mount_path     = "/mnt/nfs"
+#   resource = {
+#     platform = "cpu-e2"
+#     preset   = "16vcpu-64gb"
+#    }
+# }
+
+# endregion nfs-server
+
 #----------------------------------------------------------------------------------------------------------------------#
 #                                                                                                                      #
 #                                                                                                                      #

soperator/installations/example/variables.tf

@@ -127,6 +127,35 @@ variable "filestore_accounting" {
 
 # endregion Storage
 
+# region nfs-server
+
+variable "nfs" {
+  type = object({
+    enabled        = bool
+    size_gibibytes = number
+    mount_path     = optional(string, "/mnt/nfs")
+    resource = object({
+      platform = string
+      preset   = string
+    })
+  })
+  default = {
+    enabled        = false
+    size_gibibytes = 93
+    resource = {
+      platform = "cpu-e2"
+      preset   = "16vcpu-64gb"
+    }
+  }
+
+  validation {
+    condition     = var.nfs.enabled ? var.nfs.size_gibibytes % 93 == 0 && var.nfs.size_gibibytes <= 262074 : true
+    error_message = "NFS size must be a multiple of 93 GiB and maximum value is 262074 GiB"
+  }
+}
+
+# endregion nfs-server
+
 # region k8s
 
 variable "k8s_version" {

soperator/modules/slurm/main.tf

@@ -165,6 +165,8 @@ resource "helm_release" "slurm_cluster" {
       mount_path = submount.mount_path
     }]
 
+    nfs = var.nfs
+
     nccl_topology_type = var.nccl_topology_type
     nccl_benchmark = {
       enable         = var.nccl_benchmark_enable

soperator/modules/slurm/templates/helm_values/slurm_cluster.yaml.tftpl

@@ -108,6 +108,14 @@ volumeSources:
     emptyDir:
       sizeLimit: ${nodes.worker.resources.ephemeral_storage}Gi
 
+  %{~ if nfs.enabled ~}
+  - name: nfs
+    nfs:
+      path: ${nfs.path}
+      readOnly: false
+      server: ${nfs.host}
+  %{~ endif ~}
+
   %{~ for sub_mount in jail_submounts ~}
   - name: jail-submount-${sub_mount.name}
     persistentVolumeClaim:
@@ -214,8 +222,13 @@ slurmNodes:
       spool:
         volumeClaimTemplateSpec: null
         volumeSourceName: worker-spool
-      %{~ if length(jail_submounts) > 0 ~}
+      %{~ if length(jail_submounts) > 0 || nfs.enabled ~}
       jailSubMounts:
+        %{~ if nfs.enabled ~}
+        - mountPath: ${nfs.mount_path}
+          name: nfs
+          volumeSourceName: nfs
+        %{~ endif ~}
         %{~ for sub_mount in jail_submounts ~}
         - name: ${sub_mount.name}
           mountPath: ${sub_mount.mount_path}
@@ -241,7 +254,6 @@ slurmNodes:
       - ${key}
       %{~ endfor ~}
     %{~ endif ~}
-    %{~ if length(jail_submounts) > 0 ~}
     sshd:
       resources:
         cpu: ${nodes.login.resources.cpu * 1000}m
@@ -252,8 +264,14 @@ slurmNodes:
         cpu: ${nodes.munge.resources.cpu * 1000}m
         memory: ${nodes.munge.resources.memory}Gi
         ephemeralStorage: ${nodes.munge.resources.ephemeral_storage}Gi
+    %{~ if length(jail_submounts) > 0 || nfs.enabled ~}
     volumes:
       jailSubMounts:
+        %{~ if nfs.enabled ~}
+        - mountPath: ${nfs.mount_path}
+          name: nfs
+          volumeSourceName: nfs
+        %{~ endif ~}
         %{~ for sub_mount in jail_submounts ~}
         - name: ${sub_mount.name}
           mountPath: ${sub_mount.mount_path}

soperator/modules/slurm/variables.tf

@@ -171,6 +171,27 @@ variable "filestores" {
 
 # endregion Filestore
 
+# region nfs-server
+
+variable "nfs" {
+  type = object({
+    enabled    = bool
+    mount_path = optional(string, "/mnt/nfs")
+    path       = optional(string)
+    host       = optional(string)
+  })
+  default = {
+    enabled = false
+  }
+
+  validation {
+    condition     = var.nfs.enabled ? var.nfs.path != null && var.nfs.host != null : true
+    error_message = "NFS path and host must be set."
+  }
+}
+
+# endregion nfs-server
+
 # region Config
 
 variable "shared_memory_size_gibibytes" {