SLURM Multinode on OpenStack

Launch Login Node

Prepare User Data

When launching a login node it is worth considering what user data options to provide. While it is not required, user data can provide powerful customisation at launch that can further streamline the cluster build process.

There are several options that can be added to change how a compute node will contact nodes on startup.

• Sharing public ssh key to clients:
  • Instead of manually obtaining and sharing the root public SSH key (passwordless root ssh is required for flight profile) this can be shared over the local network with SHAREPUBKEY=true
• Add an auth key:
  • Add the line AUTH_KEY=<string>. This means that the node will only accept incoming flight hunter nodes that provide a matching authorisation key

#cloud-config
write_files:
  - content: |
      SHAREPUBKEY=true
      AUTH_KEY=banana
    path: /opt/flight/cloudinit.in
    permissions: '0600'
    owner: root:root

Deploy

To set up a cluster, you will need to import a Flight Solo image.

Before setting up a cluster on Openstack, there are several required prerequisites:

• Your own keypair
• A network
• A router
  • With an interface both on the External Gateway network and an Internal Interface on the previously created network
• A security group that allows traffic is given below (if creating the security group through the web interface then the "Any" protocol will need to be an "Other Protocol" rule with "IP Protocol" of -1)

The documentation includes instructions for importing an image to Openstack, and guides for setting up the other prerequisites can be found in the Openstack documentation

To set up a cluster:

1. Go to the Openstack instances page.

   

2. Click "Launch Instance", and the instance creation window will pop up.

3. Fill in the instance name, and leave the number of instances as 1, then click next.

   

4. Choose the desired image to use by clicking the up arrow at the end of its row. It will be displayed in the "Allocated" section when selected.

   

5. Choose the desired instance size by clicking the up arrow at the end of its row. It will be displayed in the "Allocated" section when selected.

   

6. Choose a network in the same way as an image or instance size. Note that all nodes in a cluster must be on the same network.

   

7. Choose a security group in the same way as an image or instance size. Note that all nodes in a cluster must be in the same security group.

   

8. Choose the keypair in the same way as an image or instance size.

   

9. In the "Configuration" section, there is a "Customisation Script" section with a text box. This will be used to set your user data

   

10. When all options have been selected, press the "Launch Instance" button to launch. If the button is greyed out, then a mandatory setting has not been configured.

    

11. Go to the "Instances" page in the "Compute" section. The created node should be there and be finishing or have finished creation.

    

12. Click on the down arrow at the end of the instance row. This will bring up a drop-down menu.

13. Select "Associate Floating IP", this will make the ip management window pop up.

    

14. Associate a floating IP, either by using an existing one or allocating a new one.

    1. To use an existing floating IP:

       1. Open the IP Address drop-down menu.

          

       2. Select one of the IP Addresses.

       3. Click "Associate" to finish associating an IP.

    2. To allocate a new floating IP:

       1. Click the "+" next to the drop-down arrow to open the allocation menu.

          

       2. Click "Allocate IP".

15. Click "Associate" to finish associating an IP.

Launch Compute Nodes

Prepare User Data

Setting up compute nodes is done slightly differently than a login node. The basic steps are the same except subnets, networks and security groups need to match the ones used for the login node.

This is the smallest amount of cloud init data necessary. It allows the login node to find the compute nodes as long as they are on the same network, and ssh into them from the root user (which is necessary for setup).

#cloud-config
users:
  - default
  - name: root
    ssh_authorized_keys:
      - <Content of ~/.ssh/id_alcescluster.pub from root user on login node>

There are several options that can be added to change how a compute node will contact nodes on startup.

• Sending to a specific server:
  • Instead of broadcasting across a range, add the line SERVER=<private server IP> to send to specifically that node, which would be your login node.
• Add an auth key:
  • Add the line AUTH_KEY=<string>. This means that the compute node will send it's flight hunter packet with this key. This must match the auth key provided to your login node

#cloud-config
write_files:
  - content: |
      SERVER=10.10.0.1
      AUTH_KEY=banana
    path: /opt/flight/cloudinit.in
    permissions: '0600'
    owner: root:root
users:
  - default
  - name: root
    ssh_authorized_keys:
      - <Content of ~/.ssh/id_alcescluster.pub from root user on login node>

Deploy

1. Go to the Openstack instances page.

   

2. Click "Launch Instance", and the instance creation window will pop up.

3. Fill in the instance name, and set the number of instances to create, then click next.

   

4. Choose the desired image to use by clicking the up arrow at the end of its row. It will be displayed in the "Allocated" section when selected.

   

5. Choose the desired instance size by clicking the up arrow at the end of its row. It will be displayed in the "Allocated" section when selected.

   

6. Choose a network in the same way as an image or instance size. Note that this should be the same network as the login node.

   

7. Choose a security group in the same way as an image or instance size. Note that this should be the same security group as the login node.

   

8. Choose the keypair in the same way as an image or instance size.

   

9. In the "Configuration" section, there is a "Customisation Script" section with a text box. This will be used to set user data script prepared earlier

   

10. When all options have been selected, press the "Launch Instance" button to launch. If the button is greyed out, then a mandatory setting has not been configured.

    

General Configuration

Create Node Inventory

1. Parse your node(s) with the command flight hunter parse.

   1. This will display a list of hunted nodes, for example

      [flight@login-node.novalocal ~]$ flight hunter parse
      Select nodes: (Scroll for more nodes)
      ‣ ⬡ login-node.novalocal - 10.10.0.1
        ⬡ compute-node-1.novalocal - 10.10.101.1
      

   2. Select the desired node to be parsed with Space, and you will be taken to the label editor

      Choose label: login-node.novalocal
      

   3. Here, you can edit the label like plain text

      Choose label: login1
      

      Tip

      You can clear the current node name by pressing Down in the label editor.

   4. When done editing, press Enter to save. The modified node label will appear next to the ip address and original node label.

      Select nodes: login-node.novalocal - 10.10.0.1 (login1) (Scroll for more nodes)
      ‣ ⬢ login-node.novalocal - 10.10.0.1 (login1)
        ⬡ compute-node-1.novalocal - 10.10.101.1
      

   5. From this point, you can either hit Enter to finish parsing and process the selected nodes, or continue changing nodes. Either way, you can return to this list by running flight hunter parse.

   6. Save the node inventory before moving on to the next step.

      Tip

      See flight hunter parse -h for more ways to parse nodes.

Add genders

1. Optionally, you may add genders to the newly parsed node. For example, in the case that the node should have the gender cluster and all then run the command:

   flight hunter modify-groups --add cluster,all login1
   

SLURM Multinode Configuration

1. Configure profile

   flight profile configure
   

   1. This brings up a UI, where several options need to be set. Use up and down arrow keys to scroll through options and enter to move to the next option. Options in brackets coloured yellow are the default options that will be applied if nothing is entered.
      • Cluster type: The type of cluster setup needed, in this case Slurm Multinode.
      • Cluster name: The name of the cluster.
      • Setup Multi User Environment with IPA?: Boolean value to determine whether to configure a multi-user environment with IPA. If set to true then the following will need to be filled in
        • IPA domain: The domain for the IPA server to use.
        • IPA secure admin password: The password to be used by the admin user of the IPA installation to manage the server.
      • Default user: The user that you log in with.
      • Set user password: Set a password to be used for the chosen default user.
      • IP or FQDN for Web Access: As described here, this could be the public IP or public hostname.
      • IP range of compute nodes: The IP range of the compute nodes used, remember to add the netmask. E.g. 172.31.16.0/20

2. Apply identities by running the command flight profile apply

   1. First apply an identity to the login node

      flight profile apply login1 login
      

   2. Wait for the login node identity to finish applying. You can check the status of all nodes with flight profile list.

      Tip

      You can watch the progress of the application with flight profile view login1 --watch

   3. Apply an identity to the each of the compute nodes (in this example, genders-style syntax is used to apply to node01 and node02)

      flight profile apply node[01-02] compute
      

      Tip

      You can check all available identities for the current profile with flight profile identities

Verifying Functionality

1. Create a file called simplejobscript.sh, and copy this into it:

   #!/bin/bash -l
   echo "Starting running on host $HOSTNAME"
   sleep 30
   echo "Finished running - goodbye from $HOSTNAME"
   

2. Run the script with sbatch simplejobscript.sh, and to test all your nodes try queuing up enough jobs that all nodes will have to run.

3. In the directory that the job was submitted from there should be a slurm-X.out where X is the Job ID returned from the sbatch command. This will contain the echo messages from the script created in step 1