Cluster impacts summary
 |
| |
|
This impact class contains impacts related to HP-UX clusters.
These impacts may require you to modify your source code.
|
HP's MC/ServiceGuard is a high-availability solution requiring each node
to have its own system disk and operating system. As a result, each operating
system copy must be individually maintained with up-to-date copies of
site-specific information, including configuration data and user information.
The advantage of having a large number of nodes in an MC/ServiceGuard
cluster is the ability to finely distribute the workload of a failed server
onto a large number of other servers. This strategy is likely to produce
small variations in response time in the case of a single-server outage.
Two-node clusters typically present unique problems in terms of preserving
data integrity when there is a node failure or a network failure. The cluster
nodes exchange "I'm alive" or heartbeat messages over the LANs connecting
them. If a message is not received from a node that is a cluster member
within a certain period of time, the cluster attempts to reform itself. If
there is a network failure, or if the network is too congested so that
heartbeats are missed, both nodes will attempt to reform the cluster. In
addition, if you had two very large nodes in a cluster and one failed, the
cluster would be left with just 50 percent of its total processing power.
With a three-node or larger cluster, the portion of the original cluster
with the most number of nodes is allowed to reform the cluster, and the nodes
in the disconnected subnet are removed from cluster membership. In a two-node
cluster, each cluster node represents exactly half of the cluster. If both
nodes were to succeed in their attempts to reform the cluster, data
corruption to the shared disks could easily occur since both nodes could have
applications running that need to access the same data.
To prevent this, HP introduced the concept of a cluster lock as a
tiebreaker. The cluster lock is implemented by means of a lock disk, an area
on a shared disk volume. Only one node at a time is allowed to own the
cluster lock. The node that wins the cluster lock is allowed to reform the
cluster with itself as the only node. The other cluster node is not allowed
to reform the cluster and, without network communications, cannot rejoin the
cluster.
The current release of MC/ServiceGuard introduces a quorum server as an
alternative to the cluster lock disk. The quorum server is a process that
runs on an HP-UX 11i host that can be connected to multiple MC/ServiceGuard
clusters and, for each cluster, it can provide the functions of a lock disk.
A quorum server may be used in clusters with any number of nodes, and an
individual quorum server may provide tie-breaking services to as many as 50
clusters containing up to a total of 100 nodes.
The MC/ServiceGuard interconnect between servers is made up of two
industry-standard LANs. Using LAN connections in a cluster yields a maximum
of 640 MB/sec communication between the servers in an MC/ServiceGuard
cluster.
For Oracle RAC clusters, we recommend using Hyperfabric, HP's high-speed
interconnect, where the message passing done by the database will benefit
from the high-speed link.
Note: To include only Clusters impacts when running the
scansummary and scandetail tools, use the option:
+C CL
To exclude these impacts, use the option:
-C CL
For more information on filtering output in reports, see Filtering the Output.
|
| 
