HA-LDOM live migration in Oracle Solaris Cluster 4.2
By krishna.k.murthy on Aug 07, 2014
Most Oracle Solaris Cluster resource types clearly separate their stopping and starting actions. However, the HA-LDom (HA Logical Domain / HA for Oracle VM for SPARC data service) agent has a capability of performing live migration, in which an entire switchover of the resource is performed while the cluster framework is stopping the resource. Suppose that the LDom resource starts out on node A and the administrator executes a switchover to node B. While the cluster framework is stopping the resource on node A, LDom live migration is actually migrating the LDom from node A to node B.
The problem (which previously existed) with this approach arose if another resource group declared a strong negative (SN) affinity for the LDom resource's group or if load limits were in use. The LDom is actually migrating onto the target node before the cluster framework knows that it has started there, so the cluster framework will not have had a chance to evict any resource group(s) due to strong negative affinities or load limits. There will be an interval of time in which the hard loadlimit or strong negative affinity will be violated. This may cause overload of the node, and consequent failure of the LDom live migration attempt.
With Oracle Solaris Cluster 4.2 we eliminate the above said problem by providing a mechanism in the cluster framework by which excess workload can be migrated off of the target node of the switchover before executing the live migration of the LDom. This is accomplished by making use of the following two enhancements:
A new resource property Pre_evict which ensures seamless live migration of LDom. The Pre_evict property allows the cluster framework to move excess workload off of the target node before the switchover begins, which then allows the live migration to succeed more often.
A new scha_resourcegroup_get query tag SCHA_TARGET_NODES which allows a data service to perform the live migration by informing it of the target node to be used for the migration.
For the convenience of the users / administrators, the HA-LDom agent in Oracle Solaris Cluster 4.2 has the Pre_evict set to True by default and the agent is enhanced to take advantage of these new API (Pre_evict and SCHA_TARGET_NODES query) features.
For more information, refer the following links
pre-evict resource property:
- Harish Mallya Oracle Solaris Cluster