I'd like to take advantage of having multi-cores on both my primary development machine and all except one of the auto-testers.  One of the problems with doing that is that several of the unit tests use hard coded filenames (one example, grep for "deleteme").  I'm pretty sure there are others.

Could I ask for volunteer(s) to

  1) examine phobos for these sorts of problems and fix them

  2) start testing with -j# where # is greater than 1

NOTE: just switching to random filename generation only reduces the statistical rate of failure.  The right answer is to use either mkstemp or tmpfile (or, sigh, a rewrite of them in D if you alergic to using libc). Keep in mind that unit tests are also a form of best practice documentation.

DMD and Druntime are both in good shape when it comes to both building and running the tests in parallel.  The phobos build logic appears to be solid enough to survive parallel building, though I haven't been doing parallel builds of it until today.  It's just phobos's tests that are at issue, and even there only if the tests happen to run at the same time.

I really don't want to add additional intermitent test failures to the system, we've got enough of those already.

Thoughts?

Thanks,
Brad

On 16 July 2011 00:47, Brad Roberts <braddr at puremagic.com> wrote:
> I'd like to take advantage of having multi-cores on both my primary development machine and all except one of the auto-testers. ?One of the problems with doing that is that several of the unit tests use hard coded filenames (one example, grep for "deleteme"). ?I'm pretty sure there are others.

Some of the filename tests have other problems as well. Some of them don't work on Windows. I replaced a couple of them with pragma(msg) in the distant past, just to get the unit tests passing again (grep for "broken test").


>
> Could I ask for volunteer(s) to
>
> ?1) examine phobos for these sorts of problems and fix them
>
> ?2) start testing with -j# where # is greater than 1
>
> NOTE: just switching to random filename generation only reduces the statistical rate of failure. ?The right answer is to use either mkstemp or tmpfile (or, sigh, a rewrite of them in D if you alergic to using libc). Keep in mind that unit tests are also a form of best practice documentation.
>
> DMD and Druntime are both in good shape when it comes to both building and running the tests in parallel. ?The phobos build logic appears to be solid enough to survive parallel building, though I haven't been doing parallel builds of it until today. ?It's just phobos's tests that are at issue, and even there only if the tests happen to run at the same time.
>
> I really don't want to add additional intermitent test failures to the system, we've got enough of those already.
>
> Thoughts?
>
> Thanks,
> Brad
>
> _______________________________________________
> phobos mailing list
> phobos at puremagic.com
> http://lists.puremagic.com/mailman/listinfo/phobos
>