Skip to content

Latest commit

 

History

History
295 lines (231 loc) · 11.6 KB

README.md

File metadata and controls

295 lines (231 loc) · 11.6 KB
Raw

googletest-mpi-listener

The goals of this project are:

  • to augment the googletest framework with listener(s) that yield sensible output for testing MPI-based, distributed-memory parallel software

  • to keep this software as low-maintenance as possible

  • ideally, to merge some version of this functionality into Google Test

License

This software is triple-licensed under the BSD-3, MIT, and Apache 2.0 licenses, and uses some example code from Google Test, which has the following license that must be duplicated in its entirety, per its terms:

 Copyright 2005, Google Inc.  All rights reserved.

 Redistribution and use in source and binary forms, with or without
 modification, are permitted provided that the following conditions are
 met:

     * Redistributions of source code must retain the above copyright
 notice, this list of conditions and the following disclaimer.
     * Redistributions in binary form must reproduce the above
 copyright notice, this list of conditions and the following disclaimer
 in the documentation and/or other materials provided with the
 distribution.
     * Neither the name of Google Inc. nor the names of its
 contributors may be used to endorse or promote products derived from
 this software without specific prior written permission.

 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
 "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
 LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
 A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
 OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
 SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
 LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
 DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
 THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
 (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
 OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

The MPIWrapperPrinter class design was influenced by similar code in mxx, which is distributed under an Apache 2.0 license.

Requirements

  • CMake
  • Google Test
  • a standards-conforming MPI-1.x implementation (e.g., MPICH, Open MPI, MVAPICH, Intel MPI)
  • a C++ compiler; Google Test 1.8.1 and earlier require a C++98-standard-compliant compiler, whereas later versions require a C++11-standard-compliant compiler

Building the example code/unit tests

  1. Change to the build directory:

pushd build

  1. Run CMake:

cmake ..

You should see some CMake output indicating compiler detection, etc.

  1. Run the output file(s) generated by CMake. For example, assuming the CMake generator used is "Unix Makefiles", run

make

to build the unit tests.

  1. Run the unit tests:

mpi-minimal-listener-unit-tests mpi-wrapper-listener-unit-tests

Usage

Please read the Google Test Primer and Google Test Advanced Guide for background before reading the remainder of this section.

Usage of these listeners is illustrated by the examples in test/mpi-minimal-listener-unit-tests.cpp and test/mpi-wrapper-listener-unit-tests.cpp, both of which include some very simple-minded MPI unit tests that are used for basic testing. To use this listener, you will need to write your own int main(int argc, char** argv) function -- you cannot use the stock gtest_main supplied with Google Test. Using a custom main function enables us to:

  • initialize MPI
  • add an Environment object that will finalize MPI when main terminates
  • remove the stock Google Test TestEventListener, which emits output on tests
  • add a TestEventListener that will emit sensible output (that is, rank-ordered output to stdout) for MPI-based unit tests

all of which is necessary, and not possible using the stock gtest_main function.

Directions for writing tests can be found in Google Test Primer and Google Test Advanced Guide. The design assumption in this library is that these tests will be executed in distributed fashion using the MPI distributed-memory parallel programming model (that is, a shared-nothing, process-local memory address space). This assumption implies that tests will be run by all MPI processes, or disabled on all processes. Conditionals can be used to execute parts (or all) of the body of a test on a given MPI process or set of processes.

After writing tests, your int main(int argc, char** argv) function should look like the test runner in test/mpi-wrapper-listener-unit-tests.cpp:

int main(int argc, char** argv) {
  // Filter out Google Test arguments
  ::testing::InitGoogleTest(&argc, argv);

  // Initialize MPI
  MPI_Init(&argc, &argv);

  // Add object that will finalize MPI on exit; Google Test owns this pointer
  ::testing::AddGlobalTestEnvironment(new GTestMPIListener::MPIEnvironment);

  // Get the event listener list.
  ::testing::TestEventListeners& listeners =
      ::testing::UnitTest::GetInstance()->listeners();

  // Remove default listener: the default printer and the default XML printer
  ::testing::TestEventListener *l =
        listeners.Release(listeners.default_result_printer());

  // Adds MPI listener; Google Test owns this pointer
  listeners.Append(
      new GTestMPIListener::MPIWrapperPrinter(l,
                                              MPI_COMM_WORLD)
      );
  // Run tests, then clean up and exit. RUN_ALL_TESTS() returns 0 if all tests
  // pass and 1 if some test fails.
  int result = RUN_ALL_TESTS();

  return 0;  // Run tests, then clean up and exit
}

Comments in this example describe what each line does. The ordering is also important, and should be preserved if you augment this example with additional code.

Most users should probably use a MPIWrapperPrinter wrapped around the default Google Test event listener, because this configuration provides colorized pretty-printing and works correctly with Google Test's XML and JSON output generators. Please note, however, that this colorization will add console color code characters to the output, which may be undesirable if parsing console output redirected to files. Please also note that, due to the architecture of Google Test and the use of Google Test's ADD_FAILURE_AT macro within MPIWrapperPrinter, adding more than one instance of MPIWrapperPrinter to the collection of active listeners is likely to yield garbled output.

You could also use an alternative int main(int argc, char** argv) for your test runner, taken from test/mpi-minimal-listener-unit-tests.cpp:

int main(int argc, char** argv) {
    // Filter out Google Test arguments
    ::testing::InitGoogleTest(&argc, argv);

    // Initialize MPI
    MPI_Init(&argc, &argv);

    // Add object that will finalize MPI on exit; Google Test owns this pointer
    ::testing::AddGlobalTestEnvironment(new MPIEnvironment);

    // Get the event listener list.
    ::testing::TestEventListeners& listeners =
    ::testing::UnitTest::GetInstance()->listeners();

    // Remove default listener
    delete listeners.Release(listeners.default_result_printer());

    // Adds MPI listener; Google Test owns this pointer
    listeners.Append(new MPIMinimalistPrinter);

    // Run tests, then clean up and exit. RUN_ALL_TESTS() returns 0 if all tests
    // pass and 1 if some test fails.
    int result = RUN_ALL_TESTS();

    return 0;
}

Output from MPIminimalistPrinter is not colorized, and does not work well with JSON or XML output. Consider this listener if you need more easily parseable output and you are willing to sacrifice pretty printing and pretty XML or JSON unit test reports.

Design considerations

The most important design consideration was to write something portable, quickly. This meant using MPI-1.x calls (because supercomputers sometimes run older MPI implementations), and relying on the public API calls of a portable unit testing library that is currently being maintained. As a result of these decisions:

  • this software communicates unit test results to MPI process 0 and then writes them to standard out (because MPI I/O is defined in MPI-2)

  • consequently, this software will have an output bottleneck at extreme scale, because the output is, to borrow from Bill Gropp's MPI I/O slides (see slide 3), "worse than sequential"

  • this software is header-only, to make vendoring it into software packages no more painful than vendoring in Google Test

The current setup should be usable for small numbers of MPI processes -- I've tested it on 256 MPI processes and it seems to work fine. If there is a need to write infrastructure for testing on 100,000 MPI processes, then an MPI I/O-based TestEventListener makes more sense, and would relax the output bottleneck. This new code should be written in a second header to isolate MPI-2-conforming code from MPI-1-conforming code. Although I haven't done it yet, I suspect this modification could be done in days.

If you want to hack on this code, the most important things to note are:

  • ::testing::TestEventListener::OnTestPartResult is only called after assertion failures or invocation of the SUCCESS() macro, so it may not be called on some MPI processes, and there should be no communication in this function

  • In TestEventListener::OnTestEnd, although ::testing::TestInfo has methods to query ::testing::TestPartResult objects within it, when I called those methods, the information I wanted (namely, ::testing::TestPartResult objects) was no longer present, which is why I instead store that information within a class member instead

As stated above, I aim to keep this software low-maintenance, because this package was written in a couple days in order to make the parallel software development for my work easier. I'm happy to accept bug-fixes, and I'm willing to discuss or consider feature requests/additions, but I'd like any feature to be easy to maintain. I do not get paid to write testing frameworks, which is why any responses to bug-fixes, feature requests, documentation improvements, or questions may be delayed; it is also why I want this software to be low-maintenance. In general, even though I intend to respond to correspondence about this software (I would like it to improve), users should assume I may not get back to them in a timely fashion, or at all.

TODO

  • Automate library testing instead of running tests manually. Using BATS to run executables and test captured console output is probably a reasonable solution. Captured console output could be compared to "known good" output, though it might be useful to shim in versions of the wrapper tester that can set the file name (and line number) to known, repeatable values.

Release number: LLNL-CODE-739313