Test pmrr1 reveals weird (but common) pattern of non-deterministic failure. (Direct)

[Ryan Newton]

When running the test suite generated by --enable-tests for the monad-par package, in revision 60aabe6, I see non-determinstic failures on the Direct scheduler.

Specifically, if I try to run this command 10 times:

./dist/build/test-monad-par/test-monad-par -j1

It will fail. However, the weird part is that it depends on running multiple tests together. If I run JUST the pmrr1 test, even 1000 times, I cannot reproduce the error:

./dist/build/test-monad-par/test-monad-par -j1 -t pmrr1

This is exactly the same pattern I've seen in other projects that use test-framework to drive parallel tests, including lattice-par and haskell-lockfree -- the outcome depends on what set of tests is run. I haven't figured out WHY this is yet.

[Ryan Newton]

Let's see, referring to you by name, @aturon, should bring you into this thread. This was the same problem we were having with SNZI counters right?