On Oct 11, 2013, at 4:44 PM, Jeff Hammond <[email protected]> wrote:

I apologize if this is a stupid question, but I would like MPICH to crash and burn rather than terminate gently when an error occurs so that I can gdb bt it to know where the error is in my program, since I am absolutely certain that this error is in the application and not MPICH.

Ideally, I would be able to select the failure mode at runtime, e.g. MPICH_FAILURE_MODE={hard,soft}, so that I can get either the nice MPICH trace when I think MPICH is the problem and a brutal light-the-machine-room-on-fire-and-abort-the-universe error when I think it's my fault.

For assertion failures, I think I used to change MPIR_Assert_fail/MPIR_Assert_fail_fmt to just call "abort()" instead of MPID_Abort. As long as my ulimit settings were correct then I'd get a core file that I could load in the debugger. The process manager will take care of cleaning up the other processes anyway. IIRC, you can perform a similar trick with MPIR_Err_create_code to catch error codes at the time they are created, though that may not be exactly the same spot that you want to examine. -Dave