Hello,
The JBoss clustering saga with Bill Burke and I continues.
A first version of clustering for Stateful Session Beans (SFSB) is now
available on the CVS main branch.
Features:
=========
- Sessions of each node are replicated "in-memory" to some other nodes of
the cluster
- Cluster is automagically subdivided in sub-clusters (i.e.
sub-partitions),
which optimal size is user-definable.
- Sessions are replicated in these sub-partitions and not in the whole
cluster.
For a 2-nodes cluster, it does not change anything but for a 10-nodes
cluster,
it would have meant that each node had to support the session-load
(memory,
synchro, ..) of the whole cluster i.e. of 10 nodes => doesn't scale. But
with
this scheme, each node only support the load of a few other nodes (1 or 2
ideally)
- When a node dies or starts, a new set of sub-partitions is automatically
re-computed. The result intends to minimise the number of nodes that need
to move in another sub-partition (i.e. state-transfer) and avoid
singleton groups
(singletons are added to another sub-partition even if it makes the
sub-partition
become temporarily bigger than the user-defined optimal size)
- If state doesn't change after an invocation made on a SFSB, its state is
not replicated
- When a node join a sub-partition, the state is compressed before being
sent to it.
Basic Setup:
============
As for any JBoss clustering feature, to use them you first need to copy
jboss-all/cluster/etc/cluster-service.xml in your deploy directory.
To identify a bean as clustered, just add a <clustered>True</clustered>
element to your jboss.xml SFSB definition. Example:
<jboss>
<enterprise-beans>
<session>
<ejb-name>My_Clustered_SFSB</ejb-name>
<jndi-name>My_Clustered_SFSB</jndi-name>
--> <clustered>True</clustered>
</session>
</enterprise-beans>
</jboss>
Implementation Details:
=======================
- The in-memory state management is done thanks to a new "service": a
HASessionState
- The HASessionState uses the a HAPartition (the default one) to
communicate with other nodes
- When the cluster composition changes, only one node computes the new
sub-partition topology
that will be shared between nodes.
- sub-partitions use a new, dynamically created, HAPartition. Only nodes of
a particular sub-partition
(for example 2 nodes out of 10) use the same HAPartition, thus, in this
case we
would have 1 "global" HAPartition shared by all nodes for general
communication and
5 HAPartition, each used by two different nodes for their session state
in-memory replication. But
at the end, each node only has 2 HAPartition: the global one and its
dedicated.
- To make the link between the SFSB container and the HASessionState
service, we use a new
PersistenceManager: StatefulHASessionPersistenceManager.
- A new interceptor (StatefulHASessionSynchronisationInterceptor) is added
to the
interceptor chain, just after the StatefulSessionInstanceInterceptor. Its
goal is to
call the PersistenceManager and ask it to synchronise the state of the
SFSB after any
remote interface method call or a SFSB creation (home.create (...)). In
the future,
we may add the possibility, in jboss.xml, to indicate which methods
change the state of the
bean and thus optimise many read-only calls. This feature could also be
very useful for EB.
Detailed Configuration:
=======================
A HASessionState service is started this way:
<mbean code="org.jboss.ha.hasessionstate.server.HASessionStateService"
name="JBOSS-SYSTEM:service=HASessionState">
<attribute name="JndiName">/HASessionState/Default</attribute>
<attribute name="SubPartitionProperties"> - not indicated, check
JavaGroups.com documentation - </attribute>
<attribute name="MainHaPartitionName">DefaultPartition</attribute>
<attribute name="SubPartitionSize">2</attribute>
<attribute name="SubPartitionCreationWaitTime">5000</attribute>
<attribute name="BeanCleaningDelay">1800000</attribute>
<attribute
name="TopologyComputerClassName">org.jboss.ha.hasessionstate.server.HASessio
nStateTopologyComputerImpl</attribute>
</mbean>
All attributes are optional (and thus can be omitted). Default values are
indicated.
Meaning:
- JndiName: JNDI name/path under which the HASessionState will be made
available
- SubPartitionProperties: JavaGroups property string used when creating
Sub-Partition
- MainHaPartitionName: name of the HAPartition used by all nodes of the
cluster
- SubPartitionSize: number of nodes per sub-partition. May not be respected
(i.e. bigger)
to avoid singleton groups.
- SubPartitionCreationWaitTime: When starting a new sub-partition, a member
start first to avoid singleton group
(JavaGroups needs to be extended to avoid this). This value indicates how
much miliseconds other slave nodes need
to wait before connecting to the group.
- BeanCleaningDelay: when bean state is no more used it is cleaned after
this delay. This is needed because
the cache may not know its existence if a node inherited its presence
from another node.
- TopologyComputerClassName: name of the class used to compute a new
topology after a group composition change.
A SFSB can contain some more tags in jboss.xml:
<jboss>
<enterprise-beans>
<session>
<ejb-name>My_Clustered_SFSB</ejb-name>
<jndi-name>My_Clustered_SFSB</jndi-name>
<clustered>True</clustered>
<cluster-config>
<partition-name>Default</partition-name>
<home-load-balance-policy></home-load-balance-policy>
<bean-load-balance-policy></bean-load-balance-policy>
<session-state-manager-jndi-name></session-state-manager-jndi-name>
</cluster-config>
</session>
</enterprise-beans>
</jboss>
Meaning:
- partition-name: Name of the HAPartition to be used for cluster
communication
- home-load-balance-policy: name of the Java class to be used as a policy
for the home proxy (default: load-balance)
- bean-load-balance-policy: name of the Java class to be used as a policy
for the bean proxy (default: first-available)
- session-state-manager-jndi-name: JNDI name of the HASessionState to be
used by our PersistenceManager for in-memory
replicated session state.
This is alpha code. It is really new code and need deep testing and
improvement. Feedback is really welcome.
Cheers,
Sacha
_______________________________________________
JBoss-user mailing list
[EMAIL PROTECTED]
https://lists.sourceforge.net/lists/listinfo/jboss-user
