Jim, I am not sure on what you base you understanding regarding the different signature for MPI_ERROR_RETURNS. I can’t find anything suggesting this in the standard. My understanding is that MPI_COMM_CALL_ERRHANDLER is supposed to return MPI_SUCCESS if the error handler has been successfully called, as indicated in the following snippet from the standard.

This function invokes the error handler assigned to the communicator with the error code supplied. This function returns MPI_SUCCESS in C and the same value in IERROR if the error handler was successfully called (assuming the process is not aborted and the error handler returns).

Additionally there is no reason for a user defined error handler to return an error code, it should only react to the error it is informed about. We discussed a similar issue in the WG a while back. The errhandler receives the error as a reference, fact that seems to indicate that the error handle might be allowed to change the error. Unfortunately, this is very unsettling as we can’t allow a user-defined error handler to replace a legit MPI error with MPI_SUCCESS, which basically returns MPI_SUCCESS to the upper layer when the MPI library might be in a unusable state. George.

On Nov 19, 2014, at 19:20 , Jim Dinan <[email protected]> wrote:

Hi FT Folks,

I encountered this oddity while doing some hacking today. Apparently, MPI_ERRORS_RETURN has a different signature from all other error handlers; it returns a value and all other error handlers (e.g. MPI_ERRORS_ARE_FATAL and user-defined error handlers) have void return:

typedef void MPI_Comm_errhandler_function(MPI_Comm *, int *, ...);

Because of this, MPI_Comm_call_errhandler effectively does not do anything when the error handler is set to MPI_ERRORS_RETURN. More specifically, the following code,

return MPI_Comm_call_errhandler(my_comm, my_err_code);

always returns success.

To summarize the issues: (1) It is difficult or impossible to use the MPI error subsystem when the error handler is set to MPI_ERRORS_RETURN and (2) it is impossible for a user defined error handler to return an error code. These difficulties seem like they could be troublesome to users of FT that want to create higher level resilience/resilient libraries that interact with the MPI errors subsystem.

Is what I have encountered a real issue or am I misunderstanding something? If it is a real issue, is it something we queue up for discussion at the upcoming F2F?

Interested to hear your thoughts, ~Jim. _______________________________________________ mpiwg-ft mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-ft