On 14/01/16 00:08, Fotis Georgatos wrote:
Hi Ken, Victor,
On Jan 13, 2016, at 3:25 PM, Kenneth Hoste <[email protected]> wrote:
According to our findings, "ulimit -v" must be set to 'unlimited', while "ulimit
-s" must be set to something else than 'unlimited' for the tests to work.
There may be more going on though, since that's exactly what you have according
to your debug log.
Maybe the stack size limit should be higher? Not sure…
I’ve seen this issue before. I hadn’t debugged it fully, but it appeared to
relate to glibc-2.12 of RHEL6 and friends (hello Centos6, hello SL6);
AFAIK, the issue goes away with glibc 2.13 and anything greater. If I recall
the story right, clang v3.4 and above will fail the sanitiser test,
because they are picky for a situation that can indeed imply a race condition
(if true, well done! :) - can someone else confirm this?
btw. try to install either of clang/3.3 and clang/3.4 to verify the above
statement, across glibc cases.
Ward has issued a PR to add a new option to hard disable only the sanitizer
checks, see https://github.com/hpcugent/easybuild-easyblocks/pull/806.
That should go in very soon, and so will be part of the next release (EasyBuild
v2.6.0).
Nice to have that as a tuneable.
With that in place, adding "skip_sanitizer_tests = True" will be sufficient to
dance around this issue, as opposed to having to skip *all* tests and disable
bootstrapping, as I mentioned earlier.
We should consider having that enabled by default to skip these troublesome
tests... Ward?
I’m inclined to agree, however it’s good to have a better understanding if that
introduces an “unsafe" situation.
Basically, if behaviour is the same like it used to be with clang/3.3, it may
be OK to disable as default even if imperfect;
but if the v3.4 sanitiser checks are more strict for a strong reason (race
conditions looming ahead), I’d be more skeptical...
We merged the pull request right not with having 'skip_sanitizer_tests'
disabled by default for now.
regards,
Kenneth
