Skip to content

Latest commit

 

History

History
 
 

schedmd-slurm-gcp-v5-controller

README.md

Description

Note

Slurm-gcp-v5-controller module is deprecated. See this update for specific recommendations and timelines.

This module creates a slurm controller node via the SchedMD/slurm-gcp slurm_controller_instance and slurm_instance_template modules.

More information about Slurm On GCP can be found at the project's GitHub page and in the Slurm on Google Cloud User Guide.

The user guide provides detailed instructions on customizing and enhancing the Slurm on GCP cluster as well as recommendations on configuring the controller for optimal performance at different scales.

Warning: The variables enable_reconfigure, enable_cleanup_compute, and enable_cleanup_subscriptions, if set to true, require additional dependencies to be installed on the system deploying the infrastructure.

# Install Python3 and run
pip3 install -r https://raw.githubusercontent.com/GoogleCloudPlatform/slurm-gcp/5.12.2/scripts/requirements.txt

Example

- id: slurm_controller
  source: community/modules/scheduler/schedmd-slurm-gcp-v5-controller
  use:
  - network1
  - homefs
  - compute_partition
  settings:
    machine_type: c2-standard-8

This creates a controller node with the following attributes:

  • connected to the primary subnetwork of network1
  • the filesystem with the ID homefs (defined elsewhere in the blueprint) mounted
  • One partition with the ID compute_partition (defined elsewhere in the blueprint)
  • machine type upgraded from the default c2-standard-4 to c2-standard-8

For a complete example using this module, see slurm-gcp-v5-cluster.yaml.

Live Cluster Reconfiguration (enable_reconfigure)

The schedmd-slurm-gcp-v5-controller module supports the reconfiguration of partitions and slurm configuration in a running, active cluster. This option is activated through the enable_reconfigure setting:

- id: slurm_controller
  source: community/modules/scheduler/schedmd-slurm-gcp-v5-controller
  settings:
    enable_reconfigure: true

To reconfigure a running cluster:

  1. Edit the blueprint with the desired configuration changes
  2. Call gcluster create <blueprint> -w to overwrite the deployment directory
  3. Follow instructions in terminal to deploy

The following are examples of updates that can be made to a running cluster:

  • Add or remove a partition to the cluster
  • Resize an existing partition
  • Attach new network storage to an existing partition

NOTE: Changing the VM machine_type of a partition may not work with enable_reconfigure. It is better to create a new partition and delete the old one.

This option has some additional requirements:

  • The Pub/Sub API must be activated in the target project: gcloud services enable pubsub.googleapis.com --project "<<PROJECT_ID>>"

  • The authenticated user in the local development environment (or where terraform apply is called) must have the Pub/Sub Admin (roles/pubsub.admin) IAM role.

  • Python and some python packages need to be installed with pip in the local development environment deploying the cluster. One can use following commands:

    pip3 install -r https://raw.githubusercontent.com/GoogleCloudPlatform/slurm-gcp/5.12.2/scripts/requirements.txt

    For more information, see the description of this module.

Custom Images

For more information on creating valid custom images for the controller VM instance or for custom instance templates, see our vm-images.md documentation page.

GPU Support

More information on GPU support in Slurm on GCP and other Cluster Toolkit modules can be found at docs/gpu-support.md

Placement Max Distance

When using enable_placement with Slurm, Google Compute Engine will attempt to place VMs as physically close together as possible. Capacity constraints at the time of VM creation may still force VMs to be spread across multiple racks. Google provides the max-distance flag which can used to control the maximum spreading allowed. Read more about max-distance in the official docs.

After deploying a Slurm cluster, you can use the following steps to manually configure the max-distance parameter.

  1. Make sure your blueprint has enable_placement: true setting for Slurm partitions.

  2. Deploy the Slurm cluster and wait for the deployment to complete.

  3. SSH to the deployed Slurm controller

  4. Apply the following edit to /slurm/scripts/config.yaml:

    # Replace
enable_slurm_gcp_plugins: false

# With
enable_slurm_gcp_plugins:
  max_hops:
    max_hops: 1

The max_hops parameter will be used for the max-distance argument. In the above case using a value of 1 will restrict VM to be placed on the same rack.

You can confirm that the `max-distance`` was applied by calling the following command while jobs are running:

gcloud beta compute resource-policies list \
  --format='yaml(name,groupPlacementPolicy.maxDistance)'

Warning

If a zone lacks capacity, using a lower max-distance value (such as 1) is more likely to cause VMs creation to fail.

Warning

/slurm/scripts/config.yaml will be overwritten if the blueprint is re-deployed using the enable_reconfigure flag.

Hybrid Slurm Clusters

For more information on how to configure an on premise slurm cluster with hybrid cloud partitions, see the schedmd-slurm-gcp-v5-hybrid module and our extended instructions in our docs.

Support

The Cluster Toolkit team maintains the wrapper around the slurm-on-gcp terraform modules. For support with the underlying modules, see the instructions in the slurm-gcp README.

License

Copyright 2023 Google LLC

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at

 http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.

Requirements

Name Version
terraform >= 1.1
google >= 3.83

Providers

Name Version
google >= 3.83

Modules

Name Source Version
slurm_controller_instance github.com/GoogleCloudPlatform/slurm-gcp.git//terraform/slurm_cluster/modules/slurm_controller_instance 5.12.2
slurm_controller_template github.com/GoogleCloudPlatform/slurm-gcp.git//terraform/slurm_cluster/modules/slurm_instance_template 5.12.2

Resources

Name Type
google_compute_default_service_account.default data source
google_compute_image.slurm data source

Inputs

Name Description Type Default Required
access_config Access configurations, i.e. IPs via which the VM instance can be accessed via the Internet. 
list(object({
    nat_ip       = string
    network_tier = string
  }))
 [] no
additional_disks List of maps of disks. 
list(object({
    disk_name    = string
    device_name  = string
    disk_type    = string
    disk_size_gb = number
    disk_labels  = map(string)
    auto_delete  = bool
    boot         = bool
  }))
 [] no
allow_automatic_updates If false, disables automatic system package updates on the created instances. This feature is
only available on supported images (or images derived from them). For more details, see
https://cloud.google.com/compute/docs/instances/create-hpc-vm#disable_automatic_updates		 bool true no
can_ip_forward Enable IP forwarding, for NAT instances for example. bool false no
cgroup_conf_tpl Slurm cgroup.conf template file path. string null no
cloud_parameters cloud.conf options. 
object({
    no_comma_params = bool
    resume_rate     = number
    resume_timeout  = number
    suspend_rate    = number
    suspend_timeout = number
  })
 
{
  "no_comma_params": false,
  "resume_rate": 0,
  "resume_timeout": 300,
  "suspend_rate": 0,
  "suspend_timeout": 300
}
 no
cloudsql Use this database instead of the one on the controller.
server_ip : Address of the database server.
user : The user to access the database as.
password : The password, given the user, to access the given database. (sensitive)
db_name : The database to access.		 
object({
    server_ip = string
    user      = string
    password  = string # sensitive
    db_name   = string
  })
 null no
compute_startup_script Startup script used by the compute VMs. string "" no
compute_startup_scripts_timeout The timeout (seconds) applied to the compute_startup_script. If
any script exceeds this timeout, then the instance setup process is considered
failed and handled accordingly.

NOTE: When set to 0, the timeout is considered infinite and thus disabled.		 number 300 no
controller_startup_script Startup script used by the controller VM. string "" no
controller_startup_scripts_timeout The timeout (seconds) applied to the controller_startup_script. If
any script exceeds this timeout, then the instance setup process is considered
failed and handled accordingly.

NOTE: When set to 0, the timeout is considered infinite and thus disabled.		 number 300 no
deployment_name Name of the deployment. string n/a yes
disable_controller_public_ips If set to false. The controller will have a random public IP assigned to it. Ignored if access_config is set. bool true no
disable_default_mounts Disable default global network storage from the controller
- /usr/local/etc/slurm
- /etc/munge
- /home
- /apps
Warning: If these are disabled, the slurm etc and munge dirs must be added
manually, or some other mechanism must be used to synchronize the slurm conf
files and the munge key across the cluster.		 bool false no
disable_smt Disables Simultaneous Multi-Threading (SMT) on instance. bool true no
disk_auto_delete Whether or not the boot disk should be auto-deleted. bool true no
disk_labels Labels specific to the boot disk. These will be merged with var.labels. map(string) {} no
disk_size_gb Boot disk size in GB. number 50 no
disk_type Boot disk type. string "pd-ssd" no
enable_bigquery_load Enable loading of cluster job usage into big query. bool false no
enable_cleanup_compute Enables automatic cleanup of compute nodes and resource policies (e.g.
placement groups) managed by this module, when cluster is destroyed.

NOTE: Requires Python and pip packages listed at the following link:
https://github.com/GoogleCloudPlatform/slurm-gcp/blob/3979e81fc5e4f021b5533a23baa474490f4f3614/scripts/requirements.txt

WARNING: Toggling this may impact the running workload. Deployed compute nodes
may be destroyed and their jobs will be requeued.		 bool false no
enable_cleanup_subscriptions Enables automatic cleanup of pub/sub subscriptions managed by this module, when
cluster is destroyed.

NOTE: Requires Python and pip packages listed at the following link:
https://github.com/GoogleCloudPlatform/slurm-gcp/blob/3979e81fc5e4f021b5533a23baa474490f4f3614/scripts/requirements.txt

WARNING: Toggling this may temporarily impact var.enable_reconfigure behavior.		 bool false no
enable_confidential_vm Enable the Confidential VM configuration. Note: the instance image must support option. bool false no
enable_devel Enables development mode. Not for production use. bool false no
enable_external_prolog_epilog Automatically enable a script that will execute prolog and epilog scripts
shared under /opt/apps from the controller to compute nodes.		 bool false no
enable_oslogin Enables Google Cloud os-login for user login and authentication for VMs.
See https://cloud.google.com/compute/docs/oslogin		 bool true no
enable_reconfigure Enables automatic Slurm reconfiguration when Slurm configuration changes (e.g.
slurm.conf.tpl, partition details). Compute instances and resource policies
(e.g. placement groups) will be destroyed to align with new configuration.
NOTE: Requires Python and Google Pub/Sub API.
WARNING: Toggling this will impact the running workload. Deployed compute nodes
will be destroyed and their jobs will be requeued.		 bool false no
enable_shielded_vm Enable the Shielded VM configuration. Note: the instance image must support option. bool false no
enable_slurm_gcp_plugins Enables calling hooks in scripts/slurm_gcp_plugins during cluster resume and suspend. any false no
epilog_scripts List of scripts to be used for Epilog. Programs for the slurmd to execute
on every node when a user's job completes.
See https://slurm.schedmd.com/slurm.conf.html#OPT_Epilog.		 
list(object({
    filename = string
    content  = string
  }))
 [] no
gpu DEPRECATED: use var.guest_accelerator 
object({
    type  = string
    count = number
  })
 null no
guest_accelerator List of the type and count of accelerator cards attached to the instance. 
list(object({
    type  = string,
    count = number
  }))
 [] no
instance_image Defines the image that will be used in the Slurm controller VM instance.

Expected Fields:
name: The name of the image. Mutually exclusive with family.
family: The image family to use. Mutually exclusive with name.
project: The project where the image is hosted.

For more information on creating custom images that comply with Slurm on GCP
see the "Slurm on GCP Custom Images" section in docs/vm-images.md.		 map(string) 
{
  "family": "slurm-gcp-5-12-hpc-centos-7",
  "project": "schedmd-slurm-public"
}
 no
instance_image_custom A flag that designates that the user is aware that they are requesting
to use a custom and potentially incompatible image for this Slurm on
GCP module.

If the field is set to false, only the compatible families and project
names will be accepted. The deployment will fail with any other image
family or name. If set to true, no checks will be done.

See: https://goo.gle/hpc-slurm-images		 bool false no
instance_template Self link to a custom instance template. If set, other VM definition
variables such as machine_type and instance_image will be ignored in favor
of the provided instance template.

For more information on creating custom images for the instance template
that comply with Slurm on GCP see the "Slurm on GCP Custom Images" section
in docs/vm-images.md.		 string null no
labels Labels, provided as a map. map(string) {} no
login_startup_scripts_timeout The timeout (seconds) applied to the login startup script. If
any script exceeds this timeout, then the instance setup process is considered
failed and handled accordingly.

NOTE: When set to 0, the timeout is considered infinite and thus disabled.		 number 300 no
machine_type Machine type to create. string "c2-standard-4" no
metadata Metadata, provided as a map. map(string) {} no
min_cpu_platform Specifies a minimum CPU platform. Applicable values are the friendly names of
CPU platforms, such as Intel Haswell or Intel Skylake. See the complete list:
https://cloud.google.com/compute/docs/instances/specify-min-cpu-platform		 string null no
network_ip DEPRECATED: Use static_ips variable to assign an internal static ip address. string null no
network_self_link Network to deploy to. Either network_self_link or subnetwork_self_link must be specified. string null no
network_storage An array of network attached storage mounts to be configured on all instances. 
list(object({
    server_ip             = string,
    remote_mount          = string,
    local_mount           = string,
    fs_type               = string,
    mount_options         = string,
    client_install_runner = map(string)
    mount_runner          = map(string)
  }))
 [] no
on_host_maintenance Instance availability Policy. string "MIGRATE" no
partition Cluster partitions as a list. 
list(object({
    compute_list = list(string)
    partition = object({
      enable_job_exclusive    = bool
      enable_placement_groups = bool
      network_storage = list(object({
        server_ip     = string
        remote_mount  = string
        local_mount   = string
        fs_type       = string
        mount_options = string
      }))
      partition_conf    = map(string)
      partition_feature = string
      partition_name    = string
      partition_nodes = map(object({
        access_config = list(object({
          network_tier = string
        }))
        bandwidth_tier         = string
        node_count_dynamic_max = number
        node_count_static      = number
        enable_spot_vm         = bool
        group_name             = string
        instance_template      = string
        maintenance_interval   = string
        node_conf              = map(string)
        reservation_name       = string
        spot_instance_config = object({
          termination_action = string
        })
      }))
      partition_startup_scripts_timeout = number
      subnetwork                        = string
      zone_policy_allow                 = list(string)
      zone_policy_deny                  = list(string)
      zone_target_shape                 = string
    })
  }))
 [] no
preemptible Allow the instance to be preempted. bool false no
project_id Project ID to create resources in. string n/a yes
prolog_scripts List of scripts to be used for Prolog. Programs for the slurmd to execute
whenever it is asked to run a job step from a new job allocation.
See https://slurm.schedmd.com/slurm.conf.html#OPT_Prolog.		 
list(object({
    filename = string
    content  = string
  }))
 [] no
region Region where the instances should be created. string null no
service_account Service account to attach to the controller instance. If not set, the
default compute service account for the given project will be used with the
"https://www.googleapis.com/auth/cloud-platform" scope.		 
object({
    email  = string
    scopes = set(string)
  })
 null no
shielded_instance_config Shielded VM configuration for the instance. Note: not used unless
enable_shielded_vm is 'true'.
enable_integrity_monitoring : Compare the most recent boot measurements to the
integrity policy baseline and return a pair of pass/fail results depending on
whether they match or not.
enable_secure_boot : Verify the digital signature of all boot components, and
halt the boot process if signature verification fails.
enable_vtpm : Use a virtualized trusted platform module, which is a
specialized computer chip you can use to encrypt objects like keys and
certificates.		 
object({
    enable_integrity_monitoring = bool
    enable_secure_boot          = bool
    enable_vtpm                 = bool
  })
 
{
  "enable_integrity_monitoring": true,
  "enable_secure_boot": true,
  "enable_vtpm": true
}
 no
slurm_cluster_name Cluster name, used for resource naming and slurm accounting. If not provided it will default to the first 8 characters of the deployment name (removing any invalid characters). string null no
slurm_conf_tpl Slurm slurm.conf template file path. string null no
slurmdbd_conf_tpl Slurm slurmdbd.conf template file path. string null no
source_image DEPRECATED: Use instance_image instead. string null no
source_image_family DEPRECATED: Use instance_image instead. string null no
source_image_project DEPRECATED: Use instance_image instead. string null no
static_ips List of static IPs for VM instances. list(string) [] no
subnetwork_project The project that subnetwork belongs to. string null no
subnetwork_self_link Subnet to deploy to. Either network_self_link or subnetwork_self_link must be specified. string null no
tags Network tag list. list(string) [] no
zone Zone where the instances should be created. If not specified, instances will be
spread across available zones in the region.		 string null no

Outputs

Name Description
cloud_logging_filter Cloud Logging filter to cluster errors.
controller_instance_id The server-assigned unique identifier of the controller compute instance.
pubsub_topic Cluster Pub/Sub topic.