Thursday, September 22, 2011

Reboot-less node fencing in Oracle Clusterware 11g Release 2

There has been several improvements in the node eviction and split brain scenarios with Oralce 11G R2 RAC.

Oracle Clusterware uses a STONITH (Shoot The Other Node In The Head) comparable fencing algorithm to ensure data integrity in cases, in which cluster integrity is
endangered and split-brain scenarios need to be prevented. In case of Oracle Clusterware, this means that a local process enforces the removal of one or more nodes from the cluster (fencing).
Until Oracle Clusterware 11g Release 2, Patch Set One ( the fencing of a node was performed by a “fast reboot” of the respective server. A “fast reboot” in this context summarizes a shutdown and restart procedure that does not wait for any IO to finish or for file systems to synchronize on shutdown. With Oracle Clusterware 11g Release 2, Patch Set One ( this mechanism has been changed in order to prevent such a reboot as much as possible.

Already with Oracle Clusterware 11g Release 2 this algorithm was improved so that failures of certain, Oracle RAC-required subcomponents in the cluster do not necessarily cause an
immediate fencing (reboot) of a node. Instead, an attempt is made to clean up the failure within the cluster and to restart the failed subcomponent. Only, if a cleanup of the failed component appears to be unsuccessful, a node reboot is performed in order to force a cleanup.

With Oracle Clusterware 11g Release 2, Patch Set One ( further improvements were made so that Oracle Clusterware will try to prevent a split-brain without rebooting the node. It thereby implements a standing requirement from those customers, who were requesting to preserve the node and to prevent a reboot, since the node runs applications not managed by Oracle Clusterware, which would otherwise be forcibly shut down by the reboot of a node.

With the new algorithm and when a decision is made to evict a node from the cluster, Oracle Clusterware will first attempt to shutdown all resources on the machine that was chosen to be the subject of an eviction. Especially IO generating processes are killed and it is ensured that those processes are completely stopped before continuing. If, for some reason, not all resources can be stopped or IO generating processes cannot be stopped completely, Oracle Clusterware will still perform a reboot or use IPMI to forcibly evict the node from the cluster.
If all resources can be stopped and all IO generating processes can be killed, Oracle Clusterware will shut itself down on the respective node, but will attempt to restart after the stack has been stopped. The restart is initiated by the Oracle High Availability Services Daemon, which has been introduced with Oracle Clusterware 11g Release 2.


Post a Comment


ORA-BLOG. Copyright 2008 All Rights Reserved Revolution Two Church theme by Brian Gardner Converted into Blogger Template by Bloganol dot com