HP StorageWorks P9000 Cluster Extension Software Administrator Guide (TB534-96009, February 2011)

This means that losing half the nodes in a 2-, 4-, 6-, or 8-node cluster or losing the communication
links with 50% of the nodes on each site forces every node to terminate the cluster services because
none of them have access to a majority of the configured nodes.
Therefore, a geographically dispersed MNS-based cluster requires an additional node per cluster
located at a third site so that whenever a disaster affects either the local or remote site, the other
site together with the added node has a majority.
SLE HA cluster setup considerations
Follow the guidelines in this section when you configure clusters for use with P9000 Cluster
Extension.
Quorum
In an SLE HA cluster, quorum is defined as a strict majority of the defined cluster (more than 50%).
With certain failures, a cluster might be divided into two subclusters. In an SLE HA cluster, a
subcluster with more than 50% of the nodes wins the quorum. The subcluster that wins the quorum
re-forms the cluster and fences the subcluster that lost the quorum. The behavior of the subcluster
that lost the quorum depends on the defined no-quorum policy. This behavior is in effect until the
cluster is fenced. When the cluster is fenced, the resources owned by the fenced nodes fail over
to active cluster nodes.
STONITH
STONITH is an SLE HA cluster fencing method. SLE HA cluster provides STONITH plug-ins for
devices such as UPS, PDU, Blade power control devices, and lights out devices. Some plug-ins can
STONITH more than one node (for example, Split Brain Detector STONITH) and some can STONITH
only one node (for example, HP iLO STONITH).
HP iLO STONITH uses the power control functions of an HP iLO device to STONITH a node that
has lost quorum and needs to be fenced.
IMPORTANT: If all of the iLO devices in a cluster are connected using a single network, a single
switch failure might disable iLO, preventing nodes from being fenced. This failure might be difficult
to detect, especially before a node failure where iLO features would be required.
The STONITH action can be set to power off or reset, depending on the environment requirements.
• Power off: The STONITH agent powers off the nodes in the errant subcluster.
• Reset: The STONITH agent resets the nodes in the errant subcluster, and the nodes try to
automatically rejoin the cluster.
NOTE: IPMI fencing can be used for Integrity servers that do not support RIBCL scripting.
Networking in an SLE HA cluster
Configuring redundant and independent cluster communication paths is a good way to avoid Split
Brain conditions. With redundancy in communication paths, the loss of a single interface or switch
does not break the communication between nodes and prevents Split Brain conditions.
Administrators can configure multiple independent communication paths. HP recommends using
bonded Ethernet channels.
Resource constraints
Resource constraints allow administrators to specify which cluster nodes resources can run on, the
order resources are loaded, and the other resources a specific resource is dependent on.
12 P9000 Cluster Extension features