Slurm+NFS

Slurm+NFS

This recipe shows how to orchestrate a Slurm on Batch cluster with a single Batch pool providing compute node VMs for Slurm workloads along with a shared NFS filesystem.

Configuration

Please see refer to this set of sample configuration files for this recipe.

Credentials Configuration

The credentials configuration should have management Azure Active Directory credentials defined along with a valid storage account. The management section can be supplied through environment variables instead if preferred. The batch section should also be populated which associates all of the Batch pools used by Slurm partitions. Additionally, a slurm section with the db_password must be defined.

Pool Configuration

The pool configuration can be modified as necessary for the requisite OS and other tooling that should be installed. The vm_count should be kept as 0 for both dedicated and low_priority during the initial allocation as Slurm's elastic cloud bursting will size the pools appropriately.

FS Configuration

The remote fs configuration file requires modification. Properties to modify are:

• resource_group all resource groups should be modified to fit your scenario.
• location should be modified to the Azure region where you would like the storage cluster created. If linking against Azure Batch compute nodes, it should be in the same region as your Azure Batch account.
• managed_disks should be modified for the number, size and type of managed disks to allocate for the file server.
• storage_clusters should be modified to have a unique name instead of mystoragecluster if you prefer.
• hostname_prefix should be modified to your perferred resource name prefix.
• virtual_network should be modified for the address prefixes and subnet properties that you prefer.
• network_security should be modified for inbound network security rules to apply for SSH and external NFSv4 client mounts. If no NFSv4 clients external to the virtual network are needed, then the entire nfs security rule can be omitted.
• file_server options such as mountpoint and mount_options should be modified to your scenario. Type should not be modified from nfs.
• vm_size can be modified for the file server depending upon your scenario. If using premium managed disks, then a premium VM size must be selected here.
• vm_disk_map contains all of the disks used for each VM. For nfs, there is only a single VM, thus all disks should be mapped in the "0" entry.

Slurm Configuration

The Slurm configuration should include the appropriate location and virtual network settings for the controller and login nodes, in addition to defining the appropriate elastic partitions. Please see the Slurm on Batch guide and the Slurm configuration document for more information on each option.

Commands to orchestrate the Slurm cluster

After modifying the configuration files as required, you can orchestrate the entire Slurm cluster creation with slurm cluster orchestrate. The orchestrate command wraps up the NFS disk allocation (fs disks add), NFS file server creation (fs cluster add), Batch pool allocation (pool add), and Slurm controller/login creation (slurm cluster create) into one command. The commands can be invoked separately if desired. The following assumes the configuration files are in the current working directory.

# ensure all configuration files are in the appropriate directory
export SHIPYARD_CONFIGDIR=.

# orchestrate the Slurm cluster
./shipyard slurm cluster orchestrate --storage-cluster-id nfs -y

You can log into the login nodes by issuing the command:

./shipyard slurm ssh login

which will default to logging into the first login node (since this cluster only has one login node, it is the only possible node to log in to).

There you will be able to run your Slurm commands such as sbatch, squeue, salloc, srun, etc..

To delete the Slurm cluster:

# delete the Batch pool providing Slurm compute nodes
./shipyard pool del -y

# delete the Slurm controller and login nodes
./shipyard slurm cluster destroy -y

# delete the RemoteFS shared file system
./shipyard fs cluster del nfs -y --delete-data-disks