Matlab Write Performance Tests

Writing MATLAB Performance Tests

Write performance tests for MATLAB code using the matlab.perftest.TestCase framework. This framework provides statistically rigorous timing with automatic warmup, multiple samples, and outlier handling.

When to Use

• User wants to write repeatable performance tests for their MATLAB code

• User needs to benchmark functions with statistical confidence

• User wants to detect performance regressions over time

• User is setting up continuous performance monitoring

• User asks how to use runperf or matlab.perftest.TestCase

When NOT to Use

• User wants a quick one-off timing (use timeit instead — see matlab-optimize-performance)

• User wants to optimize existing code (use matlab-optimize-performance)

• User wants to measure memory usage (use matlab-optimize-memory)

• User wants to profile code to find bottlenecks (use matlab-optimize-performance, Step 2)

Framework: matlab.perftest.TestCase

All performance tests subclass matlab.perftest.TestCase and use measurement boundaries to control what gets timed.

Basic Template

classdef MyFeaturePerformanceTest < matlab.perftest.TestCase

    properties (MethodSetupParameter)
        DataSize = struct('Small', 100, 'Medium', 1000, 'Large', 10000)
    end

    properties
        inputData
    end

    methods (TestMethodSetup)
        function setupData(testCase, DataSize)
            % ALL setup outside the measurement boundary
            testCase.inputData = randn(DataSize, 1);
        end
    end

    methods (Test)
        function testMyFunction(testCase)
            data = testCase.inputData;
            while testCase.keepMeasuring
                result = myFunction(data);
            end
            testCase.verifyNotEmpty(result);
        end
    end
end

Running Performance Tests

% Run with statistical rigor (automatic sample size)
results = runperf('MyFeaturePerformanceTest');

% View results
disp(results)

% Fixed sample count (faster, less statistical power)
import matlab.perftest.TimeExperiment;
suite = testsuite('MyFeaturePerformanceTest');
experiment = TimeExperiment.withFixedSampleSize(4);
results = run(experiment, suite);

Measurement Boundaries

The framework offers three ways to control what gets measured:

1. keepMeasuring — Needed when code is fast (<10ms)

Automatically loops the code until enough samples are collected. Required for sub-10ms operations to achieve statistical rigor; works at any speed but adds overhead for slower code where startMeasuring/stopMeasuring is preferred:

function testFastFunction(testCase)
    data = testCase.inputData;
    while testCase.keepMeasuring
        result = fastFunction(data);
    end
    testCase.verifyNotEmpty(result);
end

2. startMeasuring/stopMeasuring — For precise control

Use when you need setup between iterations or want to exclude specific code:

function testWithBoundaries(testCase)
    data = testCase.inputData;
    % Pre-computation (NOT measured)
    preparedData = preprocess(data);

    testCase.startMeasuring();
    result = functionUnderTest(preparedData);
    testCase.stopMeasuring();

    % Verification (NOT measured)
    testCase.verifyEqual(size(result), [100 1]);
end

3. No boundary — Entire method is measured

The whole Test method body is timed. Use only when the entire method IS the workload:

function testSlowFunction(testCase, DataSize) %#ok<INUSD>
    data = testCase.inputData;
    result = slowFunction(data);
    testCase.verifyNotEmpty(result);
end

Parameterization

Parameterize tests to measure across different input sizes or configurations.

MethodSetupParameter — When setup uses the parameter

properties (MethodSetupParameter)
    DataSize = struct('Small', 100, 'Medium', 1000, 'Large', 10000)
end

methods (TestMethodSetup)
    function setupData(testCase, DataSize)
        testCase.inputData = randn(DataSize, 1);
    end
end

TestParameter — When only test methods use the parameter

properties (TestParameter)
    Algorithm = {'chol', 'lu', 'qr'}
end

methods (Test)
    function testSolve(testCase, Algorithm)
        ...
    end
end

Critical gotcha: Do NOT use TestParameter for properties consumed by TestMethodSetup. MATLAB will error with "Define 'X' as a MethodSetupParameter." If your setup method needs the parameter, it must be MethodSetupParameter.

Combining Both — Size in setup, algorithm in test

properties (MethodSetupParameter)
    DataSize = struct('Small', 100, 'Medium', 1000, 'Large', 10000)
end

properties (TestParameter)
    Algorithm = {'chol', 'lu', 'qr'}
end

methods (TestMethodSetup)
    function setupData(testCase, DataSize)
        testCase.inputData = randn(DataSize);
    end
end

methods (Test)
    function testSolve(testCase, Algorithm)
        A = testCase.inputData' * testCase.inputData; % SPD matrix
        data = A;
        while testCase.keepMeasuring
            result = decomposition(data, Algorithm);
        end
        testCase.verifyNotEmpty(result);
    end
end

This produces 9 test points (3 sizes × 3 algorithms).

Lean Setup — Critical Rule

ALL setup must be outside the measurement boundary:

Never include setup/teardown inside the measured region. This inflates timing and adds noise.

Copy properties to local variables before measuring. Don't access testCase.PropertyName inside the measurement boundary — it measures the matlab.perftest.TestCase indexing overhead. Assign to a local variable outside the boundary instead:

% Correct: local variable assigned before measurement
data = testCase.inputData;
while testCase.keepMeasuring
    result = myFunction(data);
end

methods (TestMethodSetup)
    function setupData(testCase, DataSize)
        rng(42, 'twister');  % Deterministic data
        testCase.inputData = randn(DataSize, 1);
    end
end

Noise and Duration

Target Duration

• Each test point should take >10ms to execute

• If too fast, increase data size or use keepMeasuring

• Sub-millisecond tests produce unreliable results

Reducing Noise

• Use keepMeasuring for fast operations

• Pre-allocate all data in setup

• Suppress all output ('Display', 'off')

• Avoid file I/O in the measured region

• Avoid drawnow, pause, or GUI operations

• Seed RNG in setup for deterministic data

Interpreting Results

results = runperf('MyPerformanceTest');

% Access timing statistics
for i = 1:numel(results)
    samples = results(i).Samples.MeasuredTime;
    fprintf('%s: median=%.4fs, std=%.4fs, CV=%.1f%%\n', ...
        results(i).Name, median(samples), std(samples), ...
        100*std(samples)/mean(samples));
end

A coefficient of variation (CV) above 10% indicates noisy results — revisit your test setup.

Anti-Patterns

Comparing Results Over Time

Save and compare results to detect regressions:

% Save baseline
baselineResults = runperf('MyPerformanceTest');
save('perfBaseline.mat', 'baselineResults');

% Later: compare against baseline
currentResults = runperf('MyPerformanceTest');
load('perfBaseline.mat');

for i = 1:numel(currentResults)
    baseMed = median(baselineResults(i).Samples.MeasuredTime);
    currMed = median(currentResults(i).Samples.MeasuredTime);
    ratio = currMed / baseMed;
    status = "OK";
    if ratio > 1.2
        status = "REGRESSION";
    elseif ratio < 0.8
        status = "IMPROVEMENT";
    end
    fprintf('%s: %.4fs -> %.4fs (%.2fx) %s\n', ...
        currentResults(i).Name, baseMed, currMed, ratio, status);
end

Test Granularity

When asked to write performance tests, consider which level is appropriate:

When to use each:

• Unit: Function has multiple expensive operations and you need to isolate which one regressed

• System: Default choice — measures the function as users call it

• Workflow: Validates that unit optimizations translate to real-world speedups; catches cross-function bottlenecks

If unsure, generate system-level tests first, then ask whether unit-level decomposition or workflow-level benchmarks are needed.

Workflow Benchmarks (Application-Level)

For end-to-end workflow benchmarks, the structure differs from unit/system tests:

• No parameterization — one representative problem size (that's what unit tests are for)

• Realistic data — production-scale inputs, not synthetic

• Single test method — the whole workflow is one testpoint

• startMeasuring/stopMeasuring — always (never keepMeasuring)

• Setup in TestClassSetup — load data once, not per method

• Target 0.5–5s — sub-0.5s is noise-dominated; >30s slows iteration

See references/tExampleWorkflow.m for the complete template.

Quality Checklist

Verify after generating each test class:

• [ ] Setup is outside the measurement boundary (TestMethodSetup)

• [ ] Properties copied to local variables before measuring

• [ ] Each test point exceeds 10ms (or uses keepMeasuring if faster)

• [ ] RNG seeded for deterministic data

• [ ] No console output, plotting, or file I/O in measured region

• [ ] Parameterized by data size

• [ ] Single API/operation per test method (not bundling multiple)

• [ ] MethodSetupParameter used (not TestParameter) when setup consumes the parameter

Reference Files

• references/FeaturePerformanceTest.m — Performance test class template with all measurement patterns

• references/tExampleWorkflow.m — Workflow-level benchmark template (application-level)

• references/simulink-template.md — Performance test template for Simulink model benchmarks

Copyright 2026 The MathWorks, Inc.