[implicit_some]

I am learning how to use Open MPI in order to develop my own finite difference CFD solver in Fortran90.

I have read on a few tutorials and pages online that mpi_send and mpi_recv routines need to be called in order pass data from the "edge" cells from one mesh partition to its adjacent mesh partition.

What is stopping me from just using mpi_bcast? If I just broadcast the entire mesh data from the root processor to all processors and then carry out the same numerical process on all processors - but of course change the minimum/maximum cells that each processors processes.

Is this generally a less efficient way or something? Looking at the material online, it appears to me that no one uses this approach and I am wondering why. Is there a situation in which this is not possible and I have to use mpi_send/mpi_recv?

[flotus1]

The most important goal of an MPI communication scheme -right after correctness- is low overhead.
In general, you want to exchange as little data as possible, between as few processes as possible. Sending and receiving data costs time. If your solver spends a lot of time doing the MPI communication, you will get little to no speedup when increasing the number of processes.

BCAST is bad here because it sends data to all other processes, even those that don't need it. Which wastes valuable time.
Broadcasting the ENTIRE data from all cells is even worse, because you are sending data that actually none of the processes need.
Doing it this way will lead to negative speedup, on top of using tons of memory.