As I've blogged about before, SOAs are not unlike any other distributed computing systems, and thus designing a performance model should be nothing too new. At this point we understand exactly how each service behaves under an increasing load, and we have enough data to plug into a model. Now, it's just a matter of building a model.

There are very expensive performance monitoring and simulation tools that are for sale in the market, but sometimes the least expensive and most simple tools work best...in many cases, just a spreadsheet will do. For our purposes, we need to consider both information and behavior in the context of performance, as well as core features of an SOA.

Information Movement Modeling, typically asynchronous in nature, means we're attempting to simulate how information moves from point to point, point to many points, or many points to many points. Based on the information we accumulated we know the:

Information production rate from a service
Information consumption rate from a service

For example, an instance of a service is able to produce 52 messages (or similar groupings of information) per second...the source service. An instance of a service is able to consume 34 messages per second...the target service. This is a simple point-to-point relationship, but keep in mind that multi-points to multi-points are always possible, and those are a bit more complex to model since you have to determine patterns of movement between multiple points vs. all messages produced by a single service that are consumed by another.

Moreover, keep in mind transformation and routing latency is typically an issue here as well, and needs to be modeled along with consumption and production. You should have test data from these services, but the performance of transformation and routing services will be largely dependent upon the complexities of the transformations and logic associated with the routing. What many do when creating performance models is to model very complex, complex, and simple transformation scenarios, and the percentages of each.

Service Invocation Modeling, typically more synchronous in nature, means we're attempting to determine the number of times a service is able to provide a behavior (application function) in an instance of time, typically a second.

For instance, you may have a service that provides a risk calculation for the insurance business, and is perhaps abstracted into several different applications (composites). We know through testing that each composite can invoke the service up to 100 times a second before it hits a saturation point, meaning the performance of the service quickly diminishes as additional load is placed upon it. This saturation number plugs into the model, as well as the number of applications that are abstracting this service. You have to model all of these services in the same way.

Models are important because they allow you to predict performance under changing needs without having to actually build and test the system. Models, of course, are not perfect, and you must constantly adjust assumptions and modeling information as you learn more about the behavior of the architecture.

Posted by Dave Linthicum on August 22, 2006 02:07 PM