Skip to content
master
Switch branches/tags
Code

Slurm on Google Cloud Platform

The following describes setting up a Slurm cluster using Google Cloud Platform, bursting out from an on-premise cluster to nodes in Google Cloud Platform and setting a multi-cluster/federated setup with a cluster that resides in Google Cloud Platform.

The supplied scripts can be modified to work with your environment.

SchedMD provides professional services to help you get up and running in the cloud environment. SchedMD Commercial Support

Issues and/or enhancement requests can be submitted to SchedMD's Bugzilla.

Also, join comunity discussions on either the Slurm User mailing list or the Google Cloud & Slurm Community Discussion Group.

Contents

Stand-alone Cluster in Google Cloud Platform

The supplied scripts can be used to create a stand-alone cluster in Google Cloud Platform. The scripts setup the following scenario:

  • 1 - controller node
  • N - login nodes
  • Multiple partitions with their own machine type, gpu type/count, disk size, disk type, cpu platform, and maximum node count.

Instances are created from images with Slurm and dependencies preinstalled. The default, schedmd-slurm-public/schedmd-slurm-20-11-4-hpc-centos-7, is based on the Google-provided HPC-optimized CentOS 7 image.

By default, /apps and /home are mounted from the controller across all instances in the cluster. These can be overwritten, and any other controller paths or external mounts can be added.

Install using GCP Marketplace

See the following page for Marketplace instructions.

Install using Terraform

To deploy, you must have a GCP account and either have the GCP Cloud SDK and Terraform installed on your computer or use the GCP Cloud Shell.

Steps:

  1. cd to tf/examples/basic

  2. Copy basic.tfvars.example to basic.tfvars

  3. Edit basic.tfvars with the required configuration
    See the tf/examples/basic/io.tf file for more detailed information on available configuration options.

  4. Deploy the cluster

    $ terraform init
    $ terraform apply -var-file=basic.tfvars
    
  5. Tearing down the cluster

    $ terraform destroy -var-file=basic.tfvars
    

    NOTE: If additional resources (instances, networks) are created other than the ones created from the default deployment then they will need to be destroyed before deployment can be removed. This includes bursted instances that Slurm has not yet suspended.

Defining network storage mounts

There are 3 types of network storage sections that can be provided to the TF modules: network_storage, login_network_storage, and partitions[].network_storage.

  • network_storage is mounted on all instances in the cluster.
  • login_network_storage is mounted on the controller and all login nodes.
  • partitions[].network_storage is mounted on compute instances within the specified partition.

All of these have the same 5 fields:

  • server_ip
  • remote_mount
  • local_mount
  • fs_type
  • mount_options

server_ip has one special value: $controller. This indicates that the mount is on the controller, so the remote_mount path will be exported, and the server_ip will be replaced with the correct hostname so all other instances can properly access the mount.

fs_type can be one of: nfs, cifs, lustre, gcsfuse

Public Slurm images

There are currently 3 public image families available for use with Slurm-GCP:
projects/schedmd-slurm-public/global/images/family/

  • schedmd-slurm-20-11-4-hpc-centos-7
  • schedmd-slurm-20-11-4-centos-7
  • schedmd-slurm-20-11-4-debian-10

Hyperthreads

For now, hyperthreading is either enabled or disabled in the image. Slurm-GCP must know this for each compute node type when configuring the cluster so it can configure the correct number of CPUs. image_hyperthreads must be set on the partition definition to reflect the state of hyperthreads in the image. If image_hyperthreads is set to true, and the image does not have hyperthreads enabled, the compute nodes will fail to report to Slurm when created. The hpc-centos-7 image has hyperthreads disabled.
NOTE: The result of disabling hyperthreads is that half the number of CPUs will be usable, eg. c2-standard-4 compute nodes will have 2 CPUs.

Preinstalled modules: OpenMPI

OpenMPI has been compiled to work with Slurm's srun. e.g.

$ module load openmpi
$ which mpicc
/apps/ompi/v4.1.x/bin/mpicc
$ mpicc mpi_hello_world.c
$ srun -N4 a.out
Hello world from processor g1-compute-0-0, rank 0 out of 4 processors
Hello world from processor g1-compute-0-3, rank 3 out of 4 processors
Hello world from processor g1-compute-0-1, rank 1 out of 4 processors
Hello world from processor g1-compute-0-2, rank 2 out of 4 processors

Installing Custom Packages

There are two files, custom-controller-install and custom-compute-install, in the scripts directory that can be used to add custom installations for the given instance type. The files will be executed during startup of the instance types.

Since the custom install scripts must be run when starting bursted nodes, long-running customizations should be added in a custom image instead.

Accessing Compute Nodes Directly

There are multiple ways to connect to the compute nodes:

  1. If the compute nodes have external IPs you can connect directly to the compute nodes. From the VM Instances page, the SSH drop down next to the compute instances gives several options for connecting to the compute nodes.
  2. With IAP configured, you can SSH to the nodes regardless of external IPs or not. See https://cloud.google.com/iap/docs/enabling-compute-howto.
  3. Use Slurm to get an allocation on the nodes.
    $ srun --pty $SHELL
    [g1-login0 ~]$ srun --pty $SHELL
    [g1-compute-0-0 ~]$
    

OS Login

By default, all instances are configured with OS Login.

OS Login lets you use Compute Engine IAM roles to manage SSH access to Linux instances and is an alternative to manually managing instance access by adding and removing SSH keys in metadata. https://cloud.google.com/compute/docs/instances/managing-instance-access

This allows user uid and gids to be consistent across all instances.

When sharing a cluster with non-admin users, the following IAM rules are recommended:

  1. Create a group for all users in admin.google.com.
  2. At the project level in IAM, grant the Compute Viewer and Service Account User roles to the group.
  3. At the instance level for each login node, grant the Compute OS Login role to the group.
    1. Make sure the Info Panel is shown on the right.
    2. On the compute instances page, select the boxes to the left of the login nodes.
    3. Click Add Members and add the Compute OS Login role to the group.
  4. At the organization level, grant the Compute OS Login External User role to the group if the users are not part of the organization.
  5. To allow ssh to login nodes without external IPs, configure IAP for the group.
    1. Go to the Identity-Aware Proxy page
    2. Select project
    3. Click SSH AND TCP RESOURCES tab
    4. Select boxes for login nodes
    5. Add group as a member with the IAP-secured Tunnel User role
    6. Reference: https://cloud.google.com/iap/docs/enabling-compute-howto

This allows users to access the cluster only through the login nodes.

Preemptible VMs

With preemptible_bursting on, when a node is found preempted, or stopped, the slurmsync script will mark the node as "down" and will attempt to restart the node. If there were any batch jobs on the preempted node, they will be requeued -- interactive (e.g. srun, salloc) jobs can't be requeued.

Hybrid Cluster for Bursting from On-Premise

Bursting out from an on-premise cluster is done by configuring the ResumeProgram and the SuspendProgram in the slurm.conf to resume.py, suspend.py in the scripts directory. config.yaml should be configured so that the scripts can create and destroy compute instances in a GCP project. See Cloud Scheduling Guide for more information.

Pre-reqs:

  1. VPN between on-premise and GCP
  2. bidirectional DNS between on-premise and GCP
  3. Open ports to on-premise
    1. slurmctld
    2. slurmdbd
    3. SrunPortRange
  4. Open ports in GCP for NFS from on-premise

Node Addressing

There are two options: 1) setup DNS between the on-premise network and the GCP network or 2) configure Slurm to use NodeAddr to communicate with cloud compute nodes. In the end, the slurmctld and any login nodes should be able to communicate with cloud compute nodes, and the cloud compute nodes should be able to communicate with the controller.

  • Configure DNS peering

    1. GCP instances need to be resolvable by name from the controller and any login nodes.
    2. The controller needs to be resolvable by name from GCP instances, or the controller ip address needs to be added to /etc/hosts. https://cloud.google.com/dns/zones/#peering-zones
  • Use IP addresses with NodeAddr

    1. disable cloud_dns in slurm.conf
    2. add SlurmctldParameters=cloud_reg_addrs in slurm.conf
    3. disable hierarchical communication in slurm.conf: TreeWidth=65533
    4. add controller's ip address to /etc/hosts on compute image

Configuration Steps

  1. Create a base instance

    Create a bare image and install and configure the packages (including Slurm) that you are used to for a Slurm compute node. Then create an image of the base image. It's recommended to create the image in a family.

  2. Create a service account and service account key that will have access to create and delete instances in the remote project.

  3. Install scripts

    Install the resume.py, suspend.py, slurmsync.py, util.py and config.yaml.example from the slurm-gcp repository's scripts directory to a location on the slurmctld. Rename config.yaml.example to config.yaml and modify the approriate values.

    Add the path of the service account key to google_app_cred_path in config.yaml.

    Add the image URL (path to the image or family) to each instance defintion.

  4. Modify slurm.conf:

    PrivateData=cloud
    
    SuspendProgram=/path/to/suspend.py
    ResumeProgram=/path/to/resume.py
    ResumeFailProgram=/path/to/suspend.py
    SuspendTimeout=600
    ResumeTimeout=600
    ResumeRate=0
    SuspendRate=0
    SuspendTime=300
    
    # Tell Slurm to not power off nodes. By default, it will want to power
    # everything off. SuspendExcParts will probably be the easiest one to use.
    #SuspendExcNodes=
    #SuspendExcParts=
    
    SchedulerParameters=salloc_wait_nodes
    SlurmctldParameters=cloud_dns,idle_on_node_suspend
    CommunicationParameters=NoAddrCache
    LaunchParameters=enable_nss_slurm
    
    SrunPortRange=60001-63000
    
  5. Add a cronjob/crontab to call slurmsync.py to be called by SlurmUser.

    e.g.

    */1 * * * * /path/to/slurmsync.py
    
  6. Test

    Try creating and deleting instances in GCP by calling the commands directly as SlurmUser.

    ./resume.py g1-compute-0-0
    ./suspend.py g1-compute-0-0
    

Users and Groups in a Hybrid Cluster

The simplest way to handle user synchronization in a hybrid cluster is to use nss_slurm. This permits passwd and group resolution for a job on the compute node to be serviced by the local slurmstepd process rather than some other network-based service. User information is sent from the controller for each job and served by the slurm step daemon. nss_slurm needs to be installed on the compute node image, which it is when the image is created with deployment manager or Terraform. For details on how to configure nss_slurm, see https://slurm.schedmd.com/nss_slurm.html.

Multi-Cluster / Federation

Slurm allows the use of a central SlurmDBD for multiple clusters. By doing this, it also allows the clusters to be able to communicate with each other. This is done by the client commands first checking with the SlurmDBD for the requested cluster's IP address and port which the client then uses to communicate directly with the cluster.

Some possible scenarios:

  • An on-premise cluster and a cluster in GCP sharing a single SlurmDBD.
  • An on-premise cluster and a cluster in GCP each with their own SlurmDBD but having each SlurmDBD know about each other using AccountingStorageExternalHost in each slurm.conf.

The following considerations are needed for these scenarios:

  • Regardless of location for the SlurmDBD, both clusters need to be able to talk to the each SlurmDBD and controller.
    • A VPN is recommended for traffic between on-premise and the cloud.
  • In order for interactive jobs (srun, salloc) to work from the login nodes to each cluster, the compute nodes must be accessible from the login nodes on each cluster.
    • It may be easier to only support batch jobs between clusters.
      • Once a batch job is on a cluster, srun functions normally.
  • If a firewall exists, srun communications most likely need to be allowed through it. Configure SrunPortRange to define a range for ports for srun communications.
  • Consider how to present file systems and data movement between clusters.
  • NOTE: All clusters attached to a single SlurmDBD must share the same user space (e.g. same uids across all the clusters).
  • NOTE: Either all clusters and the SlurmDBD must share the same MUNGE key or use a separate MUNGE key for each cluster and another key for use between each cluster and the SlurmDBD. In order for cross-cluster interactive jobs to work, the clusters must share the same MUNGE key. See the following for more information:
    Multi-Cluster Operation
    Accounting and Resource Limits

For more information see:
Multi-Cluster Operation
Federated Scheduling Guide

Troubleshooting

  1. Nodes aren't bursting?

    1. Check /var/log/slurm/resume.log for any errors
    2. Try creating nodes manually by calling resume.py manually as the "slurm" user.
      • NOTE: If you run resume.py manually with root, subsequent calls to resume.py by the "slurm" user may fail because resume.py's log file will be owned by root.
    3. Check the slurmctld logs
      • /var/log/slurm/slurmctld.log
      • Turn on the PowerSave debug flag to get more information. e.g.
        $ scontrol setdebugflags +powersave
        ...
        $ scontrol setdebugflags -powersave
        
  2. Cluster environment not fully coming up
    For example:

    • Slurm not being installed
    • Compute images never being stopped
    • etc.
    1. Check syslog (/var/log/messages) on instances for any errors. HINT: search for last mention of "startup-script."
  3. General debugging

    • check logs
    • check GCP quotas