Slurm

Slurm is a highly configurable open source workload manager. See the Slurm project site for an overview.

Slurm can easily be enabled on a CycleCloud cluster by modifying the "run_list" in the configuration section of your cluster definition. The two basic components of a Slurm cluster are the 'master' (or 'scheduler') node which provides a shared filesystem on which the Slurm software runs, and the 'execute' nodes which are the hosts that mount the shared filesystem and execute the jobs submitted. For example, a simple cluster template snippet may look like:

[cluster custom-slurm]

[[node master]]
    ImageName = cycle.image.centos7
    MachineType = Standard_A4 # 8 cores

    [[[cluster-init cyclecloud/slurm:default]]]
    [[[cluster-init cyclecloud/slurm:master]]]
    [[[configuration]]]
    run_list = role[slurm_master_role]

[[nodearray execute]]
    ImageName = cycle.image.centos7
    MachineType = Standard_A1  # 1 core

    [[[cluster-init cyclecloud/slurm:default]]]
    [[[cluster-init cyclecloud/slurm:execute]]]
    [[[configuration]]]
    run_list = role[slurm_master_role]
    slurm.autoscale = true
    # Set to true if nodes are used for tightly-coupled multi-node jobs
    slurm.hpc = true
    slurm.default_partition = true

Slurm can easily be enabled on a CycleCloud cluster by modifying the "run_list" in the configuration section of your cluster definition. The two basic components of a Slurm cluster are the 'scheduler' node which provides a shared filesystem on which the Slurm software runs, and the 'execute' nodes which are the hosts that mount the shared filesystem and execute the jobs submitted. For example, a simple cluster template snippet may look like:

[cluster custom-slurm]

[[node scheduler]]
    ImageName = cycle.image.centos7
    MachineType = Standard_A4 # 8 cores

    [[[cluster-init cyclecloud/slurm:default]]]
    [[[cluster-init cyclecloud/slurm:scheduler]]]
    [[[configuration]]]
    run_list = role[slurm_scheduler_role]

[[nodearray execute]]
    ImageName = cycle.image.centos7
    MachineType = Standard_A1  # 1 core

    [[[cluster-init cyclecloud/slurm:default]]]
    [[[cluster-init cyclecloud/slurm:execute]]]
    [[[configuration]]]
    run_list = role[slurm_scheduler_role]
    slurm.autoscale = true
    # Set to true if nodes are used for tightly-coupled multi-node jobs
    slurm.hpc = true
    slurm.default_partition = true
`

Editing Existing Slurm Clusters

Slurm clusters running in CycleCloud versions 7.8 and later implement an updated version of the autoscaling APIs that allows the clusters to utilize multiple nodearrays and partitions. To facilitate this functionality in Slurm, CycleCloud pre-populates the execute nodes in the cluster. Because of this, you need to run a command on the Slurm scheduler node after making any changes to the cluster, such as autoscale limits or VM types.

Making Cluster Changes

The Slurm cluster deployed in CycleCloud contains a script that facilitates this. After making any changes to the cluster, run the following command as root (e.g., by running sudo -i) on the Slurm scheduler node to rebuild the slurm.conf and update the nodes in the cluster:

cd /opt/cycle/jetpack/system/bootstrap/slurm
./cyclecloud_slurm.sh scale

Removing all execute nodes

As all the Slurm compute nodes have to be pre-created, it's required that all nodes in a cluster be completely removed when making big changes (such as VM type or Image). It is possible to remove all nodes via the UI, but the cyclecloud_slurm.sh script has a remove_nodes option that will remove any nodes that aren't currently running jobs.

Creating additional partitions

The default template that ships with Azure CycleCloud has two partitions (hpc and htc), and you can define custom nodearrays that map directly to Slurm partitions. For example, to create a GPU partition, add the following section to your cluster template:

    [[nodearray gpu]]
    MachineType = $GPUMachineType
    ImageName = $GPUImageName
    MaxCoreCount = $MaxGPUExecuteCoreCount
    Interruptible = $GPUUseLowPrio
    AdditionalClusterInitSpecs = $ExecuteClusterInitSpecs

        [[[configuration]]]
        slurm.autoscale = true
        # Set to true if nodes are used for tightly-coupled multi-node jobs
        slurm.hpc = false

        [[[cluster-init cyclecloud/slurm:execute:2.0.1]]]
        [[[network-interface eth0]]]
        AssociatePublicIpAddress = $ExecuteNodesPublic

Memory settings

CycleCloud automatically sets the amount of available memory for Slurm to use for scheduling purposes. Because the amount of available memory can change slightly due to different Linux kernel options, and the OS and VM can use up a small amount of memory that would otherwise be available for jobs, CycleCloud automatically reduces the amount of memory in the Slurm configuration. By default, CycleCloud holds back 5% of the reported available memory in a VM, but this value can be overridden in the cluster template by setting slurm.dampen_memory to the percentage of memory to hold back. For example, to hold back 20% of a VM's memory:

    slurm.dampen_memory=20

Disabling autoscale for specific nodes or partitions

While the built-in CycleCloud "KeepAlive" feature does not currently work for Slurm clusters, it is possible to disable autoscale for a running Slurm cluster by editing the slurm.conf file directly. You can exclude either individual nodes or entire partitions from being autoscaled.

Excluding a node

To exclude a node or multiple nodes from autoscale, add SuspendExcNodes=<listofnodes> to the Slurm configuration file. For example, to exclude nodes 1 and 2 from the hpc partition, add the following to /etc/slurm/slurm.conf:

SuspendExcNodes=hpc-pg0-[1-2]

Then restart the slurmctld service for the new configuration to take effect.

Excluding a partition

Excluding entire partitions from autoscale is similar to excluding nodes. To exclude the entire hpc partition, add the following to /etc/slurm/slurm.conf

SuspendExcParts=hpc

Then restart the slurmctld service.

Troubleshooting

UID conflicts for Slurm and Munge users

By default, this project uses a UID and GID of 11100 for the Slurm user and 11101 for the Munge user. If this causes a conflict with another user or group, these defaults may be overridden.

To override the UID and GID, click the edit button for both the scheduler node:

Edit Scheduler

Edit Scheduler

And the execute nodearray: Edit Nodearray

and add the following attributes to the Configuration section:

Edit Configuration

    slurm.user.name = slurm
    slurm.user.uid = 11100
    slurm.user.gid = 11100
    munge.user.name = munge
    munge.user.uid = 11101
    munge.user.gid = 11101

Autoscale

CycleCloud uses Slurm's Elastic Computing feature. To debug autoscale issues, there are a few logs on the scheduler node you can check. The first is making sure that the power save resume calls are being made by checking /var/log/slurmctld/slurmctld.log. You should see lines like:

[2019-12-09T21:19:03.400] power_save: pid 8629 waking nodes htc-1

The other log to check is /var/log/slurmctld/resume.log. If the resume step is failing, there will also be a /var/log/slurmctld/resume_fail.log. If there are messages about unknown or invalid node names, make sure you haven't added nodes to the cluster without following the steps in the "Making Cluster Changes" section above.

Slurm Configuration Reference

The following are the Slurm specific configuration options you can toggle to customize functionality:

Slurm Specific Configuration Options Description
slurm.version Default: '18.08.7-1'. This is the Slurm version to install and run. This is currently the default and only option. In the future additional versions of the Slurm software may be supported.
slurm.autoscale Default: 'false'. This is a per-nodearray setting that controls whether Slurm should automatically stop and start nodes in this nodearray.
slurm.hpc Default: 'true'. This is a per-nodearray setting that controls whether nodes in the nodearray will be placed in the same placement group. Primarily used for nodearrays using VM families with InfiniBand. It only applies when slurm.autoscale is set to 'true'.
slurm.default_partition Default: 'false'. This is a per-nodearray setting that controls whether the nodearray should be the default partition for jobs that don't request a partition explicitly.
slurm.dampen_memory Default: '5'. The percentage of memory to hold back for OS/VM overhead.
slurm.suspend_timeout Default: '600'. The amount of time (in seconds) between a suspend call and when that node can be used again.
slurm.resume_timeout Default: '1800'. The amount of time (in seconds) to wait for a node to successfully boot.
slurm.user.name Default: 'slurm'. This is the username for the Slurm service to use.
slurm.user.uid Default: '11100'. The User ID to use for the Slurm user.
slurm.user.gid Default: '11100'. The Group ID to use for the Slurm user.
munge.user.name Default: 'munge'. This is the username for the MUNGE authentication service to use.
munge.user.uid Default: '11101'. The User ID to use for the MUNGE user.
munge.user.gid Default: '11101'. The Group ID to use for the MUNGE user.

CycleCloud supports a standard set of autostop attributes across schedulers:

Attribute Description
cyclecloud.cluster.autoscale.stop_enabled Is autostop enabled on this node? [true/false]
cyclecloud.cluster.autoscale.idle_time_after_jobs The amount of time (in seconds) for a node to sit idle after completing jobs before it is scaled down.
cyclecloud.cluster.autoscale.idle_time_before_jobs The amount of time (in seconds) for a node to sit idle before completing jobs before it is scaled down.