We begin the construction of the new communicator by
extracting a process group, associated with the
MPI_COMM_WORLD communicator and call it world_group:
MPI_Comm_group ( MPI_COMM_WORLD, &world_group );
The next operation creates a new group, which will be called
worker_group, by excluding processes declared
in the array ranks from the world_group:
ranks[0] = server; MPI_Group_excl ( world_group, 1, ranks, &worker_group );
These two new functions, MPI_Comm_group and
MPI_Group_excl, are described in
Chapter 5
``Groups, Contexts, and
Communicators'', of ``MPI: A Message-Passing Interface Standard''.
The number of processes to be excluded from the new group
must be passed as the second parameter
to function MPI_Group_excl. In our case we exclude only one
process, the server.
Having created the new group, we can now create the communicator
associated with that group. We shall call that communicator workers:
MPI_Comm_create ( world, worker_group, &workers );
Once the communicator has been created we can release
the names of groups worker_group and world_group:
MPI_Group_free ( &worker_group ); MPI_Group_free ( &world_group );It is important to realise that this operation only frees the names of the groups, not the groups themselves. The groups survive inside the communicators, and if needed, they can be retrieved again by calling
MPI_Comm_group.
Process groups and communicators are a very important part of MPI. They make it possible to write parallel libraries, which can be accessed by user programs without colliding with their own nomenclature, and their own messages. The idea of process groups also lies at the basis of fault tolerant and virtually synchronous programming. But MPI, which is designed specifically for scientific and engineering computations, doesn't take it that far.