In software engineering, performance testing is testing that is performed to determine how fast some aspect of a system performs under a particular workload.
Performance testing can serve different purposes. It can demonstrate that the system meets performance criteria. It can compare two systems to find which performs better. Or it can measure what parts of the system or workload cause the system to perform badly. In the diagnostic case, software engineers use tools such as profilers to measure what parts of a device or software contribute most to the poor performance or to establish throughput levels (and thresholds) for maintained acceptable response time.
In performance testing, it is often crucial (and often difficult to arrange) for the test conditions to be similar to the expected actual use.
Performance testing technology employs one or more PCs to act as injectors – each emulating the presence or numbers of users and each running an automated sequence of interactions (recorded as a script, or as a series of scripts to emulate different types of user interaction) with the host whose performance is being tested. Usually, a separate PC acts as a test conductor, coordinating and gathering metrics from each of the injectors and collating performance data for reporting purposes. The usual sequence is to ramp up the load – starting with a small number of virtual users and increasing the number over a period to some maximum.
The test result shows how the performance varies with the load, given as number of users vs response time. Various tools, including Compuware Corporation's QACenter Performance Edition, are available to perform such tests. Tools in this category usually execute a suite of tests which will emulate real users against the system. Sometimes the results can reveal oddities, e.g., that while the average response time might be acceptable, there are outliers of a few key transactions that take considerably longer to complete – something that might be caused by inefficient database queries, etc.
Performance testing can be combined with stress testing, in order to see what happens when an acceptable load is exceeded –does the system crash? How long does it take to recover if a large load is reduced? Does it fail in a way that causes collateral damage?
Performance testing is frequently not performed against a specification, i.e. no one will have expressed what is the maximum acceptable response time for a given population of users. However, performance testing is frequently used as part of the process of performance profile tuning. The idea is to identify the “weakest link” – there is inevitably a part of the system which, if it is made to respond faster, will result in the overall system running faster. It is sometimes a difficult task to identify which part of the system represents this critical path, and some test tools come provided with (or can have add-ons that provide) instrumentation that runs on the server and reports transaction times, database access times, network overhead, etc. which can be analysed together with the raw performance statistics. Without such instrumentation one might have to have someone crouched over Windows Task Manager at the server to see how much CPU load the performance tests are generating. There is an apocryphal story of a company that spent a large amount optimising their software without having performed a proper analysis of the problem. They ended up rewriting the system’s ‘idle loop’, where they had found the system spent most of its time, but even having the most efficient idle loop in the world obviously didn’t improve overall performance one iota!
Performance testing almost invariably identifies that it is parts of the software (rather than hardware) that contribute most to delays in processing users’ requests.
Performance testing can be performed across the web, and even done in different parts of the country, since it is known that the response times of the internet itself vary regionally. It can also be done in-house, although routers would then need to be configured to introduce the lag what would typically occur on public networks.
It is always helpful to have a statement of the likely peak numbers of users that might be expected to use the system at peak times. If there can also be a statement of what constitutes the maximum allowable 95 percentile response time, then an injector configuration could be used to test whether the proposed system met that specification.
Tasks to undertake
Tasks to perform such a test would include:
* Analysis of the types of interaction that should be emulated and the production of scripts to do those emulations
* Decision whether to use internal or external resources to perform the tests.
* set up of a configuration of injectors/controller
* set up of the test configuration (ideally identical hardware to the production platform), router configuration, quiet network (we don’t want results upset by other users), deployment of server instrumentation.
* Running the tests – probably repeatedly in order to see whether any unaccounted for factor might affect the results.
* Analysing the results, either pass/fail, or investigation of critical path and recommendation of corrective action.