Best way to separate parallel MPI part in sequential program - cuda

Question

I've have a huge sequntial program, in which I want to parallelize some algo with MPI and CUDA. How correctly separate sequential part from parallel? The problem lies in nesting of parallel algo, and using of slurm or loadLeveler as well(e.g. on my MPI cluster I can't write something like: mpirun -n 1 a.out: -n 2 b.out).
Example:
int main()
{
funcA();
}
void funcA()
{
funcB();
}
void funcB()
{
parallel algo starts here....
}

Answer 1

I've found a great solution, for this problem. This is sample code:
#include <iostream>
#include <mpi.h>
#include <unistd.h>
using namespace std;
int main(int argc, char** argv) {
MPI_Init(&argc, &argv);
int r;
MPI_Comm_rank(MPI_COMM_WORLD, &r);
if (r == 0) {
cout << "[GUI]Start perfoming initialization...." << endl;
sleep(2);
cout << "[GUI]Send command to start execution...." << endl;
int command = 1;
//TODO: now it's hardcoded to send data to 1 proc
MPI_Send(&command, 1, MPI_INT, 1, 0, MPI_COMM_WORLD);
cout << "[GUI]Waiting for execution results..." << endl;
int buf[5];
MPI_Recv(&buf, 5, MPI_INT, MPI_ANY_SOURCE, MPI_ANY_TAG, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
for (int i=0; i<5; i++)
{
cout << "buf["<< i << "] = " << buf[i] << endl;
}
} else {
int command;
MPI_Recv(&command, 1, MPI_INT, MPI_ANY_SOURCE, MPI_ANY_TAG, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
cout << "Received command: " << command << endl;
if (command == 1) {
cout << "[ALGO]Receive command to start execution" << endl;
sleep(2);
cout << "[ALGO]Send computed data..." << endl;
int buf[5] = {5,4,3,2,1};
MPI_Send(&buf, 5, MPI_INT, 0, 0, MPI_COMM_WORLD);
}
}
MPI_Finalize();
return 0;
}