15.2.3 Parallel Test Harness

The parallel (or concurrent) test harness is enabled by the Automake option parallel-tests. It features automatic collection of the test scripts output in .log files, concurrent execution of tests with make -j, specification of inter-test dependencies, lazy reruns of tests that have not completed in a prior run, and hard errors for exceptional failures.

This harness is still somewhat experimental and may undergo changes in order to satisfy additional portability requirements.

The parallel test harness operates by defining a set of make rules that run the test scripts listed in TESTS, and, for each such script, save its output in a corresponding .log file and its results (and other “metadata”, see API for Custom Test Drivers) in a corresponding .trs (as in Test ReSults) file. The .log file will contain all the output emitted by the test on its standard output and its standard error. The .trs file will contain, among the other things, the results of the test cases run by the script.

The parallel test harness will also create a summary log file, TEST_SUITE_LOG, which defaults to test-suite.log and requires a .log suffix. This file depends upon all the .log and .trs files created for the test scripts listed in TESTS.

As with the serial harness above, by default one status line is printed per completed test, and a short summary after the suite has completed. However, standard output and standard error of the test are redirected to a per-test log file, so that parallel execution does not produce intermingled output. The output from failed tests is collected in the test-suite.log file. If the variable ‘VERBOSE’ is set, this file is output after the summary. For best results, the tests should be verbose by default now.

Each couple of .log and .trs files is created when the corresponding test has completed. The set of log files is listed in the read-only variable TEST_LOGS, and defaults to TESTS, with the executable extension if any (see Support for executable extensions), as well as any suffix listed in TEST_EXTENSIONS removed, and .log appended. Results are undefined if a test file name ends in several concatenated suffixes. TEST_EXTENSIONS defaults to .test; it can be overridden by the user, in which case any extension listed in it must be constituted by a dot, followed by a non-digit alphabetic character, followed by any number of alphabetic characters. For example, ‘.sh’, ‘.T’ and ‘.t1’ are valid extensions, while ‘.x-y’, ‘.6c’ and ‘.t.1’ are not.

For tests that match an extension .ext listed in TEST_EXTENSIONS, you can provide a custom “test runner” using the variable ext_LOG_COMPILER (note the upper-case extension) and pass options in AM_ext_LOG_FLAGS and allow the user to pass options in ext_LOG_FLAGS. It will cause all tests with this extension to be called with this runner. For all tests without a registered extension, the variables LOG_COMPILER, AM_LOG_FLAGS, and LOG_FLAGS may be used. For example,

TESTS = foo.pl bar.py baz
TEST_EXTENSIONS = .pl .py
PL_LOG_COMPILER = $(PERL)
AM_PL_LOG_FLAGS = -w
PY_LOG_COMPILER = $(PYTHON)
AM_PY_LOG_FLAGS = -v
LOG_COMPILER = ./wrapper-script
AM_LOG_FLAGS = -d

will invoke ‘$(PERL) -w foo.pl’, ‘$(PYTHON) -v bar.py’, and ‘./wrapper-script -d baz’ to produce foo.log, bar.log, and baz.log, respectively. The foo.trs, bar.trs and baz.trs files will be automatically produced as a side-effect.

It’s important to note that, differently from what we’ve seen for the serial test harness (see Parallel Test Harness), the AM_TESTS_ENVIRONMENT and TESTS_ENVIRONMENT variables cannot be use to define a custom test runner; the LOG_COMPILER and LOG_FLAGS (or their extension-specific counterparts) should be used instead:

## This is WRONG!
AM_TESTS_ENVIRONMENT = PERL5LIB='$(srcdir)/lib' $(PERL) -Mstrict -w

## Do this instead.
AM_TESTS_ENVIRONMENT = PERL5LIB='$(srcdir)/lib'; export PERL5LIB;
LOG_COMPILER = $(PERL)
AM_LOG_FLAGS = -Mstrict -w

By default, the test suite harness will run all tests, but there are several ways to limit the set of tests that are run:

• You can set the TESTS variable. For example, you can use a command like this to run only a subset of the tests:

  env TESTS="foo.test bar.test" make -e check
  

  Note however that the command above will unconditionally overwrite the test-suite.log file, thus clobbering the recorded results of any previous testsuite run. This might be undesirable for packages whose testsuite takes long time to execute. Luckily, this problem can easily be avoided by overriding also TEST_SUITE_LOG at runtime; for example,

  env TEST_SUITE_LOG=partial.log TESTS="..." make -e check
  

  will write the result of the partial testsuite runs to the partial.log, without touching test-suite.log.

• You can set the TEST_LOGS variable. By default, this variable is computed at make run time from the value of TESTS as described above. For example, you can use the following:

  set x subset*.log; shift
  env TEST_LOGS="foo.log $*" make -e check
  

  The comments made above about TEST_SUITE_LOG overriding applies here too.

• By default, the test harness removes all old per-test .log and .trs files before it starts running tests to regenerate them. The variable RECHECK_LOGS contains the set of .log (and, by implication, .trs) files which are removed. RECHECK_LOGS defaults to TEST_LOGS, which means all tests need to be rechecked. By overriding this variable, you can choose which tests need to be reconsidered. For example, you can lazily rerun only those tests which are outdated, i.e., older than their prerequisite test files, by setting this variable to the empty value:

  env RECHECK_LOGS= make -e check
  

• You can ensure that all tests are rerun which have failed or passed unexpectedly, by running make recheck in the test directory. This convenience target will set RECHECK_LOGS appropriately before invoking the main test harness.

In order to guarantee an ordering between tests even with make -jN, dependencies between the corresponding .log files may be specified through usual make dependencies. For example, the following snippet lets the test named foo-execute.test depend upon completion of the test foo-compile.test:

TESTS = foo-compile.test foo-execute.test
foo-execute.log: foo-compile.log

Please note that this ordering ignores the results of required tests, thus the test foo-execute.test is run even if the test foo-compile.test failed or was skipped beforehand. Further, please note that specifying such dependencies currently works only for tests that end in one of the suffixes listed in TEST_EXTENSIONS.

Tests without such specified dependencies may be run concurrently with parallel make -jN, so be sure they are prepared for concurrent execution.

The combination of lazy test execution and correct dependencies between tests and their sources may be exploited for efficient unit testing during development. To further speed up the edit-compile-test cycle, it may even be useful to specify compiled programs in EXTRA_PROGRAMS instead of with check_PROGRAMS, as the former allows intertwined compilation and test execution (but note that EXTRA_PROGRAMS are not cleaned automatically, see The Uniform Naming Scheme).

The variables TESTS and XFAIL_TESTS may contain conditional parts as well as configure substitutions. In the latter case, however, certain restrictions apply: substituted test names must end with a nonempty test suffix like .test, so that one of the inference rules generated by automake can apply. For literal test names, automake can generate per-target rules to avoid this limitation.

Please note that it is currently not possible to use $(srcdir)/ or $(top_srcdir)/ in the TESTS variable. This technical limitation is necessary to avoid generating test logs in the source tree and has the unfortunate consequence that it is not possible to specify distributed tests that are themselves generated by means of explicit rules, in a way that is portable to all make implementations (see Make Target Lookup in The Autoconf Manual, the semantics of FreeBSD and OpenBSD make conflict with this). In case of doubt you may want to require to use GNU make, or work around the issue with inference rules to generate the tests.