CPSC W50 at Calvin Collge: Using C++ with MPI

Parallel Computing: Using MPI with C++

There are a number of advantages to using MPI with C++ instead of C or Fortran:

You can apply OOA/D to your problem and get reuseable code;
You can pass values via reference parameters (unlike C);
You get a stronger type system that catches more logic errors;
You get classes, templates, inline functions, overloading, and many other features that facilitate writing safe, maintainable, and efficient code.

To use MPI with C++, you must configure and install MPI to use a a C++ compiler (such as GNU's g++). The file config.help that accompanies the distribution lists the various switches you will need to set for configure, including -c++=yourCompiler, -c++linker=yourLinker, -c++flags=desiredFlags, and so on. (I don't think we got all of these set right here at Calvin, so we have to specify a few extra arguments when we compile and link, as you can see in our Makefile.)

Once that has been done, you can issue calls to MPI functions from within a C++ program, just as you would from within C or Fortran.

We think that the slick way to go is to use our MPIProcess class, which is in the file MPI_Process.h. Using this class, every MPI program has the same simple form:

   #include "MPI_Process.h"

   // ... Define MPIProcess::run() with the desired behavior ...

   int main(int argc, char ** argv)
   {
      MPIProcess aProcess(argc, argv);   // NOTE: no ampersands needed!
      aProcess.run();
   }

The file evenOdd.cpp contains a simple example program, and here is our Makefile which will need to be adjusted for your own system. (Netscape replaces tabs with spaces, so you will need to replace the spaces at the beginning of each line of the Makefile with tabs.)

To run the resulting program (e.g., evenOdd), use the usual mpirun command, for example:

  mpirun -np 8 evenOdd

Whatever code is used to define MPIProcess::run() will be then executed in SPMD fashion on your multiprocessor. The MPI_Process constructor takes care of all initialization, and the MPI_Process destructor takes care of all clean-up, without you having to do anything special.

At present, the MPI_Process interface includes the functions

MPI_Process(int & argc, char ** & argv) which constructs (initializes) an MPI_Process;
~MPI_Process() which destroys (cleans up) an MPI_Process;
char * hostName() const which returns the name of the host on which the MPI_Process receiving this message is running;
int rank() const which returns the rank of the MPI_Process receiving this message in the communicator MPI_COMM_WORLD.
int rank(const MPI_Comm & aCommunicator) const which returns the rank of the MPI_Process receiving this message
int numberOfProcesses() const which returns the number of processes for the MPI_Process receiving this message in the communicator MPI_COMM_WORLD.
int numberOfProcesses(const MPI_Comm & aCommunicator) const which returns the number of processes for the MPI_Process receiving this message in the communicator aCommunicator.

There is also a copy-constructor provided (commented out), if some computation needs one.

This is our first step, and we hope to add more components (such as an MPI_Communicator class) in the future, as time permits. Have fun!

Calvin > CS > 374 > MPI Resources > MPI and C++

This page maintained by Joel Adams.