Database Objects Fields and Attributes¶

Various ClusterWare database objects (e.g., nodes, boot configurations, image configurations, administrators, attributes) each carry with them detailed descriptors called fields. Each field consists of a name-value pair and is relevant for its database object type. Fields are predefined by ClusterWare. The cluster administrator uses the update action to change a field value.

For instance, a compute node object for each node has fields mac with the node's MAC address, name with the node's alphanumeric name, and power_uri with a value denoting how to communicate via ipmi to that node. For example, the command scyld-nodectl -i n0 ls -l displays all the defined fields' name-value pairs for node n0.

Compute node and Attribute Groups object types have special fields called attributes, where an attribute is a collection of one or more attribute name-value pairs. Attribute names that begin with an underscore "_" are called reserved attributes or system attributes. The cluster administrator uses the set action to change an attribute value. See the following section Reserved Attributes for details.

Additional attributes can be added by a cluster administrator as desired, each with a custom name and value defined by the administrator. Any script on a compute node can access the local file /etc/clusterware/attributes.ini and find that node's attributes. On the node there are helper functions in /opt/scyld/clusterware-node/functions.sh for reading attributes, specifically the function attribute_value.

Reserved Attributes¶

Within the ClusterWare attribute system, administrators are encouraged to store whatever information they find useful for labeling and customizing nodes. For ease of use, attributes names should be valid Javascript variable names, i.e., meaning that they may begin with any uppercase or lowercase letter, followed by letters, digits, or underscores. Names that start with an underscore are used by ClusterWare and should be set by administrators to affect the behavior of the system. These will be referred to as system attributes throughout this discussion.

Attributes are stored internally as a Javascript dictionary mapping strings to strings, otherwise known as name-value pairs. Administrator-defined attribute values should be strings and relatively small in size. The ClusterWare backend database enforces some document size constraints, and collections of node attributes should be no more than tens to hundreds of kilobytes in size. Individual attributes can be any length as long as the overall attribute group or node object size does not exceed this limits. Generally, if a cluster configuration is approaching these sizes, a cluster administrator pursue moving data from the database into shared storage locations referenced by database entries.

Attributes can be applied directly to nodes, but may also be collected into groups, and then these groups applied to sets of nodes. Attributes passed to nodes through groups are treated no differently than those applied directly to a node. Attribute groups help cluster administrators create more scalable and manageable configurations. See Node Attributes for more details.

The remainder of this section is a list of system attributes describing their use and allowed values.

_ansible_pull¶

Default: none

Values: reference to an ansible git repo and a playbook in that repo

Depends: none

See Appendix: Using Ansible for details about format and usage.

_ansible_pull_args¶

Default: none

Values: optional arguments for _ansible_pull

Depends: using _ansible_pull

Specify optional arguments for an _ansible_pull. See Appendix: Using Ansible for details about format and usage.

_ansible_pull_now¶

Default: none

Values: reference to an ansible git repo and a playbook in that repo

Depends: none

The cluster administrator must systemctl enable cw-ansible-pull-now and systemctl start cw-ansible-pull-now. See Appendix: Using Ansible for details about format and usage.

_busy¶

Default: undefined (see below)

Values: boolean (case-insensitive 1/0, on/off, y/n, yes/no, t/f, true/false)

Depends: none

This attribute explicitly controls the behavior of the compute node's cw-status-updater service which periodically gathers node state information every _status_secs seconds (default 10) and reports that information to its parent head node.

The cw-status-updater service can function in one of two ways:

The default manner that gathers frequently changing state (e.g., uptime and load average) and occasionally gathers (albeit more expensively) infrequently changing state information (e.g., what hardware is present and which ClusterWare packages are currently installed), or

A "busy mode" manner that severely reduces the scope of what information is gathered and reported. The service in "busy mode" is minimally invasive to performance of real-time (especially multi-node) applications that are sensitive to interruptions and to "jitter".

If _busy is undefined, then "busy mode" can be enabled or disabled by the presence or absence of /opt/scyld/clusterware-node/etc/busy.flag, which can be created in a job scheduler prologue and delete in an epilogue, or can be manually created and deleted.

If neither _busy and busy.flag are employed, then the compute node may itself heuristically determine on its own whether or not to execute in "busy mode".

A compute node in "busy mode" reports that with scyld-nodectl status -l showing "busy: True".

_boot_config¶

Default: none

Values: boot configuration identifier

Depends: none

The _boot_config attribute defines what boot configuration a given node should should use. For a detailed discussion of boot configurations and other database objects, please see Node Images and Boot Configurations.

A boot configuration identifier may be a, possibly truncated, UID or a boot configuration name.

_boot_rw_layer¶

Default: overlayfs

Values: overlayfs, rwtab

Depends: _boot_style == roram or iscsi

Use _boot_rw_layer to control the type of overlay used to provide read/write access to an otherwise read-only root file system image. The overlayfs provides a writable overlay across the entire file system, while the rwtab approach only allows write access to the locations defined in /etc/rwtab or /etc/rwtab.d in the node image.

Note that prior to kernel version 4.9, overlayfs does not support SELinux extended attributes and so cannot be used for compute nodes with SELinux in enforcing mode. The rwtab option does work with SElinux, but two additional changes need to be made when enabling rwtab. First, the cluster administrator must modify the /etc/sysconfig/readonly-root file in the node image to ensure READONLY is set to "yes":

READONLY=yes

Second, the kernel cmdline in the appropriate boot configuration must include "ro":

cmdline: enforcing=1 ro

_boot_style¶

Default: rwram

Values: rwram, roram, iscsi, disked, next, sanboot, live

Depends: none

Root file system images can be supplied to nodes through a variety of mechanisms, and this can be controlled on a per-node basis through the _boot_style attribute. In both the rwram and roram modes, the node will download the entire image into RAM and either unpack it into a tmpfs RAM file system (rwram) or apply a writable overlay (roram). These boot styles have the advantage of post-boot independence from the head node, meaning that the loss of a head node will not directly impact booted compute nodes.

The iscsi option uses less RAM as the boot image is not downloaded into node RAM, but depends on the head node even after the node is fully booted. Due to this dependence a head node crash may cause attached compute nodes to hang and lose work. This approach requires a writable overlay, as the images may be shared between multiple nodes.

With the disked option, the node boots with images read from local storage. See Appendix: Booting From Local Storage Cache for details.

Use the next option to exit the boot loader and allow the BIOS to try the next device in the BIOS boot order. Since this process depends on support in the BIOS, it may not work on every server model.

The sanboot option causes the booting node to boot using the iPXE sanboot command and defaults to booting the first hard disk. Please see the _ipxe_sanboot attribute for more details.

The live option only works for ISO-based configurations, e.g., those used for kickstart. For supported ISOs (e.g., RHEL-based) the node boots into the live installer, and the administrator needs to interact with it via the (likely graphical) system console.

_boot_tmpfs_size¶

Default: half of RAM

Values: 1g, 2g, etc.

Depends: _boot_style == rwram or _boot_rw_layer == overlayfs

During the node boot process, a tmpfs is used to provide a writable area for diskless compute nodes. For the rwram boot style this attribute controls the size of the root file system where the image is unpacked. When booting with overlayfs on a roram or iscsi style, this attribute controls the size of the writable overlay.

_coreos_ignition_url¶

Default: none

Values: The URL of a RHCOS *.ign ignition file.

Depends: none

Both _coreos_ignition_url and _coreos_install_dev are attributes that must be set to fill in variables in the associated boot config's cmdline. See Using RHCOS.

_coreos_install_dev¶

Default: none

Values: The device on the target node into which the image is installed.

Depends: none

Both _coreos_ignition_url and _coreos_install_dev are attributes that must be set to fill in variables in the associated boot config's cmdline. See Using RHCOS.

_disk_cache¶

Default: none

Values: local partition name + optional encryption

Depends: none

Specifies a persistent location where the node can store downloaded images. This location should be a local partition with sufficient size to hold a handful of compressed images.

If the specified location exists, then the node will retain there a copy of the downloaded image. During subsequent boots the node will first compare the checksum of a file previously saved with the expected checksum provided by the head node in order to avoid unnecessary downloads.

If the specified partition does not exist, then an error will be logged, although the node will download the image to RAM and still boot. If the partition exists but cannot be mounted, then it will be reformatted.

Optionally Linux Unified Key Setup (LUKS) encryption can also be specified for the partition. Append :encrypt to the partition name to encrypt with a random key, or append :encrypt=KEY to specify an encryption key.

If no key is specified, encryption is performed using standard LUKS tools with 128 bytes of data from /dev/urandom stored in a key file used as the passphrase. This key file is only briefly stored in RAM and deleted shortly before an Ext4 file system is created on the newly encrypted partition.

Alternatively, if the specified key is TPM then the random key will be stored in the booting system's Trusted Platform Module (TPM) and deleted out of RAM shortly before the file system is created. The key can also be bound to specific TPM Platform Configuration Register (PCR) values meaning that the TPM will not later reveal the key unless those PCRs hold the same values. Since these values include hashes of the BIOS code, configuration, kernel, and other boot-time binaries access to the encrypted partition can be restricted to specific boot-time configurations. If the TPM has an owner password set it must be provided in the _tpm_owner_pass attribute.

Note

The cryptsetup-luks package must be installed in the image being booted.

Specifying a KEY is essentially necessary for _disk_cache because without that after a subsequent reboot the partition contents will be lost as they were encrypted with an unknown random key.

For example:

scyld-nodectl -i n[0-63] set _disk_cache=/dev/nvme0n1p2:encrypt=Penguin

If _disk_cache is present but no _disk_root is provided, then if a roram-compatible image is downloaded, then the node will boot directly from the cached image with a writable overlay.

Important

Any data in the partition specified as a _disk_cache may be destroyed at boot time!

Similar to /etc/fstab, partitions can be identified by device path, UUID, PARTLABEL, or PARTUUID.

_disk_root¶

Default: none

Values: local partition name + optional encryption

Depends: ignored unless _boot_style == disked

Specifies a persistent location into which at boot time the node can unpack the root image. This will delete the contents of the partition before unpacking the root image. If the specified partition does not exist, then an error will be logged, although the node will still boot using the image unpacked into RAM.

Similar to _disk_cache, append :encrypt to the partition name to encrypt with a random key, or :encrypt=KEY to specify the encryption key. For _disk_root a random key is preferable, as the _disk_root contents are intended to be ephemeral across boots.

Important

All data in the partition specified as a _disk_root will be destroyed at boot time!

Similar to /etc/fstab, partitions can be identified by device path, UUID, PARTLABEL, or PARTUUID.

_disk_wipe¶

Default: none

Values: comma-separated list of local partition names + optional encryption

Depends: none

The listed partitions will be reformatted at every boot with an Ext4 file system. Similar to _disk_cache, append :encrypt to the partition name to enable "encryption at rest", or :encrypt=KEY to specify the encryption key. Like _disk_root the random key is preferable to ensure _disk_wipe partition contents are not retrievable from a physically removed storage device.

_gateways¶

Default: The default gateway for the node's interfaces

Values: <ifname>=IPaddress

Depends: None

Override the interface ifname's current gateway value with an alternative IP address. For example, _gateways=enpls0f0=10.20.30.40,enpls0f1=10.20.40.40.

_health¶

Default: none

Values: node health status

Depends: none

Cluster administrators can use a health check program that executes on a compute node (see _health_check) and relays the result back to the head node. The administrator can view a health summary with:

scyld-nodectl status --health [--refresh]

which displays one of three states: "checking" (for _health=<seconds-since-epoch), "healthy", or "unhealthy" (for _health set to anything other than <seconds> or "healthy").

The administrator can of course view _health as an attribute of a specific node or nodes, e.g.,

scyld-nodectl -in42 ls -l
scyld-nodectl -in42 ls -l | grep "_health:"

The administrator can explicitly set a custom _health value, e.g.,:

scyld-nodectl -in42 set _health="Sent back to Penguin with RMA #123456"

or the health check program on the node can return a more elaborate "unhealthy" _health value, such as "GPU2 is unhealthy". Then:

scyld-nodectl status --health

interprets that atypical elaborate value as "unhealthy", although the full elaborate unhealthy value is visible by adding -l:

scyld-nodectl status --health -l

and the GUI shows the elaborate value as a detail of the "unhealthy" node.

_health_check¶

Default: none

Values: path to health check program on node

Depends: none

Cluster administrators use the _health_check attribute to specify the path to a script or binary executable that implements the health check for display in the _health attribute (see _health).

When a health check program begins executing on the node, it should initially return a _health value of the current seconds since epoch, e.g.,

set-node-attribs _health=$(date +%s)

The ClusterWare GUI and scyld-nodectl status --health interpret this seconds-since-epoch value as "checking". At completion of the health check, the "healthy", "unhealthy", or more elaborate unhealthy result should be sent back to the head node in the using the same set-node-attribs _health=<value> mechanism.

_health_check_secs¶

default: 300

Values: seconds between executing the health check program

Depends: none

Specifies the interval in seconds for executing a health check program specified by the _health_check attribute. See _health_check.

_hostname¶

Default: none

Values: Hostname or fully-qualified domain name

Depends: none

Booting compute nodes will assign the value of _hostname as their hostname using the hostnamectl command. If the attribute value is a simple name (without periods), then the cluster domain will be appended to construct a FQDN. Changes to this variable take effect during the next reboot.

_hosts¶

Default: blank

Values: download

Depends: none

During the compute node boot process, a list of known hosts is downloaded from the head node and is appended to the compute node's /etc/hosts. By default this will only append a list of head nodes to ensure that each compute node can resolve all head nodes without DNS. If the _hosts attribute is set to 'fetch', then all compute node names and IP addresses will be appended to /etc/hosts.

_ips¶

Default: none

Values: comma-separated IP assignments

Depends: none

Compute nodes commonly define additional high-speed network interfaces other than the PXE boot network. These interfaces are commonly defined by ifcfg-XXX files located in /etc/sysconfig/network-scripts and differ between nodes only in the assigned IP address. Use the _ips attribute to specify what IP address should be assigned to an individual node on one or more interfaces. For example, a value of _ips=eno0=10.10.23.12,ib0=192.168.24.12 would cause the prenet/write_ifcfg.sh startup script to replace any IPADDR= line in /etc/sysconfig/network-scripts/ifcfg-ib0 with IPADDR=192.168.24.12 and would similarly modify the adjacent ifcfg-en0 file, replacing any IP assignment in that file with IPADDR=10.10.23.12.

_ipxe_sanboot¶

Default: none

Values: local disk or partition

Depends: _boot_style == sanboot

Use this attribute to cause a node to boot using the iPXE sanboot command. This is most commonly used to boot a locally installed disk, although administrators are cautioned to be extremely careful with stateful compute nodes as they will retain modifications from previous boots, leading to an unexpectedly heterogeneous cluster.

Nodes with this attribute set will not download an image from the head node and will instead boot based on the URL or other iPXE sanboot arguments provided. Please see the iPXE documentation for the details of what iPXE provides: http://ipxe.org/cmd/sanboot

In addition to the arguments and URLs supported by iPXE, ClusterWare also accepts a shorter URL for booting local disks of the form local://0xHH where 'HH' is a hexadecimal value specifying a local hard disk. The first disk is identified as 0x80, the second is 0x81, and so on. The provided hexadecimal value is then used in a sanboot --no-describe --drive 0xHH call.

_macs¶

Default: The default MAC address for each of the node's interfaces

Values: <ifname>=<MACaddress>

Depends: None

Override the interface ifname's current MAC address with an alternative value. For example, _macs=bond0=aa:bb:cc:dd:ee:ff. Generally only used for bonded interfaces. Ignored for the booting interface bootnet.

_no_boot¶

Default: false

Values: boolean equivalents (0 / 1, true / false, t / f, yes / no, y / n)

Depends: none

The _no_boot attribute controls whether information about a node is provided to the DHCP server. Any node with _no_boot set to true will not receive DHCP offers from any ClusterWare head node. This allows an administrator to temporarily remove a node from the cluster.

_preferred_head¶

Default: none

Values: head node UID

Depends: none

In a multihead configuration any head node can provide boot files to any compute node in the system. In most cases this is a desirable feature because the failure of any given head node will not cause any specific set of compute nodes to fail to boot. In some cases the cluster administrator may want to specify a preference of which head node should handle a given compute node. By setting a compute node's _preferred_head attribute to a specific head node's UID, all head nodes will know to point that node toward the preferred head node. This is implemented during the boot process when the iPXE script is generated and passed to the compute node. This means that any head node can still supply DHCP, the iPXE binaries, and the iPXE boot script, but the subsequent kernel, initramfs, and root file system files will be provided by the preferred head node, and thereafter the node's boot status information will be sent to that _preferred_head.

_remote_pass¶

Default: none

Values: node account password for _remote_user attribute

Depends: none

Supports an alternative to the customary ClusterWare ssh-key functionality. It is useful to support scyld-nodectl exec to non-ClusterWare compute nodes which do not have clusterware-node installed, but which do accept user/password authentication.

To use, install the sshpass RPM on the head node. Set the _remote_pass attribute to the password of the _remote_user attribute user name (default root). Subsequent executions of scyld-nodectl exec to nodes that are set up with this attribute will employ this user/password pair to authenticate access to those target node(s).

Note

Use of sshpass is discouraged and is not a best practice. A clear text password is a significant security risk.

_remote_user¶

Default: root

Values: node account name

Depends: none

The _remote_user attribute controls what account is used on the compute node when executing the scyld-nodectl reboot/shutdown commands. Please ensure the specified account can execute sudo shutdown without a password or soft power control will not work. Similarly the scyld-nodectl exec and scyld-nodectl ssh commands will also use the specified remote user account and the boot-time script that downloads head node keys will store those keys in the _remote_user's authorized_keys file.

_status_cpuset¶

Default: all available CPUs

Values: list of one or more CPU numbers

Depends: none

When a compute node boots, the status-updater and related child processes can execute by default on any of the node's CPUs, as chosen by the kernel's scheduler. The administrator may instead choose to restrict which CPUs these processes use to be a subset of all CPUs, or even to just a single CPU, in order to minimize the impact that these processes may have on a time-critical application(s) executing on the other CPUs.

The _status_cpuset value is a list of CPUs to use. For example, set _status_cpuset=0 restricts the processes to just CPU 0, set _status_cpuset="0-1" restricts to CPUs 0 and 1, and set _status_cpuset="0-1,4" restricts to CPUs 0, 1, and 4. See man 7 cpuset for details.

_status_hardware_secs¶

Default: 300

Values: seconds between checking for status hardware changes

Depends: none

A node sends its hardware state (viewed with scyld-nodectl list --long and list --long-long) as a component of its larger basic status information. See _status_secs above. This hardware component is typically only sent once at boot time. However, the node periodically reevaluates its hardware state every _status_hardware_secs seconds, and in the rare event that something has changed since it last communicated its hardware state to its parent head node, then the node includes the updated hardware information in its next periodic basic status message.

Changes to this value are communicated to an up node without needing to reboot the node.

_status_packages_secs¶

Default: 0

Values: seconds between checking for installed packages changes

Depends: none

The time interval in seconds between the relatively expensive search for what Scyld packages are installed. This value times 10 is the time interval between the even more expensive calculations of a sha256sum hash of the sorted list of names of all installed packages, distilled into a single hexidecimal value. These values are seen by executing scyld-nodectl -i<nodes> status -L on the head node.

A non-zero value should be longer than the _status_secs value, described below.

If the value is zero, then these packages searches and calculations are done just at node boot time, and additionally when (and if) the administrator executes /usr/bin/update-node-status --hardware on a compute node. Such run-time changes to a node's installed packages are relatively rare, so the default value is zero to minimize the performance impact of these operations.

Changes to this value are communicated to an up node without needing to reboot the node.

_status_secs¶

Default: 10

Values: seconds between status updates

Depends: none

Booted compute nodes periodically send basic status information to their parent head node. This value controls how often these messages are sent. Although the messages are relatively small, clusters with more compute nodes per head node will want to set this to a longer period to reduce load on the compute nodes.

Changes to this value are communicated to an up node without needing to reboot the node.

_tpm_owner_pass¶

Default: none

Values: Owner password for the compute node TPM

Depends: none

Certain TPM commands require authentication using the "owner" TPM password. This means that the clear-text password must be provided to systems using the TPM for disk encryption via this attribute.