Now comes the part where you actually run your capacity-planning test. The next question is, "How do I load the users to simulate the load?" The best way to do this is to try to emulate how users hit the server during peak hours. Does that user load happen gradually over a period of time? If so, a ramp-up-style load should be used, where x number of users are added ever y seconds. Or, do all of the users hit the system in a very short period of time all at once? If that is the case, a flat run should be used, where all of the users are simultaneously loaded onto the server. These different styles will produce different results that are not comparable. For instance, if a ramp-up run (Figure 5) is done and you find out that the system can support 5,000 users with a response time of 4 seconds or less, and then you follow that test with a flat run with 5,000 users, you'll probably find that the average response time of the system with 5,000 users is higher than 4 seconds. This is an inherent inaccuracy in ramp-up runs that prevents them from pinpointing the exact number of concurrent users that a system can support. For a portal application, for example, this inaccuracy is amplified as the size of the portal grows and as the size of the cluster is increased.

This is not to say that ramp-up tests should not be used. Ramp-up runs are great if the load on the system is slowly increased over a long period of time. This is because the system will be able to continually adjust over time. If a fast ramp-up is used, the system will lag and artificially report a lower response time than what would be seen if a similar number of users were being loaded during a flat run. So, what is the best way to determine capacity? Taking the best of both load types and running a series of tests will yield the best results. For example, using a ramp-up run to determine the range of users that the system can support should be used first. Then, once that range has been determined, doing a series of flat runs at various concurrent user loads within that range can be used to more accurately determine the capacity of the system (Figure 6).

Soak Tests
A soak test is a straightforward type of performance test. Soak tests are long-duration tests with a static number of concurrent users that test the overall robustness of the system. These tests will show any performance degradations over time via memory leaks, increased garbage collection (GC), or other problems in the system. The longer the test, the more confidence in the system you will have. It is a good idea to run this test twice - once with a fairly moderate user load (but below capacity so that there is no execute queue - Figure 7), and once with a high user load (so that there is a positive execute queue - Figure 8).

These tests should be run for several days to really get a good idea of the long-term health of the application. Make sure that the application being tested is as close to real world as possible with a realistic user scenario (how the virtual users navigate through the application), testing all of the features of the application. Ensure that all the necessary monitoring tools are running so that problems will be accurately detected and tracked down later (Figure 9).

Peak-Rest Tests
Peak-rest tests are a hybrid of the capacity-planning ramp-up-style tests and soak tests. The goal here is to determine how well the system recovers from a high load (such as one during peak hours of the system), goes back to near idle, and then goes back up to peak load and back down again.

The best way to implement this test is to do a series of quick ramp-up tests followed by a plateau (determined by the business requirements), and then a dropping off of the load. A pause in the system should then be used, followed by another quick ramp-up - then you repeat the process. A couple of things can be determined from this: Does the system recover on the second "peak" and each subsequent peak to the same level (or greater) than the first peak? And does the system show any signs of memory or GC degradation over the course of the test? The longer this test is run (repeating the peak/idle cycle over and over), the better idea you'll have of what the long-term health of the system looks like.

Summary
This article has described several approaches to performance testing. Depending on the business requirements, development cycle, and life cycle of the application, some tests will be better suited than others for a given organization. In all cases though, you should ask some fundamental questions before going down one path or another. The answers to these questions will then determine how to best test the application. These questions are:

• How repeatable do the results need to be?
• How many times do you want to run and rerun these tests?
• What stage of the development cycle are you in?
• What are your business requirements?
• What are your user requirements?
• How long do you expect the live production system to stay up between maintenance downtimes?
• What is the expected user load during an average business day?