[petsc-dev] What to do when it's hopeless?

[Jed Brown]

On Thu, 13 May 2010 13:07:16 -0500, Barry Smith <bsmith at mcs.anl.gov> wrote:
> 
> On May 13, 2010, at 10:48 AM, Jed Brown wrote:
> 
> > I'm concerned by the following which appears in
> > PetscAttachDebuggerErrorHandler() and Petsc_MPI_DebuggerOnError():
> > 
> >  ierr = PetscAttachDebugger();
> >  if (ierr) { /* hopeless so get out */
> >    MPI_Finalize();
> >    exit(*flag);
> >  }
> > 
> > Since the error handler is not guaranteed to be called collectively, it
> > seems completely unreasonable to call a collective function, especially
> > not one with broader scope than PETSC_COMM_WORLD.  It seems to me that
> > the hopeless case should instead call MPI_Abort() on the relevant
> > communicator.
> 
>    You are probably right; this code fragment is so old it may have been written before I knew what MPI_Abort() was. I am changing it.

Okay, great.

> >  Similarly, why does Petsc_MPI_AbortOnError() call abort()
> > instead of MPI_Abort()?
> 
>    I think this is clearly documented in the comment above it.  Here is the comment:
> 
>  "This is so MPI errors in the debugger will leave all the stack 
>   frames. The default abort cleans up and exits."
> 
>   You use this error handler if you are running in the debugger (for example with -start_in_debugger). If you call MPI_Abort() it cleans up and you get no useful stack information in the debugger. I will change the comment to:
> 
> "This is so MPI errors in the debugger will leave all the stack 
>   frames. The default MP_Abort() cleans up and exits thus providing no useful information
>   in the debugger hence we call abort() instead of MPI_Abort()."
> 
> Is it clear now, or am I missing something?

Much clearer, now I don't have to think to understand it.

Jed