mpi-pytest

Pytest plugin that lets you run tests in parallel with MPI.

mpi-pytest provides:

• A parallel marker indicating that the test must be run under MPI.
• A parallel_assert function for evaluating assertions in a deadlock-safe way.

parallel marker

Writing a parallel test simply requires marking the test with the parallel marker:

@pytest.mark.parallel(nprocs=5)  # run in parallel with 5 processes
def test_my_code_on_5_procs():
    ...

@pytest.mark.parallel(5)  # the "nprocs" kwarg is optional
def test_my_code_on_5_procs_again():
    ...

@pytest.mark.parallel  # run in parallel with the default number of processes (3)
def test_my_code_on_3_procs():
    ...

@pytest.mark.parallel()  # the brackets are optional
def test_my_code_on_3_procs_again():
    ...

One can also mark a test with a sequence of values for nprocs:

@pytest.mark.parallel(nprocs=[1, 2, 3])  # run in parallel on 1, 2 and 3 processes
def test_my_code_on_variable_nprocs():
    ...

@pytest.mark.parallel([1, 2, 3])  # again the "nprocs" kwarg is optional
def test_my_code_on_variable_nprocs_again():
    ...

If multiple numbers of processes are requested then the tests are parametrised and renamed to, in this case, test_my_code_on_variable_nprocs[nprocs=1], test_my_code_on_variable_nprocs[nprocs=2] and test_my_code_on_variable_nprocs[nprocs=3].

When running the code with these parallel markers, mpi-pytest adds extra markers to each test to allow one to select all tests with a particular number of processors. For example, to select all parallel tests on 3 processors, one should run:

$ mpiexec -n 3 pytest -m parallel[3]

Serial tests - those either unmarked or marked @pytest.mark.parallel(1) - can be selected by running:

$ pytest -m "not parallel or parallel[1]"

Forking mode

mpi-pytest can be used in one of two modes: forking or non-forking. The former works as follows:

1. The user calls pytest (not mpiexec -n <# proc> pytest). This launches the "parent" pytest process.
2. This parent pytest process collects all the tests and begins to run them.
3. When a test is found with the parallel marker, rather than executing the function as before, a subprocess is forked calling mpiexec -n <# proc> pytest this_specific_test_file.py::this_specific_test. This produces <# proc> 'child' pytest processes that execute the test together.
4. If this terminates successfully then the test is considered to have passed.

This is convenient for development for a number of reasons:

• The plugin composes better with other pytest plugins like pytest-xdist.
• It is not necessary to wrap pytest invocations with mpiexec calls, and all parallel and serial tests can be run at once.

There are however a number of downsides:

• Only one mainstream MPI distribution (MPICH) supports nested calls to MPI_Init. If your 'parent' pytest process initialises MPI (for instance by executing from mpi4py import MPI) then this will cause non-MPICH MPI distributions to crash.
• Forking a subprocess can be expensive since a completely fresh Python interpreter is launched each time.
• Sandboxing each test means that polluted global state at the end of a test cannot be detected.

Non-forking mode

With these significant limitations in mind, mpi-pytest therefore also supports a non-forking mode. To use it, one simply needs to wrap the pytest invocation with mpiexec, no additional configuration is necessary. For example, to run all of the parallel tests on 2 ranks one needs to execute:

$ mpiexec -n 2 pytest -m parallel[2]

parallel_assert

Using regular assert statements can be unsafe in parallel codes. Consider the code:

@pytest.mark.parallel(2)
def test_something():
    # this will only fail on *some* ranks
    assert COMM_WORLD.rank == 0

    # this will hang
    COMM_WORLD.barrier()

One can see that failing assertions on some ranks but not others will violate SPMD and lead to deadlocks. To avoid this, mpi-pytest provides a parallel_assert function used as follows:

from pytest_mpi import parallel_assert

@pytest.mark.parallel(2)
def test_something():
    # this will fail on *all* ranks
    parallel_assert(lambda: COMM_WORLD.rank == 0)
    ...

Configuration

mpi-pytest respects the environment variable PYTEST_MPI_MAX_NPROCS, which defines the maximum number of processes that can be requested by a parallel marker. If this value is exceeded an error will be raised.