MPI runtime weird example

Hi guys,

I have a weird mpi example where I don't understand the used runtimes.

#include <iostream>
#include <mpi.h>
#include <chrono>
#include <cmath>

void compute(int iterations) {
    double sum = 0.0;
    for (long i = 0; i < iterations; ++i) {
        sum += std::sin(i) * std::cos(i);
    }
    std::cout << "Computation result: " << sum << std::endl; // Optional: Just to use the result and avoid optimization
}

int main(int argc, char *argv[]) {
    MPI_Init(&argc, &argv); // Initialize MPI
    MPI_Comm comm = MPI_COMM_WORLD;
    int rank, nr_proc;
    MPI_Comm_rank(comm, &rank);
    MPI_Comm_size(comm, &nr_proc);

    long N = atol(argv[1]); // Number of iterations for computation

    auto start = std::chrono::high_resolution_clock::now();

    compute(N); // Perform the computation

    auto end = std::chrono::high_resolution_clock::now();
    std::chrono::duration<double> duration = end - start;
    std::cout << "Process " << rank << " took " << duration.count() << " seconds." << std::endl;

    MPI_Finalize(); // Finalize MPI
    return 0;
}

Running this with mpirun --oversubscribe -n 1 mpi_test.exe 1000000000 prints a runtime of 28s.
But running it with `-n 5` yields

Computation result: -0.133023
Process 3 took 36.1116 seconds.
Computation result: -0.133023
Process 4 took 36.3062 seconds.
Computation result: -0.133023
Process 1 took 55.0856 seconds.
Computation result: -0.133023
Process 0 took 55.2955 seconds.
Computation result: -0.133023
Process 2 took 64.1076 seconds.

Why do some processes need so much more time to do the same thing? My CPU has 12 logical cores.

edit:
Apparently, my CPU has 2 performance cores (allowing hyperthreading) and 8 efficiency cores with no hyperthreading.

But then I would expect the 4 processes to have the same runtime. These are the processes run on the hyperthreads of the 2 performance cores. Only the last process should run slower.

Last edited on

jonnin (11437)

its hard to know what is really going on with caches, other processes, the os scheduler, and so on.

Topic archived. No new replies allowed.

C++

Forum

MPI runtime weird example