Service Oriented Architectures are increasingly being used to implement high-performance and real-time systems. Traditional systems operate in "human real-time," where human patience is the limit. Increasingly, however, systems operate in "computer real-time," where the only limits are imposed by the operational speed of the computers and networks.

For example, next-generation Air Traffic Management systems are being developed to accommodate the huge increase in air traffic and link the operational capabilities of agencies such as the Federal Aviation Authority, the Department of Defense (DOD) and the Department of Homeland Security (DHS). These systems require higher information bandwidth (to track more aircraft or more complex "free-flight" trajectories) as well as much lower latencies or delays on the information (to detect flight abnormalities quickly). Similar demands are being made in healthcare, SCADA, network monitoring, energy distribution, transportation, and other critical infrastructure systems.

Best-of-Breed SOA Components
Demanding real-time applications require best-of-breed service-oriented foundational components. There are three kinds of foundational components in a SOA system: A messaging fabric/bus, information transformation/processing engines, and persistence/storage services (see Figure 1). Often these components are integrated into an Enterprise Service Bus (ESB) and hosted in a J2EE Application Server.

Of these foundational components, the Messaging Fabric/Bus is the most critical, since it mediates all interactions between components.

Low-performance SOA systems may use HTTP as the "messaging fabric/bus" to exchange messages between components. This approach is only suitable for non-demanding applications: HTTP isn't reliable, has limited bandwidth, introduces very high latencies, and can't buffer and queue messages and deliver them to systems that are either temporarily unavailable or join at a later time.

The solution is to deploy a high-performance messaging middleware such as RTI Data-Distribution Service, IBM WebSphere MQ, TIBCO, or SonicMQ. These middleware platforms have been developed with scalability and performance in mind. However, they each employ a different architecture optimized for different application scenarios.

Why Does Messaging Performance Matter?
The requirements and expectations of computer-speed real-time far exceed traditional human-speed real-time. Whereas in systems with a human in the loop, real-time meant that the information was available anywhere from fractions of a second to few seconds in the computer-to-computer world, real-time means decisions should be made in milliseconds or even microseconds.

Computer real-time puts more stringent requirements on the messaging infrastructure: Each processing and storage component must get hundreds of thousands of messages/events per second with microsecond or at worst millisecond latencies. This means that the messaging middleware must be able to deliver millions of messages a second system-wide.

And the capacity of the messaging fabric must be able to scale with the capacity of the underlying hardware and not impose any limits beyond those of the underlying hardware infrastructure (CPU speed, cores, speed, and bandwidth of the network) itself. As the CPU and network speeds increase those systems able to take advantage of what the hardware provides will deliver a competitive advantage. In an automated trading system, for instance, the critical metric is not the absolute time it takes to make a decision, but rather whether a decision is taken and the trade executed before competitive trades occur. The same is true in a combat management system.

One final aspect of computer real-time SOA systems is their "inverted performance-load utility curve." This means that the ability to respond in a timely manner becomes more important when the system is experiencing a high load. In a normal utility curve, such as in human real-time systems, degraded performance is acceptable under an increased load. This is because human expectations and patience adjust based on the circumstances (e.g., they understand that on a peak holiday period they may endure longer hold times when calling to make a flight reservation). In contrast, computer-speed real-time systems often have the opposite demands. It is precisely at the moments of high load when the "most critical action" is taking place and it is then when it is most critical to deliver top performance (e.g., it is precisely when market action is heavy that trading decisions must be made quickly).

The differences between human-speed real-time systems and computer-speed real-time systems are summarized in Table 1.

Selecting Messaging Middleware in SOA Systems
Messaging middleware is the key enabler of real-time SOA. However, there are many options. How can you choose the best messaging middleware for a particular real-time SOA system? Five areas distinguish messaging middleware: architecture, quality of service (QoS) control and filters, performance-boosting technologies, real-time determinism, and metrics.

Architecture
The four basic architectures employed by messaging middleware are: centralized (hub-and-spoke), clustered, federated, and peer-to-peer. (see Figure 2)

A centralized (hub-and-spoke) architecture routes every message though a single server that implements the message "service," contains all the message queues, and brokers every message.
A clustered architecture uses a collection of servers and assigns to each responsibility for some of the messages (like ownership of some of the message queues or topics). Each message is relayed by a server but not all messages use the same server.

A federated architecture also uses a collection of servers, but it uses them as a "resource pool" where queues may appear in multiple servers, and messages may be brokered by one or more servers.

A peer-to-peer architecture doesn't employ any brokers in the critical path. Messages are routed directly from the sender to the receiver.

Each has strengths and weaknesses. Centralized is easiest to administer and can provide stronger transactional semantics but suffers from poor performance, reduced tolerance to faults, and doesn't scale. Clustered is more scalable than centralized but also has reduced fault tolerance and can only offer good performance in a grid environment with all the clients co-located close to the grid. Federated is more scalable, but suffers from higher latency and jitter as each message is brokered by at least two servers. P2P offers the best scalability, performance, lowest jitter, and highest resilience, but is difficult for vendors to implement and offers limited transactional support.

As demands become more real-time, the need for performance, predictability, and balance tips the scale towards P2P architecture. That's why, for example, demanding networks like Voice over IP and Video over IP (like Skype) use peer-to-peer designs.

Quality of Service Control & Filters
QoS control is critical to deliver timely data with low latency and high throughput. CPU, memory, and network bandwidth resources must be shared among all the traffic. However, not all traffic requires the same bandwidth or has the same urgency or level or criticality. Without QoS control, the application has no way to differentiate different traffic classes and their corresponding constraints. As a consequence, the middleware can't make intelligent decisions, prioritize traffic, or ultimately meet the application requirements.