As the number of clients or the number of connections grows, you might need to scale a message service to eliminate bottlenecks or to improve performance. In general, you can scale a message service both vertically (increasing the number of client applications that are supported by a single broker) and horizontally (distributing client applications among a number of interconnected brokers).
Vertical scaling usually requires adding more processing power for a broker and by expanding available resources. You can do this by adding more processors or memory, by switching to a shared thread model, or by running the Java VM in 64 bit mode.
Horizontal scaling is generally achieved using a broker cluster. While it is possible to scale horizontally by simply redistributing clients among additional brokers that are not in a cluster, this approach is appropriate only if your messaging operations can be divided into independent work groups. However, if producer clients must produce messages to be consumed by consumer clients connected to remote brokers, then brokers must work collectively, as part of a broker cluster, to achieve horizontal scaling.
In a broker cluster, each broker is connected to every other broker in the cluster. Brokers can reside on the same host, but more often are distributed across a network. Each broker can route messages from producers to which it is directly connected to consumers that are connected to remote brokers in the cluster.
If you are using the point-to-point domain, you can scale the consumer side by allowing multiple consumers to access a queue. This is a Message Queue feature (the JMS specification defines messaging behavior in the case of only one consumer accessing a queue). When multiple consumers access a queue, the load-balancing among them takes into account each consumer’s capacity and message processing rate.