[llvm-dev] Timeout tests timing out
David Blaikie via llvm-dev
Pass: http://lab.llvm.org:8011/builders/clang-cmake-armv8-lld/builds/5672
Fail: http://lab.llvm.org:8011/builders/clang-cmake-armv8-lld/builds/5673
Is there anything we can do to the buildbot? Or the tests? (bump up the time limits or maybe remove the tests as unreliable?)
David Blaikie via llvm-dev
_______________________________________________
LLVM Developers mailing list
llvm...@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
Dan Liew via llvm-dev
Unfortunately writing a reliable test is tricky given that the
functionality we're trying to test involves timing. I would advise
against disabling the test entirely because it actually tests
functionality that people use. I'd suggest bumping up the time limits.
This is what I've done in the past. See
commit 6dfcc78364fa3e8104d6e6634733863eb0bf4be8
Author: Dan Liew <d...@su-root.co.uk>
Date: Tue May 22 15:06:29 2018 +0000
[lit] Try to make `shtest-timeout.py` test more reliable by using a
larger timeout value. This really isn't very good because it will
still be susceptible to machine performance.
While we are here also fix a bug in validation of
`maxIndividualTestTime` where previously it wasn't checked if the
type was an int.
rdar://problem/40221572
llvm-svn: 332987
HTH,
Dan.
David Blaikie via llvm-dev
doesn't pull its weight. Is the code that implements the feature
liable to failure/often touched? If it's pretty static/failure is
unlikely, possibly the time and flaky failures aren't worth the value
of possibly catching a low-chance bug.
Another option might be to reduce how often/in which configurations
the test is run - LLVM_ENABLE_EXPENSIVE_CHECKS presumably only works
for code within LLVM itself, and not test cases - but maybe I'm wrong
there & this parameter could be used (& then the timing bumped up
quite a bit to try to make it much more reliable), or something
similar could be implemented at the lit check level?
Ah, compiler-rt tests use EXPENSIVE_CHECKS to disable certain tests:
./compiler-rt/test/lit.common.configured.in:set_default("expensive_checks",
@LLVM_ENABLE_EXPENSIVE_CHECKS_PYBOOL@)
./compiler-rt/test/fuzzer/large.test:UNSUPPORTED: expensive_checks
Could you bump the timeouts a fair bit and disable the tests except
under expensive checks?
Mehdi AMINI via llvm-dev
Dan Liew via llvm-dev
>
> I appreciate the value of the feature - but it's possible the test
> doesn't pull its weight. Is the code that implements the feature
> liable to failure/often touched? If it's pretty static/failure is
> unlikely, possibly the time and flaky failures aren't worth the value
> of possibly catching a low-chance bug.
I don't have many data points here. I haven't needed to touch the code
or tests since 2018. I haven't been asked to review any patches
related to the timeout feature since then so I presume nobody has
touched it since.
> Another option might be to reduce how often/in which configurations
> the test is run - LLVM_ENABLE_EXPENSIVE_CHECKS presumably only works
> for code within LLVM itself, and not test cases - but maybe I'm wrong
> there & this parameter could be used (& then the timing bumped up
> quite a bit to try to make it much more reliable), or something
> similar could be implemented at the lit check level?
>
> Ah, compiler-rt tests use EXPENSIVE_CHECKS to disable certain tests:
>
> ./compiler-rt/test/lit.common.configured.in:set_default("expensive_checks",
> @LLVM_ENABLE_EXPENSIVE_CHECKS_PYBOOL@)
> ./compiler-rt/test/fuzzer/large.test:UNSUPPORTED: expensive_checks
>
> Could you bump the timeouts a fair bit and disable the tests except
> under expensive checks?
Only running that test in the "expensive_checks" configuration kind of
abuses that configuration option because the test is actually not an
expensive check. It's just that the test is sensitive to the
performance of the running system. If the system is under heavy load
when the test is running it's more likely to fail.
One thing we could do to remove fragility in the test is to remove the
running of `short.py` in the test. This is only invoked to check that
it's possible for a command to run to completion in the presence of a
fixed timeout. If we can live without testing that part (i.e. we only
test that a timeout can be reached) then the test should be much more
robust.
If for some reason that's not enough another thing that might help
with this problem is to not run the lit tests at the same time as all
the other tests. The hope here is that the lit tests would run while
the system load is lower. We could do this by changing the build
system to
* Make all `check-*` (except `check-lit`) targets depend on the
`check-lit` target. This would ensure that lit gets tested before we
test everything else in LLVM.
* Remove the lit test directory from inclusion in all `check-*`
targets (except `check-lit`) so we don't test lit twice.
This also has a nice benefit of testing lit **before** we use it to
test everything else in LLVM. Today we don't do that, we just run all
the tests in one big lit invocation (i.e. we test lit at the same time
that we are using it). It does have some downsides though
* If any of the lit tests fail we won't run the rest of the tests so
you won't get the results of running the other tests until the lit
tests are fixed.
* When running any of the `check-*` targets (apart from `check-lit`)
you have to wait for the lit tests to run before any other tests run.
What do you think?
Dan Liew via llvm-dev
>
> Hi,
>
> Depending on timing in a test is quite brittle in general: can we mock the timeout instead and make this fully deterministic somehow?
>
> The test has this comment:
>
> > FIXME: This test is fragile because it relies on time which can
> > be affected by system performance. In particular we are currently
> > assuming that `short.py` can be successfully executed within 2
> > seconds of wallclock time.
>
> Maybe "short.py" can be replaced by adding into lit itself a "no op" which would just not really spawn a process and instead mark the task as completed immediately internally?
I think it's probably easier to just remove running "short.py" in the
test. We'd lose some test coverage because we no longer test that a
program can run to completion when a timeout is set but maybe this is
an acceptable trade-off to avoid the test being flakey?
David Blaikie via llvm-dev
Ah, sorry - my point was to make it expensive: Make it long enough
that it's more reliable even when the machine is under load. But
making it long makes it expensive, so then classify it as such.
> It's just that the test is sensitive to the
> performance of the running system. If the system is under heavy load
> when the test is running it's more likely to fail.
>
> One thing we could do to remove fragility in the test is to remove the
> running of `short.py` in the test. This is only invoked to check that
> it's possible for a command to run to completion in the presence of a
> fixed timeout. If we can live without testing that part (i.e. we only
> test that a timeout can be reached) then the test should be much more
> robust.
If you're on board with that, it's a tradeoff I think is probably
reasonable from a test coverage V reliability V development time
tradeoff.
> If for some reason that's not enough another thing that might help
> with this problem is to not run the lit tests at the same time as all
> the other tests. The hope here is that the lit tests would run while
> the system load is lower. We could do this by changing the build
> system to
>
> * Make all `check-*` (except `check-lit`) targets depend on the
> `check-lit` target. This would ensure that lit gets tested before we
> test everything else in LLVM.
> * Remove the lit test directory from inclusion in all `check-*`
> targets (except `check-lit`) so we don't test lit twice.
>
> This also has a nice benefit of testing lit **before** we use it to
> test everything else in LLVM. Today we don't do that, we just run all
> the tests in one big lit invocation (i.e. we test lit at the same time
> that we are using it). It does have some downsides though
>
> * If any of the lit tests fail we won't run the rest of the tests so
> you won't get the results of running the other tests until the lit
> tests are fixed.
> * When running any of the `check-*` targets (apart from `check-lit`)
> you have to wait for the lit tests to run before any other tests run.
Reckon that's probably not the best way to go - wouldn't totally fix
this bug (buildbot might be busy for other reasons (maybe it's running
some regular maintenance, etc) or lit tests might still saturate cores
& slow down the test) & would slow down test iteration by having
another "pinch point" in the build graph - waiting for the last of the
lit tests to run (this kind of period reduces core utilization - as
the last few tests finish and no new work is scheduled) before
starting up all the cores again to run the next batch of tests.
- Dave
Dan Liew via llvm-dev
> > running of `short.py` in the test. This is only invoked to check that
> > it's possible for a command to run to completion in the presence of a
> > fixed timeout. If we can live without testing that part (i.e. we only
> > test that a timeout can be reached) then the test should be much more
> > robust.
>
> If you're on board with that, it's a tradeoff I think is probably
> reasonable from a test coverage V reliability V development time
> tradeoff.
Sorry for the delay here. I've put a patch up for review that goes
with this approach: https://reviews.llvm.org/D88807
David Blaikie via llvm-dev
Arthur Eubanks via llvm-dev
_______________________________________________
David Blaikie via llvm-dev
Diana Picus via llvm-dev
David Blaikie via llvm-dev
Dan Liew via llvm-dev
someone else already fixed it :)