iOS8 Day-by-Day :: Day 6 :: Profiling Unit Tests

Written by Sam Davies

Updated 9 Apr 2015: This post has been updated to Swift 1.2

This post is part of a daily series of posts introducing the most exciting new parts of iOS8 for developers – #iOS8DayByDay. To see the posts you’ve missed check out the index page, but have a read through the rest of this post first!

To enjoy each of the 39 posts all in one place, you can now also download the iOS8 Day-by-Day eBook free from our website! 


Unit tests are widely accepted as incredibly useful tools to aid writing better software. They allow you to establish the behaviour of your code before putting it in front of end-users, and to ensure that everything works as expected in conditions which aren’t necessarily default.

In Xcode 5 OCUnit was replaced with XCUnit, but there was very little noticeable difference to end users. In Xcode 6, XCUnit has had a few features added to it, one of which is the ability to measure the performance of a piece of code. This is really helpful to find sticking points in your app, and then iteratively improve the performance.

The app which accompanies this article demonstrates how you can use the new performance testing features of XCUnit to improve the speed of a simple algorithm. In this instance it takes a naive implementation of a moving average, measures the performance, before creating a new implementation which has better performance.

The code is available on github at, so grab it. Any questions then hit me up on twitter @iwantmyrealname.

Measuring Test Performance

The concept behind the new performance measuring functionality in XCTest is you can specify a block which the test runner will run repeatedly, measuring the time it takes to complete each time.

Each time you run your test suite, this performance measurement will be repeated, and a comparison to the last time will be displayed. This allows you to spot regressions in performance easily, and to iterate on improving performance.

Implementing this is really simple – XCTestCase now has method called measureBlock on it. This takes a void block/closure and times it:

func testSampleTestPerformance() {
  self.measureBlock() {
    // Code under test

This might become a little clearer with a specific example.

Moving average example

Moving average is a common problem in computing, and is basically an implementation of a finite impulse response filter with unit weights. There are many ways to implement it, but the simplest is as follows:

func calculateMovingAverage(data: Double[]) -> Double[] {
  // Create an array to store the result
  var result = Double[]()

  // Now perform the calculation
  for i in 0...(data.count - windowSize) {
    let slice = data[i..(i+windowSize)]
    let partialSum = slice.reduce(0) { $0 + $1 }
    result.append(Double(partialSum) / Double(windowSize))
  return result

The moving average is the mean of a window which moves along the input array – which is exactly what is being implemented here. This method is part of the NaiveMovingAverageCalculator class, which we can then put under test.

class MovingAverageTests: XCTestCase {
  let calculatorCreator : () -> MovingAverageCalculator = { return NaiveMovingAverageCalculator() }
  var calculator: MovingAverageCalculator = NaiveMovingAverageCalculator()

  override func setUp() {
    // Reset the calculator
    calculator = calculatorCreator()

  func testMovingAveragePerformance() {
    // This is an example of a performance test case.
    calculator.windowSize = 1000
    self.measureBlock() {
      // Put the code you want to measure the time of here.
      let randomArray = self.RandomDouble(10000)
      let result = self.calculator.calculateMovingAverage(randomArray)
      XCTAssertEqual(result.count, 9000)

  func RandomDouble(length: Int) -> Array<Double> {
    var result = Double[]()
    for i in 1..length {
    return result

There is one test method in the above code sample (testMovingAveragePerformance()), which includes the key measureBlock() method. Inside this block we are using the RandomDouble() utility method to create an array of random doubles of length 10000, before requesting the calculator to generate the moving average.

When you run this test suite then that test will be run 10 times. Once completed you can click the tag on the right hand side to see how it performed:

Initial Result

You can see a column chart of the time for each test run – giving you an indication of the variance. You can also see the mean time and its standard deviation. Setting a baseline saves this result, and then subsequent test runs will be compared to this. In fact, in the above picture, the baseline was set and you can see that a subsequent test run had only a tiny difference in performance, as you would expect.

Improving the moving average

Having decided that 6 seconds is an incredibly long time to perform a moving average, you can go ahead an attempt to improve upon it. The result of this is the BetterMovingAverageCalculator class, which has the following calculateMovingAverage() method:

func calculateMovingAverage(data: Double[]) -> Double[] {
  var result = Double[]()

  var currentSum = data[0..windowSize].reduce(0) { $0 + $1 }
  result.append(Double(currentSum) / Double(windowSize))
  for i in 0..(data.count - windowSize) {
    // Remove the first entry
    currentSum -= data[i]
    // And add the new one
    currentSum += data[i + windowSize]
    // Save it off
    result.append(Double(currentSum) / Double(windowSize))

  return result

This approach is a common improvement to a moving average – keeping a running sum means that the same sum operation is not repeated for every output value.

To see how this fairs in comparison to the original implemention, you can just update the test class as follows:

let calculatorCreator : () -> MovingAverageCalculator = { return BetterMovingAverageCalculator() }
var calculator: MovingAverageCalculator = BetterMovingAverageCalculator()

Now running the test suite again will cause the following balloon to appear:

Small improvement

If you click on it, you can reveal more result data:

Large improvement

That’s quite a huge difference – well over 95%. It’d be great if it was that easy to optimise all code.

If you’re happy with this new code, then you can reset the baseline to the new calculator, and you’ll have changed your benchmark. The results can be checked in to source control and shared with your team, so you’re all working to the same baselines.


There is a popular adage in software engineering, that you shouldn’t optimise prematurely, so you probably shouldn’t be throwing this new functionality all over your tests.

However, there are a couple of cases that this new tools really caters for:

  • Ensuring that performance regressions are noticed, caught and dealt with or accepted.
  • Making optimisation a lot easier when the time comes.

It can be used in conjuction with a more fine-grained profiler, such as that provided by instruments to really assist with improving the performance of your app.

The code for this moving average project is available on github at Feel free to fork it, and write a more optimal moving average calculator 🙂 As ever, questions / comments below or via @iwantmyrealname.