TensorFlow tests may require to be run on an even number of cores.

[akesandgren]

When doing a CPU-only build of TF 2.6.0 in a batch job with 7 cores it repeatedly failed //tensorflow/python/data/kernel_tests:interleave_test When running it on 14 cores the test passed.

I haven't had time to verify this fully but we might need to have the easyblock check that it has an even number of cores available so to not fail this test.

@Flamefire thoughts on this?

[boegel]

Maybe we can force the tests to run on an even number of cores? So if 7 cores are available, run on 6? That only leaves a single-core installation of TensorFlow as a problematic case.

If we can confirm this though, seems weird (but I guess I shouldn't be surprised).

[Flamefire]

Can I have some log files with the exact error to maybe find out what is wrong? And maybe report it as an issue to TF upstream to see what they say about this and for us to have something to reference in the upcoming workaround

[akesandgren]

Do you want the complete (debug) build log or just the snippet around interleave_test?

[branfosj]

Extra tests from me:

• 6 fail
• 7 fail
• 8 success
• 9 success

It looks like a minimum of 8 cores is needed.

[akesandgren]

Yeah, might be the problem, running on 17 cores works, so it's not a pure even/odd problem, nor a prime number problem.

[branfosj]

Error seen (from TensorFlow/2.6.0/foss-2021a/tmpkuca27g9-bazel-tf/0db3977216c8a32952b67b0996300dd3/execroot/org_tensorflow/bazel-out/k8-opt/testlogs/tensorflow/python/data/kernel_tests/interleave_test/shard_1_of_24/test.log:

======================================================================
FAIL: testInterleaveDataset_test_mode_graph_tfapiversion_2_blocklength_3_cyclelength_None_inputvalues_4567_numparallelcalls_None (__main__.InterleaveTest)
InterleaveTest.testInterleaveDataset_test_mode_graph_tfapiversion_2_blocklength_3_cyclelength_None_inputvalues_4567_numparallelcalls_None
testInterleaveDataset_test_mode_graph_tfapiversion_2_blocklength_3_cyclelength_None_inputvalues_4567_numparallelcalls_None(mode='graph', tf_api_version=2, block_length=3, cycle_length=None, input_values=array([4, 5, 6, 7]), num_parallel_calls=None)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/rds/bear-apps/devel/eb-sjb-up/EL8/EL8-cas/software/TensorFlow/2.6.0-foss-2021a/lib/python3.9/site-packages/absl/testing/parameterized.py", line 314, in bound_param_test
    return test_method(self, **testcase_params)
  File "/dev/shm/build-branfosj-admin/branfosj-admin-up/TensorFlow/2.6.0/foss-2021a/tmpzwc1qwnh-bazel-tf/0db3977216c8a32952b67b0996300dd3/execroot/org_tensorflow/bazel-out/k8-opt/bin/tensorflow/python/data/kernel_tests/interleave_test.runfiles/org_tensorflow/tensorflow/python/framework/test_combinations.py", line 366, in decorated
    execute_test_method()
  File "/dev/shm/build-branfosj-admin/branfosj-admin-up/TensorFlow/2.6.0/foss-2021a/tmpzwc1qwnh-bazel-tf/0db3977216c8a32952b67b0996300dd3/execroot/org_tensorflow/bazel-out/k8-opt/bin/tensorflow/python/data/kernel_tests/interleave_test.runfiles/org_tensorflow/tensorflow/python/framework/test_combinations.py", line 349, in execute_test_method
    test_method(**kwargs_to_pass)
  File "/dev/shm/build-branfosj-admin/branfosj-admin-up/TensorFlow/2.6.0/foss-2021a/tmpzwc1qwnh-bazel-tf/0db3977216c8a32952b67b0996300dd3/execroot/org_tensorflow/bazel-out/k8-opt/bin/tensorflow/python/data/kernel_tests/interleave_test.runfiles/org_tensorflow/tensorflow/python/data/kernel_tests/interleave_test.py", line 197, in testInterleaveDataset
    self.assertDatasetProduces(dataset, expected_output)
  File "/dev/shm/build-branfosj-admin/branfosj-admin-up/TensorFlow/2.6.0/foss-2021a/tmpzwc1qwnh-bazel-tf/0db3977216c8a32952b67b0996300dd3/execroot/org_tensorflow/bazel-out/k8-opt/bin/tensorflow/python/data/kernel_tests/interleave_test.runfiles/org_tensorflow/tensorflow/python/data/kernel_tests/test_base.py", line 232, in assertDatasetProduces
    self._compareOutputToExpected(result, expected_output, assert_items_equal)
  File "/dev/shm/build-branfosj-admin/branfosj-admin-up/TensorFlow/2.6.0/foss-2021a/tmpzwc1qwnh-bazel-tf/0db3977216c8a32952b67b0996300dd3/execroot/org_tensorflow/bazel-out/k8-opt/bin/tensorflow/python/data/kernel_tests/interleave_test.runfiles/org_tensorflow/tensorflow/python/data/kernel_tests/test_base.py", line 151, in _compareOutputToExpected
    self.assertValuesEqual(expected_value, result_value)
  File "/dev/shm/build-branfosj-admin/branfosj-admin-up/TensorFlow/2.6.0/foss-2021a/tmpzwc1qwnh-bazel-tf/0db3977216c8a32952b67b0996300dd3/execroot/org_tensorflow/bazel-out/k8-opt/bin/tensorflow/python/data/kernel_tests/interleave_test.runfiles/org_tensorflow/tensorflow/python/data/kernel_tests/test_base.py", line 92, in assertValuesEqual
    self.assertAllEqual(expected, actual)
  File "/dev/shm/build-branfosj-admin/branfosj-admin-up/TensorFlow/2.6.0/foss-2021a/tmpzwc1qwnh-bazel-tf/0db3977216c8a32952b67b0996300dd3/execroot/org_tensorflow/bazel-out/k8-opt/bin/tensorflow/python/data/kernel_tests/interleave_test.runfiles/org_tensorflow/tensorflow/python/framework/test_util.py", line 1275, in decorated
    return f(*args, **kwds)
  File "/dev/shm/build-branfosj-admin/branfosj-admin-up/TensorFlow/2.6.0/foss-2021a/tmpzwc1qwnh-bazel-tf/0db3977216c8a32952b67b0996300dd3/execroot/org_tensorflow/bazel-out/k8-opt/bin/tensorflow/python/data/kernel_tests/interleave_test.runfiles/org_tensorflow/tensorflow/python/framework/test_util.py", line 2979, in assertAllEqual
    np.testing.assert_array_equal(a, b, err_msg="\n".join(msgs))
  File "/rds/bear-apps/devel/eb-sjb-up/EL8/EL8-cas/software/SciPy-bundle/2021.05-foss-2021a/lib/python3.9/site-packages/numpy/testing/_private/utils.py", line 932, in assert_array_equal
    assert_array_compare(operator.__eq__, x, y, err_msg=err_msg,
  File "/rds/bear-apps/devel/eb-sjb-up/EL8/EL8-cas/software/SciPy-bundle/2021.05-foss-2021a/lib/python3.9/site-packages/numpy/testing/_private/utils.py", line 842, in assert_array_compare
    raise AssertionError(msg)
AssertionError:
Arrays are not equal

not equal lhs = array(6)
not equal rhs = array(4)
Mismatched elements: 1 / 1 (100%)
Max absolute difference: 2
Max relative difference: 0.5
 x: array(6)
 y: array(4)

----------------------------------------------------------------------

And an almost identical error for shard 9 of 24, with only the top few lines being different:

FAIL: testInterleaveDataset_test_mode_eager_tfapiversion_2_blocklength_3_cyclelength_None_inputvalues_4567_numparallelcalls_None (__main__.InterleaveTest)
InterleaveTest.testInterleaveDataset_test_mode_eager_tfapiversion_2_blocklength_3_cyclelength_None_inputvalues_4567_numparallelcalls_None
testInterleaveDataset_test_mode_eager_tfapiversion_2_blocklength_3_cyclelength_None_inputvalues_4567_numparallelcalls_None(mode='eager', tf_api_version=2, block_length=3, cycle_length=None, input_values=array([4, 5, 6, 7]), num_parallel_calls=None)

[akesandgren]

The "simple" solution here would maybe be to add a min_parallel config opt so we can set it to 8 in the easyconfig, which the framework can then check for ....

[branfosj]

Or the easyblock could detect < 8 cores and disable this test in that case, with a suitable warning.

[smoors]

the problem is probably due to multiple uses of multiprocessing.cpu_count() in the interleave_test, which gives a wrong result if not all cores are allocated to the job. see for example here: https://github.com/tensorflow/tensorflow/blob/39ff146388eb4b7f63d8e4c94786edf94433bb60/tensorflow/python/data/kernel_tests/interleave_test.py#L60

this may be fixed by replacing with len(os.sched_getaffinity(0)) (updated: fixed obvious mistake in replacement)

[akesandgren]

@smoors did you make a PR for that or?

[smoors]

not yet, too busy trying to make TF-2.7.1 pass the tests.

[smoors]

there's quite a lot of places where is used, so not sure what's the best way to deal with this. I'm thinking to do a brute-force find/replace in the source tree, but maybe that's a bit too brute? TF-2.6.0:

tensorflow/tools/test/system_info_lib.py:  cpu_info.num_cores = multiprocessing.cpu_count()
tensorflow/python/keras/integration_test/parameter_server_custom_training_loop_test.py:    if multiprocessing.cpu_count() < num_workers + 1:
tensorflow/python/keras/integration_test/parameter_server_keras_preprocessing_test.py:  if multiprocessing.cpu_count() < num_workers + 1:
tensorflow/python/data/ops/dataset_ops.py:      local_shard_func = lambda index, _: index % multiprocessing.cpu_count()
tensorflow/python/data/ops/dataset_ops.py:          cycle_length=multiprocessing.cpu_count(),
tensorflow/python/data/kernel_tests/snapshot_test.py:        num_snapshot_shards_per_run=multiprocessing.cpu_count())
tensorflow/python/data/kernel_tests/snapshot_test.py:        num_snapshot_shards_per_run=multiprocessing.cpu_count())
tensorflow/python/data/kernel_tests/snapshot_test.py:        num_snapshot_shards_per_run=multiprocessing.cpu_count())
tensorflow/python/data/kernel_tests/snapshot_test.py:        num_snapshot_shards_per_run=multiprocessing.cpu_count())
tensorflow/python/data/kernel_tests/snapshot_test.py:        num_snapshot_shards_per_run=multiprocessing.cpu_count())
tensorflow/python/data/kernel_tests/snapshot_test.py:        num_snapshot_shards_per_run=multiprocessing.cpu_count())
tensorflow/python/data/kernel_tests/snapshot_test.py:        num_snapshot_shards_per_run=multiprocessing.cpu_count())
tensorflow/python/data/kernel_tests/snapshot_test.py:        num_snapshot_shards_per_run=multiprocessing.cpu_count())
tensorflow/python/data/kernel_tests/snapshot_test.py:        num_snapshot_shards_per_run=multiprocessing.cpu_count())
tensorflow/python/data/kernel_tests/snapshot_test.py:        num_snapshot_shards_per_run=multiprocessing.cpu_count())
tensorflow/python/data/kernel_tests/interleave_test.py:      cycle_length = multiprocessing.cpu_count()
tensorflow/python/data/kernel_tests/interleave_test.py:      cycle_length = (multiprocessing.cpu_count() + 2) // 3
tensorflow/python/data/kernel_tests/interleave_test.py:      cycle_length = min(num_parallel_calls, multiprocessing.cpu_count())
tensorflow/lite/tools/pip_package/setup.py:    return multiprocessing.cpu_count()
tensorflow/python/data/experimental/ops/io.py:          cycle_length=multiprocessing.cpu_count(),

[smoors]

fix for interleave_test: https://github.com/easybuilders/easybuild-easyconfigs/blob/develop/easybuild/easyconfigs/t/TensorFlow/TensorFlow-2.7.1_fix_cpu_count.patch