C++ test Core dump DROPOUT_PERF.TimingGPU

[marcoabreu]

http://jenkins.mxnet-ci.amazon-ml.com/blue/organizations/jenkins/incubator-mxnet/detail/master/399/pipeline

Timing: 50 iterations of 10 calls, shape = [50,3,18,32] = 86,400 items 

Dropout Operator CPU:  Timing [Forward] 28.45100 ms, avg: 0.05690 ms X 500 passes



[       OK ] DROPOUT_PERF.TimingCPU (2745 ms)

[ RUN      ] DROPOUT_PERF.TimingGPU

Timing: 50 iterations of 10 calls, shape = [1,1,28,28] = 784 items 

terminate called after throwing an instance of 'dmlc::Error'

  what():  [22:22:55] ../mshadow/mshadow/./stream_gpu-inl.h:182: Check failed: e == cudaSuccess CUDA: an illegal memory access was encountered



Stack trace returned 10 entries:

[bt] (0) build/tests/mxnet_unit_tests(dmlc::StackTrace[abi:cxx11]()+0x56) [0xe15776]

[bt] (1) build/tests/mxnet_unit_tests(dmlc::LogMessageFatal::~LogMessageFatal()+0x28) [0xe15d78]

[bt] (2) build/tests/mxnet_unit_tests(void mshadow::DeleteStream<mshadow::gpu>(mshadow::Stream<mshadow::gpu>*)+0xb9) [0xe183f9]

[bt] (3) build/tests/mxnet_unit_tests(mxnet::test::op::CoreOpExecutor<float, float>::~CoreOpExecutor()+0x151) [0xe5f321]

[bt] (4) build/tests/mxnet_unit_tests(std::_Sp_counted_base<(__gnu_cxx::_Lock_policy)2>::_M_release()+0x46) [0xe1b316]

[bt] (5) build/tests/mxnet_unit_tests() [0xf0435e]

[bt] (6) build/tests/mxnet_unit_tests(DROPOUT_PERF_TimingGPU_Test::TestBody()+0x5e1) [0xf06401]

[bt] (7) build/tests/mxnet_unit_tests(void testing::internal::HandleExceptionsInMethodIfSupported<testing::Test, void>(testing::Test*, void (testing::Test::*)(), char const*)+0x43) [0xf36f03]

[bt] (8) build/tests/mxnet_unit_tests(testing::Test::Run()+0xba) [0xf27f3a]

[bt] (9) build/tests/mxnet_unit_tests(testing::TestInfo::Run()+0x118) [0xf28088]





/workspace/tests/ci_build/with_the_same_user: line 47: 83345 Aborted                 (core dumped) sudo -u "#${CI_BUILD_UID}" --preserve-env "LD_LIBRARY_PATH=${LD_LIBRARY_PATH}" "HOME=${CI_BUILD_HOME}" ${COMMAND[@]}

script returned exit code 134

[marcoabreu]

@cjolivier01

[cjolivier01]

looks like someone broke it, eh?

[marcoabreu]

Yep :(

[cjolivier01]

then it's not "flaky", right? I mean, it's never broken before, right? I also noticed it breaking in another build today.
I wonder if it ever broke pre-mkldnn merge?

[marcoabreu]

It's flaky because otherwise it would have been reported during PR stage.

Well the number of test failures drastically increased since the mkldnn merge. But that's why we merged it, right? To get some data early on.

[anirudh2290]

just a side note, should we add try catch around for TESTs and fail so that consecutive tests continue to execute ?

[marcoabreu]

No, this would waste too many resources. Generally, multiple failures can be fixed by one change. If we wouldn't follow fail-fast, a lot of time would be wasted. Considering the fact that PRs should be small, the chance of multiple independent bugs causing multiple test failures is quite small compared to the number of total tests executed within every PR.

The right approach is to fix one test failure, create a new commit and then fix the next test failure. While this actually sounds pretty wasteful as well, past experience has shown that this is rarely required.

Additionally, we're currently working on something that allows you to run tests locally in exactly the same way the CI does. This will allow you to test yourself before pushing it to the CI.

[rajanksin]

@marcoabreu I ran this test ~1000 times, couldnt replicate the failure. Can we close this issue ?

[KellenSunderland]

@spidydev @marcoabreu I think there were some crossed wires on this one. This test is actually enabled. Was it at one time disabled?

[rajanksin]

@KellenSunderland From the history, I dont think the test was ever disabled.

[KellenSunderland]

@spidydev agree. @marcoabreu think we could close this one?