Parallel PHPUnit
Join the DZone community and get the full member experience.
Join For FreePHPUnit is the standard testing framework for PHP code: always available through Pear or Composer, following xUnit conventions on tests and providing many features from grouping to code coverage to logging of results. There's even an extension for running Selenium tests (that I maintain), which allows you to run browser-based tests.
Parallelism
What PHPUnit lacks is parallelism: tests are run one after the other, usually in the same process. This means that when you have more available resources, such as a multicore CPU, some of the computational power is not used as the PHPUnit process may reach 100% utilization while the other cores are not working at all. This is not surprising.
PHP does not have multithreading capabilities, but it can start new processes at the OS level. So many developers have came up with the same idea: starting multiple PHPUnit processes, each working on a different subset of tests, and aggregate the results. This could theoretically give you a N times speedup when working with N different cores, for example passing from 10 minutes on a single core to 2'30'' on a quad core CPU.
Caveats
Of course the cost of coordinating different processes is always going to be present, so we will never reach the theoretical speedup. I'll report later in this article some simulations.
The most important constraints come from the design of our test suites. I can only think of two categories of tests as easily parallelizable:
- unit tests, which only use memory and CPU as resources and not disk or other external infrastructure.
- Selenium tests, which run against a live HTTP server that must be able to serve multiple requests without race conditions if your application is going to work.
By design, these two kinds of tests are always capable to run in parallel. However, other intensive and long-running tests such as end-to-end tests and integration ones usually conflict with each other:
public function setUp() { $this->pdo = new PDO(...); $this->pdo->query('DELETE FROM users'); } public function testUsersCanBeAddedWithAllDetails() { $this->request->post('/users', ...); $this->assertEquals(1, $this->request->get('/users')); } public function testUsersCanBeDeletedByAnAdmin() { $this->insertAnUser(); $this->assertEquals(1, $this->request->get('/users')); $this->request->delete('/users', ...); $this->assertEquals(0, $this->request->get('/users')); }
These API-based tests are never going to run in parallel (on the same machine) when written in this way, due to the race condition on the users table. If you have a slow suite that you want to speed up, chances are that it contains many end-to-end tests like these. Some of these tests can be isolated with RDBMS transactions, but it's difficult for black-box tests to intervene on the transaction isolation inside the application.
The tools
PHPUnit is due to support parallelism since 2007, but it has never come up in the package and pull requests for the feature have never been accepted. So we have to resort to external tools.
Probably the most complete tool working on top of PHPUnit is Paratest , which has two peculiarities:
- It uses reflection to compose a list of all of your tests instead of grepping *Test.php files.
- It reads PHPUnit JUnit-format logs to aggregate results from different tests, which makes it difficult to break than tools that parse the output of the command itself.
The only limitations of it are that it poses some stronger constraints on your tests, for example they have to follow the PSR-0 convention. However, it delegates much to PHPUnit and lets you use many of the same command line switches such as --configuration and --bootstrap.
Experiments
To experiment with Paratest, I created a simulated unit test suite that only works with the CPU. I have 10 test of the form :
public function testExample() { for ($i = 0; $i < 1024*1024; $i++) { $this->assertTrue(true); } }
I then tried to run this suite on a dual core CPU, on a physical (not virtual) home machine. I have tried different options, too:
- vanilla PHPUnit, serial execution
- Paratest, single process execution (to find out if it has an high overhead).
- Paratest with 2 parallel processes.
These are the results:
[21:13:25][giorgio@Desmond:~/paratestexample]$ ./compare.sh PHPUnit 3.7.13-5-g6937c46 by Sebastian Bergmann. .......... Time: 03:04, Memory: 3.25Mb OK (10 tests, 20971520 assertions) Running phpunit in 1 process with /home/giorgio/paratestexample/vendor/bin/phpunit .......... Time: 03:01, Memory: 3.75Mb OK (10 tests, 20971520 assertions) Running phpunit in 2 processes with /home/giorgio/paratestexample/vendor/bin/phpunit .......... Time: 02:15, Memory: 3.75Mb OK (10 tests, 20971520 assertions)
The difference is a 25% decrease in total time, which is really worth investigating further.
Conclusions
I'm going to experiment more with Paratest to see if it's possible to speed up also batteries of end-to-end tests, for example making different processes using different databases or offloading the PHPUnit commands to different machines.
Opinions expressed by DZone contributors are their own.
Comments