HPE Cray provides Ansible plays and roles for software products deemed necessary for the system to function. Customers are free to write their own Ansible plays and roles to augment what HPE Cray provides or implement new features. Basic knowledge of Ansible is needed to write plays and roles. The information below includes recommendations and best practices for writing and running Ansible code on the system successfully with the Configuration Framework Service (CFS).
Help with Ansible can be found in the external Ansible documentation:
The Version Control Service (VCS), which is setup during the Cray System Management (CSM) product installation, is the appropriate place to store configuration content. Individual product installations include the Ansible code to properly configure each product.
The structure of the individual repository directories matches the recommended directory layout from the external Ansible documentation.
Playbooks are found at the top level, and the Ansible roles and variables are in their appropriate directories.
Inventory directories like group_vars
and host_vars
may exist, but they are empty and left for variable overrides and customizations by the user.
Within an Ansible playbook, users can designate which node groups the various tasks and roles will run against.
For example, a playbook may contain a list of roles that are run on only the Compute
nodes, or a list of roles that will run on only Application
nodes. This is designated using the hosts
parameter.
For example, hosts: Compute
would be used to target the compute nodes. Users can create additional sections that target other node types, or adjust the hosts that the included roles will run against as necessary.
It is also possible to target multiple groups within a section of a playbook, or to specify complex targets, such as nodes that are in one group and not in another group.
The syntax for this is available in the external Ansible documentation.
Hosts can also be in more than one group at a time. In this case, Ansible will run all sections where the node is in the targeted host group.
See the Ansible Inventory section for more information about groups that are made automatically available through CFS dynamic inventory.
CFS will handle scaling up Ansible to run on many hosts, but there are still places where performance can be improved by correctly writing Ansible plays.
See Create an Image Customization CFS Session for a definition of image customization.
cfs_image
host group to distinguish between image customization and node personalization.
This allows image customization to be identified in the hosts
parameter, removing the need to evaluate conditionals, and ensuring that tasks are not accidentally running in both modes needlessly.gather_facts: false
. If only a few facts are required, it is also possible to limit fact gathering by setting gather_subset
.
For more information on gather_subset
, see the external Ansible module setup documentation.group_by
and add_host
The group_by
and add_host
modules can both be used to dynamically generate new hosts groups for the Ansible inventory.
These modules prove useful when hosts can be grouped according to a common property. Then, plays can be designed to only target that particular group.
Examples of this might include grouping by operating system, hardware type, or a hardware property such as the presence of a GPU.
Ansible can then use these to skip roles and tasks more efficiently than if the when
conditional is applied.
group_by
should be used when there are multiple named groups by which hosts can be grouped.
For more information on using this module see See Ansible’s playbooks best practices guide.
Example of group_by
:
- name: group by OS
hosts: all
tasks:
- name: Classify hosts by OS
group_by:
key: os_{{ ansible_facts['distribution'] }}
- name: centOS playbook
hosts: os_CentOS
tasks:
...
add_host
is useful for cases where the property is true or false.
It allows users to create a new group consisting of only the hosts where the property is true.
See Ansible’s documentation on the add host module for more information.
Example of add_host
:
- name: group by a sample variable
hosts: all
tasks:
- name: Add all hosts where sample_var is true to the new Sample group
add_host:
name: '{{ inventory_hostname }}'
groups: sample_group
when: sample_var
- name: Sample playbook
hosts: sample_group
tasks:
...
To target only a subset of a set of nodes, plays should use the following syntax.
In this example the play is targeting only nodes in the sample group that are also in the Compute
nodes group.
&
takes the intersection of the Compute
and sample_group
groups.
hosts: Compute:&sample_group
To target a set of nodes except the ones in the new group, plays should use the following syntax.
In this example the play is targeting Compute
nodes that are not a part of the sample group.
!
negates the sample_group
group, so that only Compute nodes that are not an image are targeted.
hosts: Compute:!sample_group
include_*
Use include_*
(dynamic re-use) to skip multiple tasks at once when using conditionals. Ansible evaluates conditionals for every node in every task.
This includes when the conditional is applied to a block, or a role imported with roles:
or the import_role
task.
This is because these are static imports that are compiled at the beginning of the playbook, and the conditional is inherited by every task in the role or block.
Evaluating these conditionals for each task may only take a second or two, but across the hundreds of tasks that might be part of a playbook, this can add up to significant wasted time.
Instead use dynamic imports with the include_*
tasks. Because these are evaluated at runtime, a conditional can skip the import of the role or tasks entirely, and is only evaluated once.
For example, in the following case the role is imported statically and the when
statement will be propagated down and evaluated for each task in the role. This wastes time by running the same check many times.
- name: Sample playbook
hosts: all
roles:
- {role: sample_role, when: cray_cfs_image}
Instead the role should be imported dynamically so that the when
conditional is only evaluated once:
- name: Sample playbook
hosts: all
tasks:
- include_role:
role: sample_role
when: cray_cfs_image
See the Ansible documentation on Conditionals with re-use
and Re-using files and roles for more information.
(Dynamic re-use is not possible when importing playbooks, so instead consider using group_by
rather than a conditional static import.)
any_errors_fatal
. In addition to not working with all Ansible strategies, this can cause an Ansible run to exit early, and the nodes that did not have the error will have to start from the beginning of the playbook in the next session.free
Ansible strategy. This means avoiding situations where all nodes in a batch need to complete a task before moving onto the next, and can save time by allowing nodes to proceed through a playbook at their own pace.hosts
target that does
not match any nodes in the current inventory/limit, but when multiple types of nodes are configured at the same time with the same configuration, they may end up in the same batch and Ansible run. This would mean that Ansible
has to run through the sections for both types of nodes, taking more time than if the nodes were in separate batches and could skip past the unneeded code.For users just starting to write plays, or who just want to focus on the biggest improvements, here is a summary of the key tips:
gather_facts: false
group_by
, add_host
and include_*
rather than roles:
and when:
to avoid repeating conditionals.Because CFS splits components into multiple batches, and components may also configure at different times when they are rebooted, some keywords meant for coordinating the runs of multiple nodes may not work as expected.
Keyword | Notes |
---|---|
any_errors_fatal |
This keyword is intended to stop execution as soon as any node reports a failure. However, this will only stop execution for the current batch. |
run_once |
This keyword is intended to limit a task to running on a single node. However this will only cause the task to be run once per batch. |
serial |
This keyword is intended to limit runs to a small number of nodes at a time, such as during a rolling upgrade. |
However, this will only function within the batch, so more nodes may be running the task than intended when multiple batches are running. |
CFS supports two Ansible strategies: cfs_linear
and cfs_free
. cfs_linear
runs all task in a playbook serially, with all nodes completing a task before Ansible moves on to the next task.
cfs_free
decouples the nodes allowing each node to proceed through the playbook at its own pace.
Switching to cfs_free
from the default strategy of cfs_linear
may result in better configuration time; however, currently, not all included playbooks support the cfs_free
strategy,
so this should only be done for playbooks that are confirmed to work correctly with the cfs_free
strategy.
In addition, the cfs_free
strategy is limited by the fact that configuration in CFS is applied over multiple layers and multiple playbooks.
This means that even when using the cfs_free
strategy, all nodes must complete a playbook together before moving onto the next playbook.
The CFS Ansible strategies extend the similarly named Ansible strategy, adding reporting callbacks that are used to track components’ state. cfs_linear
and cfs_free
should always be used in place of linear
and free
to ensure that CFS functions correctly.