OpenPFC / Build a simulation

A process has a rank and a lifetime

Start MPI once, query its communicator and turn rank-local values into a global answer.

Includes the model, inputs, required headers and a short run guide.

Parallel C++ begins with a small lifetime question: who starts MPI, and who finishes it? OpenPFC’s environment wrapper owns those calls for a standalone program. A communicator then supplies the rank and group size.

pfc::mpi::environment environment;
pfc::mpi::communicator comm;
const int rank = comm.rank();
int sum = 0;
MPI_Allreduce(&rank, &sum, 1, MPI_INT, MPI_SUM, comm);
if (rank == 0)
  std::cout << "ranks,rank_sum\n" << comm.size() << ',' << sum << '\n';

The local value is the rank \(r\). The collective sums all ranks, so the independent answer is

\[\sum_{r=0}^{P-1}r=P(P-1)/2.\]

Every rank participates in the collective; only rank zero prints the result. Root-only output does not make the collective a root-only operation.

Run the same program with a different process count

openpfc run model.cpp input.json --launcher=none
openpfc run model.cpp input.json --ranks=2 --launcher=mpirun

The empty input is unused. The printed CSV reports ranks and rank_sum. The tests check \(P=1\) and \(P=2\). Changing the rank count changes the sum, not the source code or an assumed constant inside it.

The environment must be constructed before communicator operations and outlive them. Its constructor calls MPI_Init and its destructor calls MPI_Finalize; do not add a second owner in a host application that already initialised MPI. Use the host’s lifetime and pass its communicator in that case. A duplicated communicator manages a communication context, not a second MPI runtime. Continue with a class-owned session, or inspect the public environment wrapper.

The complete C++ model and starting input are available with this lesson.