Slurm cloud bursting

Slurm has the ability to support a cluster that grows and shrinks on demand, typically relying upon a service such as Amazon Elastic Computing Cloud (Amazon_EC2), Google Cloud Platform (GCP), Microsoft Azure, or Oracle Cloud Infrastructure (OCI).

Some Slurm cloud bursting resources:

Jump to our top-level Slurm page: Slurm batch queueing system

Slurm configuration for cloud nodes


Make sure to install these prerequisites before building Slurm RPM packages:

dnf install http-parser-devel json-c-devel

Check that slurmctld has JSON support:

strings `which slurmctld ` | grep HAVE_JSON

Configuring slurm.conf for power saving

According to the Slurm Power_Saving_Guide the following parameters in slurm.conf must be carefully configured, for example:

# Exceptions to the node suspend/resume logic (partitions):
# SuspendTime=3600   # Global value: the default is INFINITE.  DO NOT change the default!
TreeWidth=1000     # Or 65535, the default is 50.  Large values disable tree fanout of slurmd communications

Some important points about power saving parameters:

  • If you set SuspendTime to anything but INFINITE (or -1), power saving shutdown of nodes will commence!

  • It may be preferable to omit the global parameter and leave it with the default value SuspendTime=INFINITE. In stead define it only on any relevant partitions, for example:

    PartitionName=my_partition SuspendTime=3600
  • The SlurmctldParameters=idle_on_node_suspend causes nodes drained for maintenance purposes to become idle and available for running jobs. This is most likely not desirable.

  • An important side effect of power saving suspension of nodes pertains to on-premise nodes. Compute nodes that are, for example, drained for maintenance purposes will be suspended and later resumed when needed by jobs. This is highly undesirable!

    This issue has been resolved in Slurm 23.02 by bug_15184 which introduces a new slurm.conf parameter SuspendExcStates. This permits to configure node states which you want to be excluded from power saving suspension. Valid states for SuspendExcStates include:


    Beware: If any nodes in the cluster are running Slurm versions lower than 23.02, the slurmd will fail if the new SuspendExcStates parameter is specified!

Resume and Suspend scripts

Each site must develop their own ResumeProgram and SuspendProgram scripts. Remember that Slurm executes these scripts as the slurm user, which must therefore be able to power up/down cloud nodes!

In addition, a ResumeFailProgram script will be useful for handling nodes that do not resume (power up) correctly.

Logging from the ResumeProgram/SuspendProgram scripts must be programmed in the scripts. This example may be used:

echo "`date` User $USER invoked $action $0 $*" >>/var/log/slurm/power_save.log

See also some suspend/resume scripts in

Site-to-Site VPN connection

The local on-premise cluster network is the subnet (for example, 10.1.x.x/16) where all your compute nodes are connected together. Likewise, when you use an off-site cloud service, a number of (virtual) compute nodes are also connected together in another (virtual) subnet (for example,

When you want to connect your on-premise cluster to the cloud service compute nodes, you need to set up a virtual private network (VPN) which extends a private network across a public network and enables users to send and receive data across shared or public networks as if their computing devices were directly connected to the private network. Such a VPN connection is a prerequisite for accessing to the specific cloud services described below, and you will need to configure this early on.

For a Site-to-site VPN tunnel from a cloud service to the local on-premise network using a Linux server, see our page:

Network router devices can also be used, see the cloud service’s list of supported devices:

Resources for Slurm on Google Cloud Platform (GCP)

Resources for Slurm on Microsoft Azure

Login to Azure:

Other topics:

Azure subscription

Get started with the Azure EA portal:

To create a Microsoft Azure account related to an EA enrollment, then please follow these steps:

Azure CLI

The Azure command-line interface Azure_CLI is an environment to create and manage Azure resources. Azure_CLI is available across Azure services and is designed to get you working quickly with Azure, with an emphasis on automation. See:

Install Azure_CLI using dnf:

  • On RHEL import the Microsoft repository key:

    rpm --import
  • Create the file /etc/yum.repos.d/azure-cli.repo with contents:

    name=Azure CLI
  • Install the azure-cli package:

    dnf install azure-cli

Alternatively, the following Ansible role may be used:

- name: Create the Azure CLI YUM repository
    name: "azure-cli"
    description: "Azure CLI"
    baseurl: ""
    gpgkey: ""
    gpgcheck: yes
    enabled: yes
- name: Install the azure-cli RPM
    name: "azure-cli"
    state: present

Azure CLI login

The Linux command for Azure_CLI is:

az help        # Lists az subcommands
az --version   # Display CLI version

Append the --help option to display subcommand help, for example:

az account --help

To authenticate your CLI session with your existing Azure subscription, use the az_login command:

az login

If the CLI can open your default browser, it will do so and load an Azure sign-in page. Otherwise, open a browser page at and enter the authorization code displayed in your terminal.

Select your existing Azure subscription in the web page. Your Azure credentials will be stored in the folder $HOME/.azure/. None of your sign-in information is stored by the Azure_CLI. Instead, an authentication refresh token is generated by Azure and stored. If the token is revoked you get a message from the CLI saying you need to sign in again.

After signing in, CLI commands are run against your default subscription. If you have multiple subscriptions, you can change your default subscription.

List your Azure account information:

az account list --output table

If you want to log out from the Azure_CLI:

az logout

Virtual network in Azure

The Virtual Machine nodes should be connected to a shared virtual network (Azure_VNet) inside Azure so that a VPN subnet connection can be made later.

See the Quickstart: Create a virtual network using the Azure portal.

In Home > New > Create virtual network create a new Resource Group:

  • In Project details select an existing Resource group for your VMs, or create a new one.

  • Choose a suitable network name, for example VPNnetresourcegroup, and select the Azure region, for example West Europe.

  • Choose a sufficiently large IP address space, for example

  • Create a subnet, for example VMsubnet, and define smaller subnet rangewithin your IP address space, for example, leaving room for other subnets (the VPN gateways require their own separate subnet).

Then join all future VMs to Virtual network/subnet: VPNnetresourcegroup.

Currently existing VMs can unfortunately not be moved to a different virtual network, see However, with some tricks an existing VM can be recreated on a different virtual network, see

In conclusion, if a VM is connected to the wrong subnet, it is easier to delete the VM and create it again from scratch.

Virtual Network management

Create, change, or delete a virtual network including Change DNS servers.

In the Virtual Network you can specify the use of on-premise DNS servers in the Settings->DNS Servers menu by adding the IP-addresses of DNS servers:

  • Select Custom in stead of Default (Azure-provided) and reboot the VM.

VPN gateway to Azure

Use the Azure portal to create a Site-to-Site VPN gateway connection from your on-premises network to the Azure_VNet:

Do not choose the Basic gateway SKU because it does not support IKEv2, see vpngateways. You may select a better SKU from the list, for example VpnGw2 is recommended.

You may use Azure_CLI to list your Azure_VNet:

az network vnet list -o table

Configure VPN gateways

The VPN gateways configuration steps are as follows:

  1. Create an Azure Resource named Public IP address. You have to assign a name to this, for example PublicIPAddress. Write down the VPN gateway’s Public IP address.

  2. Create a route based Azure site-to-site VPN gateway for your selected Azure Virtual Network (Azure_VNet), for example VPNnet:

    • Do not choose the Basic gateway SKU because it does not support IKEv2, see vpngateways.

    • You may select a better SKU from the list, for example VpnGw2 on Generation1.

    • Create a new IP address and use the Public IP address assigned above.

    • Deployment may take 20 minutes or more!

  3. Create a new resource, search for Local network gateway:

    • Assign a public IP address for the on-premise local network VPN gateway.

    • Create the new Local network gateway with the local IP address.

    • For Address space use the local cluster network’s subnet which will contain the local VPN gateway router.

    • Select the VPN subnet resource group. Creation may take 20 minutes or more!

  4. Go to the VPN gateway page and click on Settings -> Connections.

    Click +Add to create a new named network connection, for example Azure2Local. You must select a Pre-shared key to be used both by the Azure and local VPN gateway routers.

You still need to configure a Local network gateway with the local public IP address which acts as a router to the local cluster network’s subnet.

Libreswan IPsec VPN router

For a Site-to-site VPN tunnel from Azure to the local on-premise network, a Libreswan IPsec VPN router can be used:

Additional Azure VPN gateway information

There are two distinct types of Azure VPN gateways:

  • Policy-based (static-routing) gateway.

  • Route-based (dynamic-routing) gateway.

Azure supports three types of Point-to-site VPN options:

  • Secure Socket Tunneling Protocol (SSTP). SSTP is a Microsoft proprietary SSL-based solution that can penetrate firewalls since most firewalls open the outbound TCP port that 443 SSL uses.

  • OpenVPN. OpenVPN is a SSL-based solution that can penetrate firewalls since most firewalls open the outbound TCP port that 443 SSL uses.

  • IKEv2 VPN. IKE (defined in RFC7296) VPN is a standards-based IPsec VPN solution that uses outbound UDP ports 500 and 4500 and IP protocol no. 50. Firewalls do not always open these ports, so there is a possibility of IKEv2 VPN not being able to traverse proxies and firewalls.

See the VPN Gateway FAQ about which type to configure. Note: Libreswan uses IKEv2.

Other resources are:

Create resources and machines in Azure home

Go to the Azure_home page and create a Resource_group with a descriptive name such as Group_compute_nodes.

Now you can create a new machine in the Virtual_machines page by clicking +Add:

  • Select the Resource_group created above.

  • Choose a machine name, for example node001.

  • Choose your region, for example West Europe.

  • Availability: no redundancy

  • Select a predefined Azure image, for example AlmaLinux-based 8. Display lists of available free Almalinux images:

    az vm image list --all -o table --publisher almalinux

    Display lists of available Rockylinux images:

    az vm image list --all -o table --publisher erockyenterprisesoftwarefoundationinc1653071250513
    az vm image list -f rocky --all -o table

    There is a free Rocky Linux 8 - Official image, see also the list of Rockylinux cloud images.

  • Azure spot instance: No

  • Size of the machine: Select from the list of available Azure machine_sizes in the Change size menu.

    For a small compute node the memory optimized Standard DS12 v2 (4 vcpus, 28 GiB memory) Intel Xeon Broadwell may be suitable. You may be hit by errors_for_resource_quotas.

  • Pricing of different VMs is listed in the Linux_Virtual_Machines_Pricing page.

  • For Administrator account:

    • For Authentication type the Password method may have to be used until we get documentation about using Linux OpenSSH public keys (only Putty_keys work).

      SSH public key for Azure is documented in

    • For the administrator username for the VM make a choice:


      which is not a regular user (since users will probably be created later) nor the root superuser (forbidden by Azure).

      The administrator user’s $HOME directory will be the OS standard /home. The home directory may later be changed inside the VM by the root user with usermod:

      root# usermod --home /var/lib/azureuser --move-home azureuser

      This will only work if azureuser is logged out and root has been logged in directly.

In the Next->Disks window select:

  • OS disk type: Standard HDD may suffice.

  • No data disks are needed because we will store user files in a separate Azure storage.

In the Next->Networking tab:

  • For the Network interface select the desired Virtual network (Azure_VNet) for VPN access (see above how to create a shared virtual network named VPNnetresourcegroup). It is very hard to change the virtual network once the VM has been created!

  • For Public IP select None because we will not use this.

  • Check the box Delete NIC when VM is deleted.

  • SSH (port 22) access should be enabled for remote management.

In the Next->Management window select:

  • Enable auto-shutdown every day at an appropriate time (choose your timezone).

In the Next->Management window the default values should be fine.

In the Next->Tags window select:

  • Tag the Name field to identify the group using this machine for billing purposes.

Finally do a Review and create. This will open a popup window:

Generate new key pair
Select: Download private key and create resource

You must save the VM’s SSH public key file <vm-name>_key.pem to disk and use as shown below to login.

Create and Manage Linux VMs with the Azure CLI

See the Azure_CLI subcommands for VMs:

az vm --help
az vm create --help
az vm image list --all -o table --publisher almalinux   # List all AlmaLinux images

Example from the above tutorial with an AlmaLinux image:

az vm create \
  --resource-group myResourceGroupVM \
  --name myVM \
  --image almalinux:almalinux:8-gen2:8.6.2022052501 \
  --admin-username azureuser \

To create multiple copies of a virtual machine (VM) for use in Azure for development and test:

To create, store and share images at scale, see Azure Compute Galleries.

The basic steps listed in the documentation are:

  1. Deprovision the VM.

  2. Create VM image.

  3. Create a VM from the captured image.

  4. Create a gallery for storing and sharing resources.

Start and stop VMs with the Azure CLI

See Experts Inside example Start, Restart and Stop Azure VM from Azure CLI.

List your existing Azure VMs (output formatted as a simple table):

az vm list --show-details -o table

Display details of a named VM in a specified resourcegroup:

az vm show -g <resourcegroup> -n <VM> [-d|--show-details]

Starting and stopping a VM:

az vm start --help
az vm start      -g <resourcegroup> -n <VM>
az vm restart    -g <resourcegroup> -n <VM>
az vm deallocate -g <resourcegroup> -n <VM>     # Shut down <VM> and deallocate resources (stops billing)
az vm stop       -g <resourcegroup> -n <VM>     # Does NOT deallocate resources and BILLING CONTINUES

Important: The VM_deallocate operation is required to release the VM resource and stop billing. The VM_stop operation only powers off, but does not release resources.

Add the flag --no-wait to avoid waiting for the long-running operation to finish. Add the flag --verbose to increase command logging.

See Azure CLI 2.0: Quickly Start / Stop ALL VMs.

Adding swap space

The Azure web interface does not seem to allow the adding of a swap space. It remains to be seen whether VM compute nodes would actually benefit from having a swap space.

How to configure a swap file in Linux Azure virtual machines:

Increasing machine size

  • Click on the ? (Help) item at the top right of the web page.

  • Click the HelpSupport page and open a New support request.

  • For the Issue type select Service and subscription limits (quotas).

  • Details needed.

Create a virtual machine

Some RedHat resources:

SSH login to Azure VM

When the VM was created, the SSH authorized key for login was offered by the Azure GUI. You must save that SSH key file to a local folder, for example $HOME/.ssh/<vm-name>_key.pem. If you wish to display the public key in the certificate use the openssl-rsa command:

openssl rsa -in $HOME/.ssh/<vm-name>_key.pem -pubout

The VM login username may be defined as azureuser.

First ensure that you can ping the VM’s IP-address (private subnet or public IP).

Next you can try to make an SSH login to the VM named <vm-name>:

ssh -i $HOME/.ssh/<vm-name>_key.pem azureuser@<vm-name>

You can now append additional SSH public keys from your on-premise hosts to the azureuser file:


Root login

Root logins to Azure VMs are not permitted.

Use sudo to run root commands, for example:

sudo dnf update

Ansible with Azure

Azure does not allow remote root logins, so root access must be done using sudo access.

Use this Ansible command to run commands as user azureuser on the VM named <vm-name>:

ansible-playbook --private-key $HOME/.ssh/<vm-name>_key.pem --user azureuser --become <Ansible-commands> -l <vm-name>


  • The SSH key defined above may be required, and this is used with the --private-key option.

  • The user azureuser should be able to run superuser commands using sudo which is used with the --become option.

Storage in Azure

NFS file service on Azure

NFS v3 is supported in Azure since (about) February 2022. Some Azure NFSv3 documentation:

Another NFS file service is Azure NetApp Files.

Azure storage accounts

For NFSv3 storage read Network File System (NFS) 3.0 protocol support for Azure Blob Storage.

Go to Azure_home and select Storage_accounts:

  • Create a new Resource group, for example My_storage (any previously defined storage resource groups can also be used).

  • For the Storage account name you must select a name that is globally unique within all of Azure! Example name: myexamplestorage (only lowercase letters and numbers are permitted).

  • For replication you may want to choose the cheapest Locally redundant storage (LRS) which is only replicated inside the given data center, not at other locations. More expensive alternatives are also available.

  • In the Advanced window enable NFSv3 storage by:

    • Check the box Enable hierarchical namespace

    • Then check the box Enable network file system v3

    Choose a suitable Access_tier (see Blob_pricing):

    • Hot Tier (The Hot tier has the highest storage costs, but the lowest access costs.)

    • Cool Tier (The Cool tier has lower storage costs and higher access costs compared to the Hot tier.)

  • In the Networking window set Public network access to Disabled. When ‘network file share v3’ is enabled, the connectivity method must be set to ‘public endpoint (selected networks)’. Then you must create a Storage_Private_Endpoint connection to grant access. See also What is a private endpoint?.

Select the appropriate Virtual network and Subnets for your subscription.

All Storage_accounts have a public IP-address associated with them. The public IP-address can be looked up with DNS as described in Azure Private Endpoint DNS configuration and the section DNS changes for private endpoints in Storage_Private_Endpoint, for example:

$ nslookup
$ nslookup

Go to the Private endpoint connections tab and Add a private endpoint with an IP-address in your Azure Vnet ( This private IP will be used for NFS mounts in VMs as well as from on-premise nodes.

For the private endpoint select Private DNS integration=No.

  • In the Tags window select the appropriate name tag.

  • Finally Review and Create the storage account.

The DNS servers used in your cluster must be configured to resolve Azure DNS names to the Private endpoint connections, for example: has address

If you use your on-premise DNS server, you have to create a new DNS zone containing addresses for the Private endpoint connections, for example this DNS name:

myexamplestorage IN A

It is a good idea to add the reverse-DNS lookup in the zone as well:


Mount Azure storage container

Mount an NFSv3 storage:

In your Storage account created above, go to the Containers item:

  • Click + Containers to create a new container. Select an appropriate name, then click Advanced and check that No Root Squash is selected.

    Now click Create.

  • In a VM machine mount the container to some example directory:

    mkdir /nfsdata
    mount -o sec=sys,vers=3,nolock,proto=tcp <storage-account-name><storage-account-name>/<container-name>  /nfsdata

    If you add this to /etc/fstab you must also add the nofail flag.

    The sec=sys is the default setting, which uses local UNIX UIDs and GIDs by means of AUTH_SYS to authenticate NFS operations.

    The same NFS mount name can also be used by on-premise nodes, provided they have been configured for routing to the Azure subnet.

NFS automount Azure storage container

In stead of the above static NFS mounting of NFSv3 storage, you can also use the Linux NFS autofs automounter. Here is a suggested method which works on both Azure VMs and on-premise servers:

  1. Create the autofs file /etc/ with contents:

    nfsdata -sec=sys,vers=3,nolock,proto=tcp <storage-account-name><storage-account-name>/<container-name>

    Make sure that DNS resolving <storage-account-name> works correctly both on Azure VMs and on-premise with your chosen <storage-account-name>, see the above DNS setup. The storage name nfsdata can be any other name that makes sense.

  2. Append to /etc/auto.master a line:

    /azure /etc/ --timeout=60

    This will automount onto the /azure directory.

  3. Optional: You may add a link in the Azure VM from /home to the NFSv3 storage:

    $ ln -s /azure/nfsdata /home/nfsdata

    This can be used, for example, to keep users’ $HOME directories under /home while actually storing the directories on the shared Azure NFSv3 filesystem.

  4. Restart the autofs service:

    systemctl restart autofs

Now you have automatic mounting of the Azure NFSv3 storage on both Azure VMs and on-premise servers:

$ cd /home/nfsdata

The autofs will unmount filesystems after some minutes of inactivity.

Cloud node configuration

When the cloud node subnet has been connected to the on-premise cluster subnet through the above described VPN tunnel, there are some additional network configurations which you need to consider:

  1. DNS servers configured in the cloud subnet. It is recommended to configure the on-premise DNS server IP-addresses in the cloud nodes as the primary DNS servers, see this file:

    cat /etc/resolv.conf

    which should contain the same DNS server IP-addresses as for the on-premise nodes. The cloud’s default DNS server should be added last (a maximum of 3 IP-addresses is allowed, see the resolv.conf man-page).

  2. The DNS domain name configured in the cloud nodes should be the same as your on-premise nodes. This can be accomplished with the hostnamectl command, for example:

    cloudnode001# hostnamectl set-hostname
  3. Add the cloud nodes’ static IP-addresses to your on-premise DNS server so that DNS lookups work, for example:

  4. Cloud nodes should relay SMTP E-mails via an SMTP server in your on-premise cluster subnet. For example, a Sendmail server could have this in the /etc/mail/ file:

    define(`SMART_HOST', `')dnl

    Therefore it is recommended to set up an SMTP relay server on your cluster subnet for relaying mails to your organization’s mail service. Note: The SMTP relay server’s firewall must permit SMTP connections from the cloud nodes.

  5. Set up IP routing from on-premise servers and compute nodes to the cloud nodes as described in