Compiling and Running C++ MPI Applications

Compiling and Running C++ MPI Applications

This guide explains how to load the Open MPI module, compile a distributed C++ program using the MPI wrapper, and submit an MPI job to the unite partition on the cluster.

IMPORTANT: Compilation Policy

WARNING: Do not run compilation commands directly on the Login Node!
High-performance libraries like Open MPI can cause heavy overhead during compilation. Always allocate an interactive compute node session on the unite partition using `srun` before building your code.

Available Modules

To see the available Open MPI configurations, run:

$ module avail unite/mpi

Step-by-Step Guide

1. Request an Interactive Node for Compilation

Switch to a safe environment on a compute node under the unite partition:

$ srun --partition=unite --cpus-per-task=2 --time=00:30:00 --pty bash

Once your command prompt updates, you are active on a compute node.

2. Load the Required Modules

MPI modules require a compatible compiler toolchain. Load both the GCC compiler and the corresponding Open MPI library:

$module load unite/compilers/gcc-14.3.0
$ module load unite/mpi/4.1.0

To verify your environment's active C++ MPI compiler wrapper, run:

$ mpicxx --version

3. Create a Sample MPI C++ File

Create a parallel C++ file named `mpi_hello.cpp`. This example initializes the MPI environment, determines the rank of each process, and identifies how many total processes are running.

mpi_hello.cpp

#include <mpi.h>
#include <iostream>
 
int main(int argc, char** argv) {
    // Initialize the MPI environment
    MPI_Init(&argc, &argv);
 
    // Get the number of processes
    int world_size;
    MPI_Comm_size(MPI_COMM_WORLD, &world_size);
 
    // Get the rank (ID) of the process
    int world_rank;
    MPI_Comm_rank(MPI_COMM_WORLD, &world_rank);
 
    // Get the name of the processor node
    char processor_name[MPI_MAX_PROCESSOR_NAME];
    int name_len;
    MPI_Get_processor_name(processor_name, &name_len);
 
    // Print off a hello world message from each process
    std::cout << "Hello world from process rank " << world_rank 
              << " out of " << world_size 
              << " processors on node " << processor_name << std::endl;
 
    // Finalize the MPI environment.
    MPI_Finalize();
    return 0;
}

4. Compilation

Compile your application using the `mpicxx` wrapper script, which handles adding all necessary linking flags for Open MPI automatically.

$ mpicxx -O3 -std=c++17 mpi_hello.cpp -o mpi_hello_executable

* `mpicxx`: The C++ compilation wrapper script for Open MPI. * `-O3`: High-level optimization flag. * `-o mpi_hello_executable`: The resulting binary output.

Exit your interactive session to return to the login node when the build completes successfully:

$ exit

Running Distributed Jobs via Slurm Batch

To launch an MPI job across multiple processors or nodes via Slurm, create a batch script named `submit_mpi.sh`:

<file bash submit_mpi.sh> #!/bin/bash #SBATCH –job-name=mpi_cpp_job #SBATCH –partition=unite # Target partition for the UNITE cluster #SBATCH

Table of Contents