hpc-enterprise-slurm.yaml

# Copyright 2024 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.

---

blueprint_name: hpc-enterprise-slurm-v6

vars:
  project_id:  ## Set GCP Project ID Here ##
  deployment_name: hpc01
  region: us-central1
  zone: us-central1-a
  gpu_zones: [us-central1-a, us-central1-b, us-central1-c, us-central1-f]
  slurm_image:
    # Visit https://github.com/GoogleCloudPlatform/slurm-gcp/blob/master/docs/images.md#published-image-family
    # for a list of valid family options with Slurm
    family: slurm-gcp-6-8-hpc-rocky-linux-8
    project: schedmd-slurm-public
  # If image above is changed to use custom image, then setting below must be set to true
  instance_image_custom: false
  metadata: # Workaround for https://github.com/GoogleCloudPlatform/cluster-toolkit/discussions/3243
    VmDnsSetting: GlobalOnly

# Recommended to use GCS backend for Terraform state
# See https://github.com/GoogleCloudPlatform/hpc-toolkit/tree/main/examples#optional-setting-up-a-remote-terraform-state
#
# terraform_backend_defaults:
#  type: gcs
#  configuration:
#    bucket: <<BUCKET_NAME>>

# Documentation for each of the modules used below can be found at
# https://github.com/GoogleCloudPlatform/hpc-toolkit/blob/main/modules/README.md

deployment_groups:
- group: primary
  modules:
  # Source is an embedded module, denoted by "modules/*" without ./, ../, /
  # as a prefix. To refer to a local or community module, prefix with ./, ../ or /
  - id: network
    source: modules/network/vpc

  - id: private_service_access
    source: community/modules/network/private-service-access
    use: [network]

  - id: controller_sa
    source: community/modules/project/service-account
    settings:
      name: controller
      project_roles:
      - compute.instanceAdmin.v1
      - iam.serviceAccountUser
      - logging.logWriter
      - monitoring.metricWriter
      - pubsub.admin
      - storage.objectViewer

  - id: login_sa
    source: community/modules/project/service-account
    settings:
      name: login
      project_roles:
      - logging.logWriter
      - monitoring.metricWriter
      - storage.objectViewer

  - id: compute_sa
    source: community/modules/project/service-account
    settings:
      name: compute
      project_roles:
      - logging.logWriter
      - monitoring.metricWriter
      - storage.objectCreator

  - id: homefs
    source: modules/file-system/filestore
    use: [network, private_service_access]
    settings:
      local_mount: /home

  - id: projectsfs
    source: modules/file-system/filestore
    use: [network, private_service_access]
    settings:
      local_mount: /projects

  # This file system has an associated license cost.
  # https://console.developers.google.com/marketplace/product/ddnstorage/exascaler-cloud
  - id: scratchfs
    source: community/modules/file-system/DDN-EXAScaler
    use: [network]
    settings:
      local_mount: /scratch

  - id: n2_nodeset
    source: community/modules/compute/schedmd-slurm-gcp-v6-nodeset
    use: [network, compute_sa]
    settings:
      node_count_dynamic_max: 4
      machine_type: n2-standard-2
      instance_image: $(vars.slurm_image)
      enable_placement: false  # the default is: true
      allow_automatic_updates: false

  - id: n2_partition
    source: community/modules/compute/schedmd-slurm-gcp-v6-partition
    use: [n2_nodeset]
    settings:
      partition_name: n2
      exclusive: false  # allows nodes to stay up after jobs are done
      is_default: true
      partition_conf:
        SuspendTime: 300 # time (in secs) the nodes in this partition stay active after their tasks have completed

  - id: c2_nodeset
    source: community/modules/compute/schedmd-slurm-gcp-v6-nodeset
    use: [network, compute_sa]
    settings:
      node_count_dynamic_max: 20
      machine_type: c2-standard-60  # this is the default
      instance_image: $(vars.slurm_image)
      enable_placement: true
      bandwidth_tier: tier_1_enabled
      disk_type: pd-ssd
      disk_size_gb: 100
      allow_automatic_updates: false

  # use `-p c2` to submit jobs to this partition:
  # ex: `srun -p c2 -N 1 hostname`
  - id: c2_partition
    source: community/modules/compute/schedmd-slurm-gcp-v6-partition
    use: [c2_nodeset]
    settings:
      partition_name: c2
      # the following two are true by default
      exclusive: true  # this must be true if nodeset.enable_placement is true

  - id: c2d_nodeset
    source: community/modules/compute/schedmd-slurm-gcp-v6-nodeset
    use: [network, compute_sa]
    settings:
      node_count_dynamic_max: 20
      machine_type: c2d-standard-112
      instance_image: $(vars.slurm_image)
      bandwidth_tier: tier_1_enabled
      disk_type: pd-ssd
      disk_size_gb: 100
      allow_automatic_updates: false

  - id: c2d_partition
    source: community/modules/compute/schedmd-slurm-gcp-v6-partition
    use: [c2d_nodeset]
    settings:
      partition_name: c2d

  - id: c3_nodeset
    source: community/modules/compute/schedmd-slurm-gcp-v6-nodeset
    use: [network, compute_sa]
    settings:
      node_count_dynamic_max: 20
      machine_type: c3-highcpu-176
      instance_image: $(vars.slurm_image)
      bandwidth_tier: tier_1_enabled
      disk_type: pd-ssd
      disk_size_gb: 100
      allow_automatic_updates: false

  - id: c3_partition
    source: community/modules/compute/schedmd-slurm-gcp-v6-partition
    use: [c3_nodeset]
    settings:
      partition_name: c3

  - id: a2_8_nodeset
    source: community/modules/compute/schedmd-slurm-gcp-v6-nodeset
    use: [network, compute_sa]
    settings:
      node_count_dynamic_max: 16
      machine_type: a2-ultragpu-8g
      # This makes this nodeset look for machines in any of the following zones
      # https://github.com/GoogleCloudPlatform/hpc-toolkit/tree/develop/community/modules/compute/schedmd-slurm-gcp-v6-nodeset#compute-vm-zone-policies
      zones: $(vars.gpu_zones)
      bandwidth_tier: gvnic_enabled
      instance_image: $(vars.slurm_image)
      disk_type: pd-ssd
      disk_size_gb: 100
      node_conf:
        SocketsPerBoard: 2
        CoresPerSocket: 24
      allow_automatic_updates: false

  # use `-p a208` to submit jobs to this partition:
  # ex: `srun -p a208 --gpus-per-node=8 -N 1 nvidia-smi`
  - id: a2_8_partition
    source: community/modules/compute/schedmd-slurm-gcp-v6-partition
    use: [a2_8_nodeset]
    settings:
      partition_name: a208
      # The following allows users to use more host memory without specifying cpus on a job
      partition_conf:
        DefMemPerGPU: 160000
        DefMemPerCPU: null

  - id: a2_16_nodeset
    source: community/modules/compute/schedmd-slurm-gcp-v6-nodeset
    use: [network, compute_sa]
    settings:
      node_count_dynamic_max: 16
      machine_type: a2-megagpu-16g
      # This makes this nodeset look for machines in any of the following zones
      # https://github.com/GoogleCloudPlatform/hpc-toolkit/tree/develop/community/modules/compute/schedmd-slurm-gcp-v6-nodeset#compute-vm-zone-policies // !!!
      zones: $(vars.gpu_zones)
      bandwidth_tier: gvnic_enabled
      instance_image: $(vars.slurm_image)
      disk_type: pd-ssd
      disk_size_gb: 100
      node_conf:
        SocketsPerBoard: 2
        CoresPerSocket: 24
      allow_automatic_updates: false

  # use `-p a216` to submit jobs to this partition:
  # ex: `srun -p a216 --gpus-per-node=16 -N 1 nvidia-smi`
  - id: a2_16_partition
    source: community/modules/compute/schedmd-slurm-gcp-v6-partition
    use: [a2_16_nodeset]
    settings:
      partition_name: a216
      # The following allows users to use more host memory without specifying cpus on a job
      partition_conf:
        DefMemPerGPU: 160000
        DefMemPerCPU: null

  - id: h3_nodeset
    source: community/modules/compute/schedmd-slurm-gcp-v6-nodeset
    use: [network, compute_sa]
    settings:
      node_count_dynamic_max: 16
      # Note that H3 is available in only specific zones. https://cloud.google.com/compute/docs/regions-zones
      machine_type: h3-standard-88
      bandwidth_tier: gvnic_enabled  # https://cloud.google.com/compute/docs/compute-optimized-machines#h3_network
      instance_image: $(vars.slurm_image)
      # H3 does not support pd-ssd and pd-standard
      # https://cloud.google.com/compute/docs/compute-optimized-machines#h3_disks
      disk_type: pd-balanced
      disk_size_gb: 100
      allow_automatic_updates: false

  # use `-p h3` to submit jobs to this partition:
  # ex: `srun -p h3  -N 1 hostname`
  - id: h3_partition
    source: community/modules/compute/schedmd-slurm-gcp-v6-partition
    use: [h3_nodeset]
    settings:
      partition_name: h3

  - id: slurm_login
    source: community/modules/scheduler/schedmd-slurm-gcp-v6-login
    use: [network, login_sa]
    settings:
      instance_image: $(vars.slurm_image)
      machine_type: n2-standard-4
      # we recommend disabling public IPs if possible
      # but that requires your network to have a NAT or
      # private access configured
      enable_login_public_ips: true

  - id: slurm_controller
    source: community/modules/scheduler/schedmd-slurm-gcp-v6-controller
    use: [network, controller_sa, homefs, projectsfs, scratchfs, slurm_login,
          n2_partition, c2_partition, c2d_partition, c3_partition, a2_8_partition, a2_16_partition, h3_partition]
    settings:
      instance_image: $(vars.slurm_image)
      # the following allow for longer boot time
      # which is useful for large GPU nodes
      cloud_parameters:
        no_comma_params: false
        resume_rate: 0
        resume_timeout: 600
        suspend_rate: 0
        suspend_timeout: 600
      # we recommend disabling public IPs if possible
      # but that requires your network to have a NAT or
      # private access configured
      enable_controller_public_ips: true

  - id: hpc_dashboard
    source: modules/monitoring/dashboard
    outputs: [instructions]