|
A cluster is a group of application server instances
that run simultaneously, which act like a single system, enabling
high availability and load balancing. Within a cluster, multiple
server instances can run on the same computer (known as a vertical cluster)
or can be located on different computers (known as a horizontal cluster),
or they can form a combination of both horizontal and vertical clusters. With
clustering, client work can be distributed across several nodes
instead of handled by a single application server.
In a clustered configuration, application server instances are
server members of the cluster, all of which must have identical
application components deployed on them. However, other than the
configured applications, cluster members do not have to share any
other configuration parameters. For example, you can cluster multiple
server instances on one computer, with a single instance on another computer,
provided they are all running the same version of application server.
By clustering, you can achieve one or more of the following benefits.
How you implement clustering determines which benefits are achieved:
Failover
Load balancing
Scalability
2.1.1 FailoverFailover allows one or more application server instances
to act as backup to a failed application server instance and resume
processing the task, therefore enabling another application server
to carry on processing. However, if an application server fails
during a transaction, the backup application server does not recover
the state of the failed instance. If a server fails when a user
enters data into a form, for example, the data may have to be reentered.
2.1.2 Load balancingLoad balancing is a technique used to distribute
work across a number of systems so that no single device is overwhelmed.
If one server starts to get congested or overloaded, requests are
forwarded to another server that has more capacity.
2.1.2.1 Application server load balancingApplication server load balancing is useful for
managing the load between application servers. Application servers
can be configured to use a weighted round-robin routing policy that
ensures a balanced routing distribution based on the set of server
weights that are assigned to the members of a cluster. Configuring all
servers in the cluster to have the same weight produces a load distribution
in which all servers receive approximately the same number of requests.
Weighting some servers more heavily sends more requests to these
servers than to those that are weighted less heavily.
Preferred routing configurations can also be configured to ensure,
for example, that only cluster members on that node are selected
(using the round-robin weight method) and cluster members on remote
nodes are selected only if a local server is not available. Application
server load balancing is best used when balancing is needed between
tiers.
2.1.2.2 Web server load balancingWeb server load balancing is useful for queuing
and throttling requests. For the Apache HTTP Server, the most commonly
used method for load balancing is Round-Robin DNS.
Round-Robin DNS is a relatively simple method of load
balancing, where a domain name system (DNS) server provides a name
to address resolution and is always involved when a host name is
included in a URL. A Round-Robin DNS server can resolve one single
host name into multiple IP addresses such that requests for a single
URL (containing a host name) actually reference different web servers.
The client requests a name resolution for the host name but, in
fact, receives different IP addresses, therefore spreading the load
among the web servers. In a simple configuration, the Round-Robin
DNS server cycles through the list of available servers.
2.1.3 Scalability and availabilityScalability in a cluster means that an administrator
can increase the capacity of the application dynamically to meet
the demand without interrupting or negatively impacting service.
Clusters allow administrators to remove nodes from a cluster in
order to upgrade components, such as memory, or to add nodes to
the cluster without bringing down the cluster itself.
|
|
|
