Slurm Rebuild

.github/workflows/stackhpc.yml

@@ -178,12 +178,13 @@ jobs:
           ansible-playbook -v ansible/site.yml
           ansible-playbook -v ansible/ci/check_slurm.yml
 
-      - name: Test reimage of compute nodes and compute-init (via rebuild adhoc)
+      - name: Test compute node reboot and compute-init
         run: |
           . venv/bin/activate
           . environments/.stackhpc/activate
           ansible-playbook -v --limit compute ansible/adhoc/rebuild.yml
           ansible-playbook -v ansible/ci/check_slurm.yml
+          ansible-playbook -v ansible/adhoc/reboot_via_slurm.yml
 
       - name: Check sacct state survived reimage
         run: |

ansible/.gitignore

@@ -80,3 +80,5 @@ roles/*
 !roles/slurm_stats/**
 !roles/pytools/
 !roles/pytools/**
+!roles/rebuild/
+!roles/rebuild/**

ansible/adhoc/reboot_via_slurm.yml

@@ -0,0 +1,24 @@
+# Reboot compute nodes via slurm. Nodes will be rebuilt if `image_id` in inventory is different to the currently-provisioned image.
+# Example:
+#   ansible-playbook -v ansible/adhoc/reboot_via_slurm.yml
+
+- hosts: login
+  run_once: true
+  become: yes
+  gather_facts: no
+  tasks:
+    - name: Submit a Slurm job to reboot compute nodes
+      ansible.builtin.shell: |
+        set -e
+        srun --reboot -N 2 uptime
+      become_user: root
+      register: slurm_result
+      failed_when: slurm_result.rc != 0
+
+    - name: Fetch Slurm controller logs if reboot fails
+      ansible.builtin.shell: |
+        journalctl -u slurmctld --since "10 minutes ago" | tail -n 50
+      become_user: root
+      register: slurm_logs
+      when: slurm_result.rc != 0
+      delegate_to: "{{ groups['control'] | first }}"

ansible/roles/compute_init/README.md

@@ -1,11 +1,104 @@
-# EXPERIMENTAL: compute-init
-
-Experimental / in-progress functionality to allow compute nodes to rejoin the
-cluster after a reboot.
-
-To enable this add compute nodes (or a subset of them into) the `compute_init`
-group.
-
+# EXPERIMENTAL: compute_init
+
+Experimental functionality to allow compute nodes to rejoin the cluster after
+a reboot without running the `ansible/site.yml` playbook.
+
+To enable this:
+1. Add the `compute` group (or a subset) into the `compute_init` group. This is
+   the default when using cookiecutter to create an environment, via the
+   "everything" template.
+2. Build an image which includes the `compute_init` group. This is the case
+   for StackHPC-built release images.
+3. Enable the required functionalities during boot, by setting the
+   `compute_init_enable` property for a compute group in the
+   OpenTofu `compute` variable to a list which includes "compute", plus the
+   other roles/functionalities required, e.g.:
+
+   ```terraform
+   ...
+   compute = {
+      general = {
+         nodes = ["general-0", "general-1"]
+         compute_init_enable = ["compute", ... ] # see below
+      }
+   }
+   ...
+   ```
+
+## Supported appliance functionalities
+
+The string "compute" must be present in the `compute_init_enable` flag to enable
+this functionality. The table below shows which other appliance functionalities
+are currently supported - use the name in the role column to enable these.
+
+| Playbook                 | Role (or functionality) | Support         |
+| -------------------------|-------------------------|-----------------|
+| hooks/pre.yml            | ?                       | None at present |
+| validate.yml             | n/a                     | Not relevant during boot |
+| bootstrap.yml            | (wait for ansible-init) | Not relevant during boot |
+| bootstrap.yml            | resolv_conf             | Fully supported |
+| bootstrap.yml            | etc_hosts               | Fully supported |
+| bootstrap.yml            | proxy                   | None at present |
+| bootstrap.yml            | (/etc permissions)      | None required - use image build |
+| bootstrap.yml            | (ssh /home fix)         | None required - use image build |
+| bootstrap.yml            | (system users)          | None required - use image build |
+| bootstrap.yml            | systemd                 | None required - use image build |
+| bootstrap.yml            | selinux                 | None required - use image build |
+| bootstrap.yml            | sshd                    | None at present |
+| bootstrap.yml            | dnf_repos               | None at present (requirement TBD) |
+| bootstrap.yml            | squid                   | Not relevant for compute nodes |
+| bootstrap.yml            | tuned                   | None |
+| bootstrap.yml            | freeipa_server          | Not relevant for compute nodes |
+| bootstrap.yml            | cockpit                 | None required - use image build |
+| bootstrap.yml            | firewalld               | Not relevant for compute nodes |
+| bootstrap.yml            | fail2ban                | Not relevant for compute nodes |
+| bootstrap.yml            | podman                  | Not relevant for compute nodes |
+| bootstrap.yml            | update                  | Not relevant during boot |
+| bootstrap.yml            | reboot                  | Not relevant for compute nodes |
+| bootstrap.yml            | ofed                    | Not relevant during boot |
+| bootstrap.yml            | ansible_init (install)  | Not relevant during boot |
+| bootstrap.yml            | k3s (install)           | Not relevant during boot |
+| hooks/post-bootstrap.yml | ?                       | None at present |
+| iam.yml                  | freeipa_client          | None at present [1] |
+| iam.yml                  | freeipa_server          | Not relevant for compute nodes |
+| iam.yml                  | sssd                    | None at present |
+| filesystems.yml          | block_devices           | None required - role deprecated |
+| filesystems.yml          | nfs                     | All client functionality |
+| filesystems.yml          | manila                  | All functionality |
+| filesystems.yml          | lustre                  | None at present |
+| extras.yml               | basic_users             | All functionality [2] |
+| extras.yml               | eessi                   | All functionality [3] |
+| extras.yml               | cuda                    | None required - use image build [4] |
+| extras.yml               | persist_hostkeys        | Not expected to be required for compute nodes |
+| extras.yml               | compute_init (export)   | Not relevant for compute nodes |
+| extras.yml               | k9s (install)           | Not relevant during boot |
+| extras.yml               | extra_packages          | None at present. Would require dnf_repos |
+| slurm.yml                | mysql                   | Not relevant for compute nodes |
+| slurm.yml                | rebuild                 | Not relevant for compute nodes |
+| slurm.yml                | openhpc [5]             | All slurmd-related functionality |
+| slurm.yml                | (set memory limits)     | None at present |
+| slurm.yml                | (block ssh)             | None at present |
+| portal.yml               | (openondemand server)   | Not relevant for compute nodes |
+| portal.yml               | (openondemand vnc desktop) | None required - use image build |
+| portal.yml               | (openondemand jupyter server) | None required - use image build |
+| monitoring.yml           | (all monitoring)        | None at present [6] |
+| disable-repos.yml        | dnf_repos               | None at present (requirement TBD) |
+| hooks/post.yml           | ?                       | None at present |
+
+
+Notes:
+1. FreeIPA client functionality would be better provided using a client fork
+   which uses pkinit keys rather than OTP to reenrol nodes.
+2. Assumes home directory already exists on shared storage.
+3. Assumes `cvmfs_config` is the same on control node and all compute nodes
+4. If `cuda` role was run during build, the nvidia-persistenced is enabled
+  and will start during boot.
+5. `openhpc` does not need to be added to `compute_init_enable`, this is
+   automatically enabled by adding `compute`.
+5. Only node-exporter tasks are relevant, and will be done via k3s in a future release.
+
+
+## Approach
 This works as follows:
 1. During image build, an ansible-init playbook and supporting files
 (e.g. templates, filters, etc) are installed.
@@ -31,21 +124,7 @@ The check in 4b. above is what prevents the compute-init script from trying
 to configure the node before the services on the control node are available
 (which requires running the site.yml playbook).
 
-The following roles/groups are currently fully functional:
-- `resolv_conf`: all functionality
-- `etc_hosts`: all functionality
-- `nfs`: client functionality only
-- `manila`: all functionality
-- `basic_users`: all functionality, assumes home directory already exists on
-  shared storage
-- `eessi`: all functionality, assumes `cvmfs_config` is the same on control
-  node and all compute nodes.
-- `openhpc`: all functionality
-
-The above may be enabled by setting the compute_init_enable property on the
-tofu compute variable.
-
-# Development/debugging
+## Development/debugging
 
 To develop/debug changes to the compute script without actually having to build
 a new image:
@@ -83,7 +162,7 @@ reimage the compute node(s) first as in step 2 and/or add additional metadata
 as in step 3.
 
 
-# Design notes
+## Design notes
 - Duplicating code in roles into the `compute-init` script is unfortunate, but
   does allow developing this functionality without wider changes to the
   appliance.

ansible/roles/compute_init/tasks/export.yml

@@ -2,9 +2,9 @@
   file:
     path: /exports/cluster
     state: directory
-    owner: root
+    owner: slurm
     group: root
-    mode: u=rwX,go=
+    mode: u=rX,g=rwX,o=
   run_once: true
   delegate_to: "{{ groups['control'] | first }}"
 
@@ -23,21 +23,27 @@
   file:
     path: /exports/cluster/hostvars/{{ inventory_hostname }}/
     state: directory
-    mode: u=rwX,go=
-    # TODO: owner,mode,etc
+    owner: slurm
+    group: root
+    mode: u=rX,g=rwX,o=
   delegate_to: "{{ groups['control'] | first }}"
 
 - name: Template out hostvars
   template:
     src: hostvars.yml.j2
     dest: /exports/cluster/hostvars/{{ inventory_hostname }}/hostvars.yml
-    mode: u=rw,go=
+    owner: slurm
+    group: root
+    mode: u=r,g=rw,o=
   delegate_to: "{{ groups['control'] | first }}"
 
 - name: Copy manila share info to /exports/cluster
   copy:
     content: "{{ os_manila_mount_share_info_var | to_nice_yaml }}"
     dest: /exports/cluster/manila_share_info.yml
+    owner: root
+    group: root
+    mode: u=rw,g=r
   run_once: true
   delegate_to: "{{ groups['control'] | first }}"
   when: os_manila_mount_share_info is defined

ansible/roles/rebuild/README.md

@@ -0,0 +1,30 @@
+rebuild
+=========
+
+Enables reboot tool from https://github.com/stackhpc/slurm-openstack-tools.git to be run from control node.
+
+Requirements
+------------
+
+clouds.yaml file
+
+Role Variables
+--------------
+
+- `openhpc_rebuild_clouds`: Directory. Path to clouds.yaml file.
+
+
+Example Playbook
+----------------
+
+    - hosts: control
+      become: yes
+      tasks:
+        - import_role:
+            name: rebuild
+
+License
+-------
+
+Apache-2.0
+

ansible/roles/rebuild/defaults/main.yml

@@ -0,0 +1,2 @@
+---
+openhpc_rebuild_clouds: ~/.config/openstack/clouds.yaml

ansible/roles/rebuild/tasks/main.yml

@@ -0,0 +1,21 @@
+---
+
+- name: Create /etc/openstack
+  file:
+    path: /etc/openstack
+    state: directory
+    owner: slurm
+    group: root
+    mode: u=rX,g=rwX
+
+- name: Copy out clouds.yaml
+  copy:
+    src: "{{ openhpc_rebuild_clouds }}"
+    dest: /etc/openstack/clouds.yaml
+    owner: slurm
+    group: root
+    mode: u=r,g=rw
+
+- name: Setup slurm tools
+  include_role:
+    name: slurm_tools

ansible/roles/slurm_stats/README.md

@@ -21,7 +21,7 @@ Example Playbook
     - hosts: compute
       tasks:
         - import_role:
-            name: stackhpc.slurm_openstack_tools.slurm-stats
+            name: slurm_stats
 
 
 License

ansible/roles/slurm_tools/tasks/main.yml

@@ -27,7 +27,7 @@
   module_defaults:
     ansible.builtin.pip:
       virtualenv: /opt/slurm-tools
-      virtualenv_command: python3 -m venv
+      virtualenv_command: "{{ 'python3.9 -m venv' if ansible_distribution_major_version == '8' else 'python3 -m venv' }}"
       state: latest
   become: true
   become_user: "{{ pytools_user }}"

ansible/slurm.yml

@@ -9,6 +9,16 @@
     - include_role:
         name:  mysql
 
+- name: Setup slurm-driven rebuild
+  hosts: rebuild:!builder
+  become: yes
+  tags:
+    - rebuild
+    - openhpc
+  tasks:
+    - import_role:
+        name: rebuild
+
 - name: Setup slurm
   hosts: openhpc
   become: yes

environments/.stackhpc/inventory/extra_groups

@@ -1,10 +1,6 @@
 [basic_users:children]
 cluster
 
-[rebuild:children]
-control
-compute
-
 [etc_hosts:children]
 cluster
 
@@ -35,3 +31,6 @@ builder
 [sssd:children]
 # Install sssd into fat image
 builder
+
+[rebuild:children]
+control

environments/.stackhpc/tofu/SMS.tfvars

@@ -1,4 +1,8 @@
-cluster_net = "stackhpc-ipv4-geneve"
-cluster_subnet = "stackhpc-ipv4-geneve-subnet"
+cluster_networks = [
+    {
+        network = "stackhpc-ipv4-geneve"
+        subnet = "stackhpc-ipv4-geneve-subnet"
+    }
+]
 control_node_flavor = "general.v1.small"
 other_node_flavor = "general.v1.small"

environments/.stackhpc/tofu/main.tf

@@ -81,7 +81,7 @@ module "cluster" {
             nodes: ["compute-0", "compute-1"]
             flavor: var.other_node_flavor
             compute_init_enable: ["compute", "etc_hosts", "nfs", "basic_users", "eessi"]
-            # ignore_image_changes: true
+            ignore_image_changes: true
         }
         # Example of how to add another partition:
         # extra: {