Using Serviceguard Extension for RAC, 6th Edition, April 2008

communicates over link level protocol (DLPI) and supported over Serviceguard
heartbeat subnet networks, including primary and standby links.
The most common network configurations is to have all interconnect traffic for cluster
communications to go on a single heartbeat network that is redundant so that
Serviceguard monitors the network and resolves interconnect failures by cluster
reconfiguration.
The following are situations when it is not possible to place all interconnect traffic on
a single network:
• RAC GCS (cache fusion) traffic may be very high, so an additional dedicated
heartbeat network for Serviceguard needs to be configured.
• Some networks, such as Infiniband, are not supported by CFS/CVM, so the
CSS-HB/RAC-IC traffic may need to be on a separate network that is different
from SG-HB network.
• SGeFF requires a dual Serviceguard heartbeat network, and CSS-HB/RAC-IC does
not support multiple networks for HA purposes.
• In a multiple database configuration, RAC-IC traffic of one database may interfere
with RAC-IC traffic of another database, so the RAC-IC traffic of databases may
need to be separated.
In the above cases, you will see a longer time to recover some network failures beyond
those protected by primary and standby, unless Serviceguard is configured to monitor
the network.
A failure of CSS-HB/RAC-IC network in such configuration does not force Serviceguard
to reform the cluster. If Serviceguard is not configured to monitor the network, Oracle
will take at least CSS misscount time interval to resolve the network failure. The default
value of CSS misscount in SGeRAC configurations is 600 seconds.
To avoid longer recovery times, manage Oracle Clusterware and RAC-DB instances
using Serviceguard multi-node packages. In addition, configure the
CLUSTER_INTERCONNECT_SUBNET package configuration parameter (as done with
a standard SUBNET package configuration parameter) in the respective multi-node
packages to monitor the CSS-HB/RAC-IC networks.
Planning Storage for Oracle Cluster Software
Oracle Cluster Software requires shared storage for the Oracle Cluster Registry (OCR)
and a vote device. Automatic Storage Management can not be used for the OCR and
vote device since these files must be accessible before Oracle Cluster Software starts.
The minimum required size for the OCR is 100MB and for the vote disk is 20 MB. The
Oracle OCR and vote device can be created on supported shared storage, including
SLVM logical volumes, CVM raw volumes, and CFS file systems (on HP-UX releases
that support Veritas CFS and CVM; see “About Veritas CFS and CVM from Symantec”
(page 25)).
Planning Storage for Oracle Cluster Software 41