SPARC: Guidelines for Oracle VM Server for SPARC Logical Domains as Cluster Nodes

Consider the following points when you create an Oracle VM Server for SPARC logical domain or guest domain on a physically clustered machine that is SPARC hypervisor capable, for use as a cluster node:

  • SR-IOV devices – An SR-IOV device is supported with a logical domain that is configured to run as a cluster node. See the Oracle Solaris Cluster 4 Compatibility Guide (http://www.oracle.com/technetwork/server-storage/solaris-cluster/overview/solariscluster4-compatibilityguide-1429037.pdf) for information about supported SR-IOV devices.

  • SCSI LUN requirement – The virtual shared storage device, or virtual disk back end, of an Oracle VM Server for SPARC guest domain must be a full SCSI LUN in the I/O domain. You cannot use an arbitrary virtual device.

  • Fencing – Do not export a storage LUN to more than one guest domain cluster node on the same physical machine unless you also disable fencing for that device from the guest domain nodes that you export the LUN to. Otherwise, the SCSI reservation on the device can be modified by one of the clusters and exclude nodes of other clusters. When that happens any other guest domain cluster nodes that subsequently try to access the device will panic. Such a situation could be triggered for example when one node of a cluster is halted. When you export a storage LUN to more than one guest domain cluster node on the same physical machine, login to each guest domain cluster node that you have exported the LUN to and use cldev command to disable fencing of the device.

    For example:

    # cldev set -p default_fencing=nofencing d2"
  • Network isolation – Guest domains that are located on the same physical machine but are configured in different clusters must be network isolated from each other. Use one of the following methods:

    • Configure the clusters to use different network interfaces in the I/O domain for the private network.

    • Use different network addresses for each of the clusters when you perform initial configuration of the clusters.

  • Networking in guest domains – Network packets to and from guest domains must traverse service domains to reach the network drivers through virtual switches. Virtual switches use kernel threads that run at system priority. The virtual-switch threads must be able to acquire needed CPU resources to perform critical cluster operations, including heartbeats, membership, checkpoints, and so forth. Configuring virtual switches with the mode=sc setting enables expedited handling of cluster heartbeat packets. However, the reliability of other critical cluster operations can be enhanced by adding more CPU resources to the service domain under the following workloads:

    • High-interrupt load, for example, due to network or disk I/O. Under extreme load, virtual switches can preclude system threads from running for a long time, including virtual-switch threads.

    • Real-time threads that are overly aggressive in retaining CPU resources. Real-time threads run at a higher priority than virtual-switch threads, which can restrict CPU resources for virtual-switch threads for an extended time.

  • Non-shared storage – For non-shared storage, such as for Oracle VM Server for SPARC guest-domain OS images, you can use any type of virtual device. You can back such virtual devices by any implement in the I/O domain, such as files or volumes. However, do not copy files or clone volumes in the I/O domain for the purpose of mapping them into different guest domains of the same cluster. Such copying or cloning would lead to problems because the resulting virtual devices would have the same device identity in different guest domains. Always create a new file or device in the I/O domain, which would be assigned a unique device identity, then map the new file or device into a different guest domain.

  • Exporting storage from I/O domains – If you configure a cluster that is composed of Oracle VM Server for SPARC I/O domains, do not export its storage devices to other guest domains that also run Oracle Solaris Cluster software.

  • Live migration restriction – The live migration and cold migration features of Oracle VM Server for SPARC are not supported for logical domains that are configured to run as cluster nodes.

    However, guest domains that are configured to be managed by the HA for Oracle VM Server for SPARC data service do support cold migration and live migration. But no other type of logical domain in this data service configuration is supported to use cold migration or live migration.

  • mpgroup boot drive support – Oracle Solaris Cluster supports mpgroups for boot drives with fencing disabled using the cldev command. You must disable fencing before configuring the vdisk for the boot drive in an mpgroup.

    For example:

    # cldev set -p default_fencing=nofencing d2
  • Virtual disk multipathing – With the exception of boot drives, do not configure the virtual disk multipathing feature of Oracle VM Server for SPARC on a logical domain that is configured as a cluster node.

For more information about Oracle VM Server for SPARC, see the Oracle VM Server for SPARC 3.6 Administration Guide.