Assert function for MPI Programs

Question

I have to validate state of some data in a MPI program. The program will run on a super computer with a distributed memory system. A quick research about C Standard's assert function revealed that assert internally uses abort() function for program termination. I haven't found much information as to how abort() works on a multi-process program, especially in MPI's context - it is very different than POSIX enviroment. Does abort() only terminate the process in which it is called, or can it terminate all the processes?

And finally how would I really terminate all processes of a MPI program when a condition fails? Is there a built-in assert in MPI library?

Answer 1

abort() only terminates the MPI task that invokes it.

It is very likely that will be detected by mpirun and/or the resource manager, and kill all the MPI job (e.g. all the MPI tasks on all nodes) after that.

That being said, this is library/system dependent, and you should double check that first.

The right way to terminate a MPI job is to

MPI_Abort(MPI_COMM_WORLD, errorcode)

errorcode is an int and is generally assigned a strictly positive value.