Overview of Performance Testing Framework

The performance test interface leverages the script, function, and class-based unit testing interfaces. You can perform qualifications within your performance tests to ensure correct functional behavior while measuring code performance. Also, you can run your performance tests as standard regression tests to ensure that code changes do not break performance tests.

Determine Bounds of Measured Code

This table indicates what code is measured for the different types of tests.

Type of Test	What Is Measured	What Is Excluded
Script-based	Code in each section of the script	Code in the shared variables section Measured estimate of the framework overhead
Function-based	Code in each test function	Code in the following functions: setup, setupOnce, teardown, and teardownOnce Measured estimate of the framework overhead
Class-based	Code in each method tagged with the Test attribute	Code in the methods with the following attributes: TestMethodSetup, TestMethodTeardown, TestClassSetup, and TestClassTeardown Shared fixture setup and teardown Measured estimate of the framework overhead
Class-based deriving from matlab.perftest.TestCase and using startMeasuring and stopMeasuring methods	Code between calls to startMeasuring and stopMeasuring in each method tagged with the Test attribute	Code outside of the startMeasuring/stopMeasuring boundary Measured estimate of the framework overhead
Class-based deriving from matlab.perftest.TestCase and using the keepMeasuring method	Code inside each keepMeasuring-while loop in each method tagged with the Test attribute	Code outside of the keepMeasuring-while boundary Measured estimate of the framework overhead

Types of Time Experiments

You can create two types of time experiments.

• A frequentist time experiment collects a variable number of measurements to achieve a specified margin of error and confidence level. Use a frequentist time experiment to define statistical objectives for your measurement samples. Generate this experiment using the runperf function or the limitingSamplingError static method of the TimeExperiment class.

• A fixed time experiment collects a fixed number of measurements. Use a fixed time experiment to measure first-time costs of your code or to take explicit control of your sample size. Generate this experiment using the withFixedSampleSize static method of the TimeExperiment class.

This table summarizes the differences between the frequentist and fixed time experiments.

	Frequentist time experiment	Fixed time experiment
Warm-up measurements	5 by default, but configurable through TimeExperiment.limitingSamplingError	0 by default, but configurable through TimeExperiment.withFixedSampleSize
Number of samples	Between 4 and 256 by default, but configurable through TimeExperiment.limitingSamplingError	Defined during experiment construction
Relative margin of error	5% by default, but configurable through TimeExperiment.limitingSamplingError	Not applicable
Confidence level	95% by default, but configurable through TimeExperiment.limitingSamplingError	Not applicable
Framework behavior for invalid test result	Stops measuring a test and moves to the next one	Collects specified number of samples

Write Performance Tests with Measurement Boundaries

If your class-based tests derive from matlab.perftest.TestCase instead of matlab.unittest.TestCase, then you can use the startMeasuring and stopMeasuring methods or the keepMeasuring method multiple times to define boundaries for performance test measurements. If a test method has multiple calls to startMeasuring, stopMeasuring and keepMeasuring, then the performance testing framework accumulates and sums the measurements. The performance testing framework does not support nested measurement boundaries. If you use these methods incorrectly in a Test method and run the test as a TimeExperiment, then the framework marks the measurement as invalid. Also, you still can run these performance tests as unit tests. For more information, see Test Performance Using Classes.

Run Performance Tests

There are two ways to run performance tests:

• Use the runperf function to run the tests. This function uses a variable number of measurements to reach a sample mean with a 0.05 relative margin of error within a 0.95 confidence level. It runs the tests 5 times to warm up the code and between 4 and 256 times to collect measurements that meet the statistical objectives.

• Generate an explicit test suite using the testsuite function or the methods in the TestSuite class, and then create and run a time experiment.

  • Use the withFixedSampleSize method of the TimeExperiment class to construct a time experiment with a fixed number of measurements. You can specify a fixed number of warm-up measurements and a fixed number of samples.

  • Use the limitingSamplingError method of the TimeExperiment class to construct a time experiment with specified statistical objectives, such as margin of error and confidence level. Also, you can specify the number of warm-up measurements and the minimum and maximum number of samples.

You can run your performance tests as regression tests. For more information, see Test Performance Using Classes.

Understand Invalid Test Results

In some situations, the MeasurementResult for a test result is marked invalid. A test result is marked invalid when the performance testing framework sets the Valid property of the MeasurementResult to false. This invalidation occurs if your test fails or is filtered. Also, if your test incorrectly uses the startMeasuring and stopMeasuring methods of matlab.perftest.TestCase, then the MeasurementResult for that test is marked invalid.

When the performance testing framework encounters an invalid test result, it behaves differently depending on the type of time experiment:

• If you create a frequentist time experiment, then the framework stops measuring for that test and moves to the next test.

• If you create a fixed time experiment, then the framework continues collecting the specified number of samples.