OpenMP

Learning Objectives

  • Describe the OpenMP fork-join execution model

  • Add #pragma omp parallel and #pragma omp parallel for to a serial loop

  • Control thread count and query thread ID at runtime

  • Time a parallel region and compute speedup

  • Recognize data race conditions and use the reduction clause to avoid them

OpenMP Basics

OpenMP is a directive-based API for shared-memory parallelism. It requires no changes to data structures. You annotate loops and parallel regions with #pragma directives, link with -fopenmp, and the compiler generates multithreaded code. The same source file compiles correctly without -fopenmp and runs serially.

The Fork-Join Model

OpenMP programs follow a fork-join execution model:

  1. The program begins as a single master thread.

  2. When execution reaches a #pragma omp parallel block, the runtime forks a team of threads. All threads execute the block concurrently.

  3. At the closing }, threads join and the master thread continues alone.

Controlling Threads

Mechanism

Effect

OMP_NUM_THREADS=N (env var)

Set default team size before launching the program

omp_set_num_threads(N)

Set team size in code before entering a parallel region

omp_get_num_threads()

Query team size from inside a parallel region

omp_get_thread_num()

Query caller’s zero-based rank (0 = master thread)

Warning

Outside a #pragma omp parallel block, omp_get_num_threads() always returns 1 regardless of OMP_NUM_THREADS.

Parallel Vector Addition

#include <stdio.h>
#include <omp.h>

#define ARRAY_SIZE 8000000
static double a[ARRAY_SIZE], b[ARRAY_SIZE], c[ARRAY_SIZE];

void vector_add(double *c, double *a, double *b, int n)
{
#pragma omp parallel for
  for (int i = 0; i < n; i++){
    c[i] = a[i] + b[i];
  }
}

int main(int argc, char *argv[]){
  for (int i = 0; i < ARRAY_SIZE; i++) {
    a[i] = 1.0;
    b[i] = 2.0;
  }

  #pragma omp parallel
  if (omp_get_thread_num() == 0)
    printf("Running with %d thread(s)\n", omp_get_num_threads());

  vector_add(c, a, b, ARRAY_SIZE);
}

Building and Running

# Compile with OpenMP
gcc -fopenmp -O3 vecadd.c -o vecadd

# Set thread count and run
export OMP_NUM_THREADS=8
./vecadd

Reduction: Avoiding Data Races

When multiple threads update a shared variable, you get a data race. Use reduction to safely accumulate:

double sum = 0.0;
#pragma omp parallel for reduction(+:sum)
for (int i = 0; i < n; i++) {
  sum += a[i];  // Each thread gets a private copy; combined at end
}

Reduction operators: +, -, *, &, |, ^, min, max.

Running on ARC Clusters

For OpenMP jobs on ARC clusters:

  1. Request CPU cores:

    #SBATCH --cpus-per-task=8  # Number of OpenMP threads

  2. Set thread count:

    export OMP_NUM_THREADS=8

  3. Submit:

    sbatch run.slurm

See Running Jobs for SLURM details.

References