From 6ec3a7317a4a3ec7a644b4aa91ca384546762e87 Mon Sep 17 00:00:00 2001
From: Steve Brasier <33413598+sjpb@users.noreply.github.com>
Date: Fri, 6 Sep 2024 09:52:44 +0100
Subject: [PATCH] Add RL9 cuda build variant (#428)

* determine cuda distro automatically

* fix typo in CUDA samples

* make facts available for cuda

* add RL9 cuda build variant

* fix typo in build definitions

* set packer build volume sizes depending on build variant

* fix volume size definition

* fix cuda verfsion to workaround issue with 12-6-0-1

* don't fail all builds if one fails

* bump CUDA builder disk size (build ran out of space)

* download cuda image to /mnt on gh runner

* download cuda image to /mnt on gh runner

* fix fatimage.yml mnt permissions

* Update main.yml

* switch to open nvidia drivers

* bump CI images

* make packer build volume-backed optional again

---------

Co-authored-by: bertiethorpe <bertie443@gmail.com>
Co-authored-by: bertiethorpe <84867280+bertiethorpe@users.noreply.github.com>
---
 .github/workflows/fatimage.yml               | 16 +++++++++-----
 ansible/extras.yml                           |  2 +-
 ansible/roles/cuda/defaults/main.yml         |  7 +++---
 ansible/roles/cuda/tasks/main.yml            | 13 ++---------
 environments/.stackhpc/ARCUS.pkrvars.hcl     |  3 ---
 environments/.stackhpc/LEAFCLOUD.pkrvars.hcl |  3 ---
 environments/.stackhpc/terraform/main.tf     |  4 ++--
 packer/openstack.pkr.hcl                     | 23 +++++++++++++++-----
 8 files changed, 37 insertions(+), 34 deletions(-)

diff --git a/.github/workflows/fatimage.yml b/.github/workflows/fatimage.yml
index 31fcc789a..7e2fc35b1 100644
--- a/.github/workflows/fatimage.yml
+++ b/.github/workflows/fatimage.yml
@@ -10,16 +10,20 @@ jobs:
     name: openstack-imagebuild
     runs-on: ubuntu-22.04
     strategy:
-      matrix:
+      fail-fast: false # allow other matrix jobs to continue even if one fails
+      matrix: # build RL8, RL9+OFED, RL9+CUDA versions
         os_version:
           - RL8
           - RL9
         build:
           - openstack.openhpc
           - openstack.openhpc-ofed
+          - openstack.openhpc-cuda
         exclude:
           - os_version: RL8
             build: openstack.openhpc-ofed
+          - os_version: RL8
+            build: openstack.openhpc-cuda
           - os_version: RL9
             build: openstack.openhpc
     env:
@@ -81,7 +85,9 @@ jobs:
       - name: Download image
         run: |
           . venv/bin/activate
-          openstack image save --file ${{ steps.manifest.outputs.image-name }}.qcow2 ${{ steps.manifest.outputs.image-name }}
+          sudo mkdir /mnt/images
+          sudo chmod 777 /mnt/images
+          openstack image save --file /mnt/images/${{ steps.manifest.outputs.image-name }}.qcow2 ${{ steps.manifest.outputs.image-name }}
 
       - name: Set up QEMU
         uses: docker/setup-qemu-action@v3
@@ -95,13 +101,13 @@ jobs:
         run: sudo mkdir -p './${{ steps.manifest.outputs.image-name }}'
 
       - name: mount qcow2 file
-        run: sudo guestmount -a ${{ steps.manifest.outputs.image-name }}.qcow2 -i --ro -o allow_other './${{ steps.manifest.outputs.image-name }}'
+        run: sudo guestmount -a /mnt/images/${{ steps.manifest.outputs.image-name }}.qcow2 -i --ro -o allow_other './${{ steps.manifest.outputs.image-name }}'
 
       - name: Run Trivy vulnerability scanner
         uses: aquasecurity/trivy-action@0.17.0
         with:
           scan-type: fs
-          scan-ref: "./${{ steps.manifest.outputs.image-name }}"
+          scan-ref: "${{ steps.manifest.outputs.image-name }}"
           scanners: "vuln"
           format: sarif
           output: "${{ steps.manifest.outputs.image-name }}.sarif"
@@ -117,7 +123,7 @@ jobs:
         uses: aquasecurity/trivy-action@0.16.1
         with:
           scan-type: fs
-          scan-ref: "./${{ steps.manifest.outputs.image-name }}"
+          scan-ref: "${{ steps.manifest.outputs.image-name }}"
           scanners: "vuln"
           format: table
           exit-code: '1'
diff --git a/ansible/extras.yml b/ansible/extras.yml
index 445a0cc16..c32f51c32 100644
--- a/ansible/extras.yml
+++ b/ansible/extras.yml
@@ -21,7 +21,7 @@
 - name: Setup CUDA
   hosts: cuda
   become: yes
-  gather_facts: no
+  gather_facts: yes
   tags: cuda
   tasks:
     - import_role:
diff --git a/ansible/roles/cuda/defaults/main.yml b/ansible/roles/cuda/defaults/main.yml
index 6b377a10b..33a25d9b4 100644
--- a/ansible/roles/cuda/defaults/main.yml
+++ b/ansible/roles/cuda/defaults/main.yml
@@ -1,11 +1,12 @@
-cuda_distro: rhel8
+cuda_distro: "rhel{{ ansible_distribution_major_version }}"
 cuda_repo: "https://developer.download.nvidia.com/compute/cuda/repos/{{ cuda_distro }}/x86_64/cuda-{{ cuda_distro }}.repo"
 cuda_driver_stream: default
+cuda_package_version: 'latest'
 cuda_packages:
-  - cuda
+  - "cuda{{ ('-' + cuda_package_version) if cuda_package_version != 'latest' else '' }}"
   - nvidia-gds
 # _cuda_version_tuple: # discovered from installed package e.g. ('12', '1', '0')
-cuda_version_short: "{{ _cuda_version_tuple[0] }}.{{ cuda_version_tuple[1] }}"
+cuda_version_short: "{{ _cuda_version_tuple[0] }}.{{ _cuda_version_tuple[1] }}"
 cuda_samples_release_url: "https://github.com/NVIDIA/cuda-samples/archive/refs/tags/v{{ cuda_version_short }}.tar.gz"
 cuda_samples_path: "/home/{{ ansible_user }}/cuda_samples"
 cuda_samples_programs:
diff --git a/ansible/roles/cuda/tasks/main.yml b/ansible/roles/cuda/tasks/main.yml
index b323cfc04..22f8e9e8e 100644
--- a/ansible/roles/cuda/tasks/main.yml
+++ b/ansible/roles/cuda/tasks/main.yml
@@ -24,22 +24,13 @@
   failed_when: false
   register: _cuda_driver_module_enabled
 
-- name: List nvidia driver dnf module stream versions
-  shell:
-    cmd: dnf module list nvidia-driver | grep -oP "\d+-dkms" | sort -V
-  # Output of interest from command is something like (some whitespace removed):
-  # "nvidia-driver 418-dkms   default [d], fm, ks    Nvidia driver for 418-dkms branch "
-  changed_when: false
-  register: _cuda_driver_module_streams
-  when: "'No matching Modules to list' in _cuda_driver_module_enabled.stderr"
-
 - name: Enable nvidia driver module
-  ansible.builtin.command: "dnf module enable -y nvidia-driver:{{ _cuda_driver_module_streams.stdout_lines | last }}"
+  ansible.builtin.command: "dnf module enable -y nvidia-driver:open-dkms"
   register: _cuda_driver_module_enable
   when: "'No matching Modules to list' in _cuda_driver_module_enabled.stderr"
   changed_when: "'Nothing to do' not in _cuda_driver_module_enable.stdout"
 
-- name: Install nvidia drivers # TODO: make removal possible?
+- name: Install nvidia drivers
   ansible.builtin.command: dnf module install -y nvidia-driver
   register: _cuda_driver_install
   when: "'No matching Modules to list' in _cuda_driver_module_enabled.stderr"
diff --git a/environments/.stackhpc/ARCUS.pkrvars.hcl b/environments/.stackhpc/ARCUS.pkrvars.hcl
index c07717156..6fd80e7a6 100644
--- a/environments/.stackhpc/ARCUS.pkrvars.hcl
+++ b/environments/.stackhpc/ARCUS.pkrvars.hcl
@@ -1,7 +1,4 @@
 flavor = "vm.ska.cpu.general.small"
-use_blockstorage_volume = true
-volume_size = 15 # GB
-image_disk_format = "qcow2"
 networks = ["4b6b2722-ee5b-40ec-8e52-a6610e14cc51"] # portal-internal (DNS broken on ilab-60)
 ssh_keypair_name = "slurm-app-ci"
 ssh_private_key_file = "~/.ssh/id_rsa"
diff --git a/environments/.stackhpc/LEAFCLOUD.pkrvars.hcl b/environments/.stackhpc/LEAFCLOUD.pkrvars.hcl
index da2d96d38..5adf4199c 100644
--- a/environments/.stackhpc/LEAFCLOUD.pkrvars.hcl
+++ b/environments/.stackhpc/LEAFCLOUD.pkrvars.hcl
@@ -1,8 +1,5 @@
 flavor = "ec1.large"
-use_blockstorage_volume = true
-volume_size = 15 # GB
 volume_type = "unencrypted"
-image_disk_format = "qcow2"
 networks = ["909e49e8-6911-473a-bf88-0495ca63853c"] # slurmapp-ci
 ssh_keypair_name = "slurm-app-ci"
 ssh_private_key_file = "~/.ssh/id_rsa"
diff --git a/environments/.stackhpc/terraform/main.tf b/environments/.stackhpc/terraform/main.tf
index ac588930c..96e04538b 100644
--- a/environments/.stackhpc/terraform/main.tf
+++ b/environments/.stackhpc/terraform/main.tf
@@ -30,8 +30,8 @@ variable "cluster_image" {
     type = map(string)
     default = {
         # https://github.com/stackhpc/ansible-slurm-appliance/pull/413
-        RL8: "openhpc-RL8-240813-1317-1b370a36"
-        RL9: "openhpc-ofed-RL9-240813-1317-1b370a36"
+        RL8: "openhpc-RL8-240904-1509-1687368f"
+        RL9: "openhpc-ofed-RL9-240904-1509-1687368f"
     }
 }
 
diff --git a/packer/openstack.pkr.hcl b/packer/openstack.pkr.hcl
index bb6af1a38..5f66c0320 100644
--- a/packer/openstack.pkr.hcl
+++ b/packer/openstack.pkr.hcl
@@ -120,7 +120,7 @@ variable "manifest_output_path" {
 
 variable "use_blockstorage_volume" {
   type = bool
-  default = false
+  default = true
 }
 
 variable "volume_type" {
@@ -129,13 +129,18 @@ variable "volume_type" {
 }
 
 variable "volume_size" {
-  type = number
-  default = null # When not specified use the size of the builder instance root disk
+  type = map(number)
+  default = {
+    # fat image builds, GB:
+    openhpc = 15
+    openhpc-ofed = 15
+    openhpc-cuda = 30
+  }
 }
 
 variable "image_disk_format" {
   type = string
-  default = null # When not specified use the image default
+  default = "qcow2"
 }
 
 variable "metadata" {
@@ -150,6 +155,7 @@ variable "groups" {
     # fat image builds:
     openhpc = ["control", "compute", "login"]
     openhpc-ofed = ["control", "compute", "login", "ofed"]
+    openhpc-cuda = ["control", "compute", "login", "ofed", "cuda"]
   }
 }
 
@@ -158,11 +164,11 @@ source "openstack" "openhpc" {
   flavor = var.flavor
   use_blockstorage_volume = var.use_blockstorage_volume
   volume_type = var.volume_type
+  volume_size = var.volume_size[source.name]
   metadata = var.metadata
   networks = var.networks
   floating_ip_network = var.floating_ip_network
   security_groups = var.security_groups
-  volume_size = var.volume_size
   
   # Input image:
   source_image = "${var.source_image[var.os_version]}"
@@ -178,7 +184,7 @@ source "openstack" "openhpc" {
   ssh_bastion_private_key_file = var.ssh_bastion_private_key_file
   
   # Output image:
-  image_disk_format = var.image_disk_format
+  image_disk_format = "qcow2"
   image_visibility = var.image_visibility
   image_name = "${source.name}-${var.os_version}-${local.timestamp}-${substr(local.git_commit, 0, 8)}"
 }
@@ -195,6 +201,11 @@ build {
     name = "openhpc-ofed"
   }
 
+  # CUDA fat image:
+  source "source.openstack.openhpc" {
+    name = "openhpc-cuda"
+  }
+
   # Extended site-specific image, built on fat image:
   source "source.openstack.openhpc" {
     name = "openhpc-extra"