- removed comment
CarpetIOHDF5 may segfault when one_file_per_group is selected
Issue #410
closed
I was facing one of those cryptic MPI error messages on Ranger, where your job quits without any useful error message. It turns out that this parameter file had
CarpetIOHDF5::one_file_per_group = "yes" CarpetIOHDF5::out2D_vars = " ADMBase::gxx"
and CarpetIOHDF5 would try to loop over the group variables when just one of them was selected. As a result it has accessed memory address that it was not supposed to, consequently killing the job.
I have attached the stack back trace.
Keyword: Carpet
Keyword: git
Keyword: version
Comments (5)
-
-
- changed status to open
- removed comment
-
- removed comment
This looks fine -- please apply.
-
- changed status to resolved
- removed comment
applied. Thanks for pointing out what to add (was really simple this way).
-
- edited description
- changed status to closed
- Log in to comment
The corresponding code in CarpetIOHDF5.cc, which outputs data that are not slices, uses the same algorithm. However, it contains an additional check ensuring that a default request is used if the corresponding request does not exist. Look for calls to IOUtil_DefaultIORequest to find this code. I believe that an equivalent logic would correct this problem in OutputSlice.cc.